Content Operations

What happens when you feed years of messy content into AI? In this episode, Bill Swallow and Alan Pringle dig into the content debt crisis, including increased system costs, neglected localization, and the fallout of “just use AI” mandates. They share practical insights to help organizations get back on track.

Alan Pringle: Is your content updated? Does it reflect the latest information? Is it created for all the different locales that your company serves? Is it in different languages? That is another pile of debt that when you start looking at AI, all the problems will be very brutally magnified, and you’re going to have to address them to really have a large language model that works at all.

Related links:

Balancing automation, accuracy, and authenticity: AI in localization (podcast)

Forbes: AI Costs More Than The People It Replaced

Taming AI: Using AI for content conversion at scale (podcast)

LinkedIn:

Bill Swallow

Alan Pringle

Transcript:

Disclaimer: This is a machine-generated transcript with edits.

Introduction with ambient background music

Christine Cuellar: From Scriptorium, this is Content Operations, a show that delivers industry-leading insights for global organizations.

Bill Swallow: In the end, you have a unified experience so that people aren’t relearning how to engage with your content in every context you produce it.

Sarah O’Keefe: Change is perceived as being risky; you have to convince me that making the change is less risky than not making the change.

Alan Pringle: And at some point, you are going to have tools, technology, and processes that no longer support your needs, so if you think about that ahead of time, you’re going to be much better off.

End of introduction

Bill Swallow: Hi everybody, I’m Bill Swallow.

Alan Pringle: And I’m Alan Pringle.

BS: And today’s episode is going to focus more on the content debt crisis, AI edition.

AP: And it’s also going to be the complaint edition, surprise, surprise, because there’s a lot of things in the AI world right now that are still making me cranky. And I am sure we will talk about them at some point.

BS: Yeah. So with the rise of AI, I think it’s kind of holding a microscope to a lot of the technical debt that we’ve been seeing over the years in content operations in general, whether you have outdated authoring formats or content that’s not being updated on a regular basis, new delivery formats not necessarily meeting the needs of the users, and so forth. And all of that is kind of compiling or snowballing into a bigger problem once you start feeding all of this stuff into AI.

AP: Right. And it’s interesting to me how everyone’s talking about AI as being this productivity tool. In a lot of ways it is, but that’s not what the focus of this is. In a way, it is also sort of a consultant for you. As you just mentioned, Bill, when you start looking at AI and delivering it, treating it as a delivery endpoint for your content, a distribution endpoint, you are going to start to discover that your processes on the back end for creating and distributing your content are not what they should be. So it’s kind of like this consultant saying, Hey, you need to do better over here. And that is where a lot of this debt is coming from, from my point of view.

BS: Mm-hmm. The unfortunate part of that consultant is that it’s not offering advice on how to fix it, but it certainly is pointing out the issues.

AP: It’s like, “This is screwed up. Full stop.” So, I mean, part of why we’re here is to talk about some of those kinds of debt. And let’s just start with one technical debt. And I’m saying technical in the sense of the way that you perhaps use software to put together your content. Let’s kind of focus on the content operations world.

BS: Mm-hmm.

AP: For example, if you are delivering content via unstructured desktop authoring tools, of which there are many, and you can templatize things and make your content seem more consistent, but there’s a problem with a lot of desktop published or content generated from the desktop publishing world. It’s more focused on look and feel and fit and finish, particularly if you’re delivering for PDF. And yes, people are still doing that. So there’s a lot of time and effort spent on that look and feel, that fit and finish. And frankly, that time should have been invested in adding intelligence to the content to explain, you know, things under the covers about what user is this for? What is the model of this particular thing? All of that kind of metadata, that kind of categorization. Desktop publishing, at least from my point of view, doesn’t really do a great job of helping you catalog that kind of stuff. So that’s a problem.

BS: No. It is a problem. Also, with desktop publishing, you can kind of confuse AI a bit if you’re using desktop publishing inconsistently. So if you’re using formatting tools to override formatting to make things look like headings or make certain paragraphs look like children of another paragraph, doing those manual finesses is great for print because, as you know, most people will look at that and understand the hierarchy of information, understand what’s going on with the content that they’re reading. But anything digital isn’t necessarily going to pick that up. 

AP: Right. Yeah.

BS: Especially if you’re looking at something like bare bones HTML, if you’re using CSS to override the size and prominence of a standard paragraph as opposed to using a heading, that is not necessarily going to be picked up as a heading, even though a reader would actually see that as a heading while looking at the HTML page.

AP: Right. What a human reader can figure out from formatting cues, if those cues aren’t set up in a way that a large language model, a computer, can understand, there’s that huge disconnect, and that’s where that debt starts piling up. And then another angle here in this content creation world, if you are using multiple different tools to create your content, there’s a good chance that content under the covers is not going to be processed the same by a large language model. So there’s another deficiency right there on top of that. And this is very common, for example, if you have had mergers, acquisitions, and you’ve basically created a larger company from many different companies. And they all still have, especially in legacy content, things created the “old way,” and all the old ways start to pile up and cause problems because your large language model can’t properly basically figure out what is the heading in this particular chunk of docs versus what it is over here. So it can’t parse it as well. And again, this all goes all the way back to the way that you created that content. And it’s a clue, hey, you need you need to fix this. And I I think beyond that more technical, the way that you create it, there also there are also issues with the content itself.

BS: Mm-hmm.

AP: Is it updated? Does it reflect the latest information? Is it created for all the different locales that your company serves? Is it in different languages? That is another pile of debt that when you start looking at AI, it’s gonna be all the problems in regard to that are also gonna be just very brutally magnified, and you’re going to have to address them to really have a large language model that works at all.

BS: Yeah, because not only do you have the technical debt on the source content side and on the published content side, but you also now have technical debt growing on the AI side because you need to spend more time and energy refining the how that model works with your content in order to achieve the correct results.

AP: Right. So you’re having to do a lot of overrides and we I don’t know if overrides is the right word, but you’re having to do a lot of additional processing and figuring out so it will parse things correctly. And there are a lot of companies today that still have problems keeping content updated to the latest and greatest. So unfortunately, people turn around and call support. Or today they start hitting up the chatbot. But guess what? 

BS: Mm-hmm.

AP: If the chatbot doesn’t have access to the latest and greatest because frankly it doesn’t exist or it’s not hooked up to it. It’s it’s just like the poor people in support. It’s not going to know what to do and it’s going to spit out probably very authoritatively wrong outdated information. Again, yeah, it’s be it’s goes all the way back to

BS: That’s yeah.

AP: In your content creation process, how are you accounting for updates? How quickly are you getting them in place? How are you handling them in both your source language and how are you handling them in your other languages? It’s one thing I do want to bring up here, and and and I may be biased here, but I don’t think localization is getting enough attention on the AI distribution side. It’s been talked about for a very long time in regard to machine translation, AI assisted translation. 

BS: Mm-hmm.

AP: But I don’t just like I think sometimes localization is a second thought for a lot of companies, which still blows my mind in 2026. And by the way, we’re recording this in July 2026. So what we say right now may be outdated next month. Who knows? 

BS: Who knows?

AP: There is that issue. I’m curious, you have a more of a localization background than I do, but I do see that being a potential debt problem, part of this debt crisis that we’re talking about here.

BS: Definitely, because if you’re pushing your content out to AI, do you have targeted audiences in mind? Or is it going to be a free-for-all of people going in and using that AI to get answers to their questions? if you are localizing your content, I think probably the best practice here would be to somehow bundle for consumption all of the guides in all of the different languages together. So you have product XYZ and you have it translated in three languages. Then you put all of those copies together and give that to the AI so it can draw its associations as well as it builds out you know its understanding or I hate to use understanding because the AI doesn’t understand things. It’s very easy to make slips in that way. But so that AI can actually draw those relationships. So, you know, if you’re looking in English or you’re looking in Spanish or you’re looking in German, the same query in those three languages will retrieve pretty much the same result. That’s proper for that language. And we talked with someone last year on the podcast, Steve Maule from Acclaro, about AI in translation. 

AP: Yeah.

BS: And his focus was more on using AI as not so much a generative tool, but you know, a tool for aiding in AI, basically the next step from machine translation to neural machine translation to AI translation. And kind of the benefits and drawbacks at the time. Again, that was a year ago. So many of these things have probably changed again. But if you are translating content you may want to go back to that podcast and take a listen to what he had to say. and we’ll put a link to that in the show notes for this one.

AP: Yeah, again, I keep going back to this idea that we talked about at the beginning, how delivering via AI just puts this magnifying glass to what you’re doing and uncovers things that are wrong. But as you had mentioned, it doesn’t necessarily offer advice on how to fix it. And we will talk about that probably to wrap up the podcast. But what I also want to mention too. More and more, there’s actual real debt involved, not just the more indirect debt we’re talking about with the content debt, the technical debt. The AI companies, the providers, at one time had very generous offerings as far as the amount you could hit their APIs, the amount of tokens that they offered, whatever else. 

BS: Definitely.

AP: But now, the days of I guess you could say subsidized token use, they’re pretty much over. And now the actual real cost is being passed on to the people using these large language models. And as a result, the actual price, the cost of using AI is skyrocketing in all these organizations. And there have been a lot of news reports lately about very

BS: Mm-hmm.

AP: And there have been a lot of news reports lately about very big firms, and I will not name names, but you know them all, their household names, basically having to do a 180 and tell their employees, hey, you need to chill a little bit on your AI use because you’re burning up the tokens, converting PDF files to PowerPoint. And yeah, I get it.

BS: Mm-hmm.

AP: Doing that conversion is probably not the most efficient use, productive use of AI tokens. But this gets into kind of, I don’t know if there’s such a thing as cultural debt, but if your company is sending this message, and a lot of companies are, use AI for everything. And you know, they’ve got these leaderboards up showing these people use this much AI this week, you know, that sort of thing. You are encouraging a free-for-all. So, instead of saying, here are the good uses, this is where we think you need to focus your use of AI in coding, in content creation, in whatever else, this whole idea of just use it. You got to use it a lot. And now that’s coming back and biting people in the backside. So companies are scrambling and telling people, calm down.

BS: Mm-hmm.

AP: Chill out. Don’t use the AI as much. So it, there’s a mixed message there. And it’s because there was not good communication, nor was there good thought put into how workflows should incorporate AI. And I think that cultural thing right now is a huge problem. And everybody’s focus on the cost and the non-subsidized token use when they need to be going back and looking at, well, look at your comms. Look at what you were telling people six, twelve months ago about AI. It’s kinda on you.

BS: Mm-hmm. Yeah, basically don’t let your nine-year-old run around in a toy store with your credit card.

AP: Yeah, pretty much.

BS: That’s pretty much what’s been happening. And yeah, we’ve been seeing lots of reports. There was one in Forbes last week, where, you know, companies are saying that the cost of using the compute power is more expensive now than the people it was supposed to augment.

AP: Exactly. So I think the reality of what AI can do maybe is starting to sink in, maybe a little, but at least exposing people to the true cost of it on a corporate level, it may

BS: Mm-hmm.

AP: I’m probably being too positive here. What? Me being positive? It may result in some recalibration about how companies are telling people to use AI and maybe thinking a little more on a cultural level, all the way starting with communication. Here are the workflows you need to apply AI to. This is where you can apply it.

BS: Mm-hmm.

AP: And then there’s also the whole vibe coding discussion. I don’t know how much we want to get into it, but we all know, right, there is now an industry, you know, people are now going in to clean up vibe coding because, yeah, just because you can vibe code shouldn’t mean doesn’t mean you should be vibe coding. Because who was going to maintain and clean up that mess? So again.

BS: Well, we’re cleaning up a lot of vibe coding now.

AP: It’s another cultural issue that should have been addressed, but in this AI rush, everybody was like, use it, use it, use it, without thinking about how they should really how they should really apply it and basically sort of reimagine the way they do work in a productive way. And that does not mean let the AI do everything from my point of view.

BS: Mm-hmm.

AP: So I guess we probably need to kind of wrap up and talk about what people can do when all of their ugliness is amplified and magnified by AI. And it really it’s about rewinding and going back and looking at from the start, especially in the content world and the content operations world, how are you creating that content? Are you doing it in a consistent way? Are you building in intelligence into your content that an LLM can pick up to better understand the context of that content you’re providing it? Is your localization workflow efficient? Is it getting content turned around quickly so people in other markets are not six or eight months behind in the latest and greatest information. And by the way, yes, that still happens today, believe it or not, that people are months out and getting it because they are in another location. It’s shameful, but it still happens. It does.

BS: Mm-hmm. Yeah. And there’s also, you know, how are you feeding your content over to AI? How are you providing it? Are you just having the AI team you know scrape web pages and PDFs, or are you supplying something that’s a little bit more targeted and tied to the content itself, and maybe supplies some of that intelligence over?

AP: Yeah, because I think Sarah’s talked about this on a previous podcast. A company was noticing that their LLM was kind of not using PDFs or not weighing them, or and I’m again I’m personifying big time here. my apologies.

BS: Yeah. It’s easy.

AP: Was yeah, was not really it was not giving the weight to the content in PDFs. And when the company kind of did some reverse engineering, they realized it was because that PDF what they didn’t have a lot of context. 

BS: Mm-hmm.

AP: There wasn’t a lot of basically metadata built in that the LLM could basically parse. So it’s like, I’m gonna kind of put that to the side because it doesn’t have the richness that I need to do things well. So it it’s about looking at how you write. How are you delivering this content? And one possibility here, and it’s, well, there’s several structured content where you have metadata built into your content. You have tagging that offers semantic context.

BS: Mm-hmm.

AP: That’s one way to do it. And it’s not the only way. I mean, we have a lot of clients who use it, but it absolutely is not the only way to do it. Knowledge graphs can also be part of this solution. And sometimes knowledge graphs and structured content can play together to provide that rich feed of information that LLMs prefer and can do a better job with. So it’s not a one-size-fits-all solution when it comes to how to create that content or how to best hand it over to AI, but I think it is kind of a one-size-fits-all that if you aren’t doing things right foundationally, AI is going to kick your tail.

BS: Mm-hmm. Pretty much, yeah. And I think the best way to look at it is to consider AI another delivery target, just like you would a portal, just like you would a PDF or what have you, a help system. When you consider it as another endpoint for your content, another, you know, place to deliver to, it makes it a lot easier to start scoping what you need to do to reach, the requirements for that particular target.

AP: Agreed. So content people look at it as a delivery point and also look at as a job aid too to help enforce style guides, to help enforce taxonomy, whatever else. So it’s not just about content creation. 

BS: Mm-hmm.

AP: It’s also that part of your thinking needs to be a content distribution point, like Bill mentioned. And it can be hard to think of it as both of those things, but in the content world, it absolutely is.

BS: And I think that’s a good place to leave it. So thank you, Alan.

AP: Thanks, Bill.

BS: And we’ll see you on the next one.

Get more insights on AI and content operations in our monthly newsletter!

var gform;gform||(document.addEventListener("gform_main_scripts_loaded",function(){gform.scriptsLoaded=!0}),document.addEventListener("gform/theme/scripts_loaded",function(){gform.themeScriptsLoaded=!0}),window.addEventListener("DOMContentLoaded",function(){gform.domLoaded=!0}),gform={domLoaded:!1,scriptsLoaded:!1,themeScriptsLoaded:!1,isFormEditor:()=>"function"==typeof InitializeEditor,callIfLoaded:function(o){return!(!gform.domLoaded||!gform.scriptsLoaded||!gform.themeScriptsLoaded&&!gform.isFormEditor()||(gform.isFormEditor()&&console.warn("The use of gform.initializeOnLoaded() is deprecated in the form editor context and will be removed in Gravity Forms 3.1."),o(),0))},initializeOnLoaded:function(o){gform.callIfLoaded(o)||(document.addEventListener("gform_main_scripts_loaded",()=>{gform.scriptsLoaded=!0,gform.callIfLoaded(o)}),document.addEventListener("gform/theme/scripts_loaded",()=>{gform.themeScriptsLoaded=!0,gform.callIfLoaded(o)}),window.addEventListener("DOMContentLoaded",()=>{gform.domLoaded=!0,gform.callIfLoaded(o)}))},hooks:{action:{},filter:{}},addAction:function(o,r,e,t){gform.addHook("action",o,r,e,t)},addFilter:function(o,r,e,t){gform.addHook("filter",o,r,e,t)},doAction:function(o){gform.doHook("action",o,arguments)},applyFilters:function(o){return gform.doHook("filter",o,arguments)},removeAction:function(o,r){gform.removeHook("action",o,r)},removeFilter:function(o,r,e){gform.removeHook("filter",o,r,e)},addHook:function(o,r,e,t,n){null==gform.hooks[o][r]&&(gform.hooks[o][r]=[]);var d=gform.hooks[o][r];null==n&&(n=r+"_"+d.length),gform.hooks[o][r].push({tag:n,callable:e,priority:t=null==t?10:t})},doHook:function(r,o,e){var t;if(e=Array.prototype.slice.call(e,1),null!=gform.hooks[r][o]&&((o=gform.hooks[r][o]).sort(function(o,r){return o.priority-r.priority}),o.forEach(function(o){"function"!=typeof(t=o.callable)&&(t=window[t]),"action"==r?t.apply(null,e):e[0]=t.apply(null,e)})),"filter"==r)return e[0]},removeHook:function(o,r,t,n){var e;null!=gform.hooks[o][r]&&(e=(e=gform.hooks[o][r]).filter(function(o,r,e){return!!(null!=n&&n!=o.tag||null!=t&&t!=o.priority)}),gform.hooks[o][r]=e)}});







Name*



First




Last




Email*



Data collection*

I consent to my submitted data being collected and stored.


Review our privacy policy

Consent to subscribe*

I consent to the use of my submitted data for marketing emails. I understand that I can unsubscribe at any time.


Review our privacy policy




















gform.initializeOnLoaded( function() {gformInitSpinner( 2, 'https://www.scriptorium.com/wp-content/plugins/gravityforms/images/spinner.svg', true );jQuery('#gform_ajax_frame_2').on('load',function(){var contents = jQuery(this).contents().find('*').html();var is_postback = contents.indexOf('GF_AJAX_POSTBACK') >= 0;if(!is_postback){return;}var form_content = jQuery(this).contents().find('#gform_wrapper_2');var is_confirmation = jQuery(this).contents().find('#gform_confirmation_wrapper_2').length > 0;var is_redirect = contents.indexOf('gformRedirect(){') >= 0;var is_form = form_content.length > 0 && ! is_redirect && ! is_confirmation;var mt = parseInt(jQuery('html').css('margin-top'), 10) + parseInt(jQuery('body').css('margin-top'), 10) + 100;if(is_form){jQuery('#gform_wrapper_2').html(form_content.html());if(form_content.hasClass('gform_validation_error')){jQuery('#gform_wrapper_2').addClass('gform_validation_error');} else {jQuery('#gform_wrapper_2').removeClass('gform_validation_error');}setTimeout( function() { /* delay the scroll by 50 milliseconds to fix a bug in chrome */ }, 50 );if(window['gformInitDatepicker']) {gformInitDatepicker();}if(window['gformInitPriceFields']) {gformInitPriceFields();}var current_page = jQuery('#gform_source_page_number_2').val();gformInitSpinner( 2, 'https://www.scriptorium.com/wp-content/plugins/gravityforms/images/spinner.svg', true );jQuery(document).trigger('gform_page_loaded', [2, current_page]);window['gf_submitting_2'] = false;}else if(!is_redirect){var confirmation_content = jQuery(this).contents().find('.GF_AJAX_POSTBACK').html();if(!confirmation_content){confirmation_content = contents;}jQuery('#gform_wrapper_2').replaceWith(confirmation_content);jQuery(document).trigger('gform_confirmation_loaded', [2]);window['gf_submitting_2'] = false;wp.a11y.speak(jQuery('#gform_confirmation_message_2').text());}else{jQuery('#gform_2').append(contents);if(window['gformRedirect']) {gformRedirect();}}jQuery(document).trigger("gform_pre_post_render", [{ formId: "2", currentPage: "current_page", abort: function() { this.preventDefault(); } }]); if (event && event.defaultPrevented) { return; } const gformWrapperDiv = document.getElementById( "gform_wrapper_2" ); if ( gformWrapperDiv ) { const visibilitySpan = document.createElement( "span" ); visibilitySpan.id = "gform_visibility_test_2"; gformWrapperDiv.insertAdjacentElement( "afterend", visibilitySpan ); } const visibilityTestDiv = document.getElementById( "gform_visibility_test_2" ); let postRenderFired = false; function triggerPostRender() { if ( postRenderFired ) { return; } postRenderFired = true; gform.core.triggerPostRenderEvents( 2, current_page ); if ( visibilityTestDiv ) { visibilityTestDiv.parentNode.removeChild( visibilityTestDiv ); } } function debounce( func, wait, immediate ) { var timeout; return function() { var context = this, args = arguments; var later = function() { timeout = null; if ( !immediate ) func.apply( context, args ); }; var callNow = immediate && !timeout; clearTimeout( timeout ); timeout = setTimeout( later, wait ); if ( callNow ) func.apply( context, args ); }; } const debouncedTriggerPostRender = debounce( function() { triggerPostRender(); }, 200 ); if ( visibilityTestDiv && visibilityTestDiv.offsetParent === null ) { const observer = new MutationObserver( ( mutations ) => { mutations.forEach( ( mutation ) => { if ( mutation.type === 'attributes' && visibilityTestDiv.offsetParent !== null ) { debouncedTriggerPostRender(); observer.disconnect(); } }); }); observer.observe( document.body, { attributes: true, childList: false, subtree: true, attributeFilter: [ 'style', 'class' ], }); } else { triggerPostRender(); } } );} );

The post The debt crisis: AI edition appeared first on Scriptorium.

There are five levels of maturity for AI-driven content operations. Which level are you in? In this episode, Sarah O’Keefe and Bill Swallow walk through the AI content ops maturity model, from ad hoc experimentation to fully autonomous workflows.

Sarah O’Keefe: We want this automation, right? We want the ability to go in and extract release notes and do something with them. We have to have a certain level of maturity on the software development process so that we can grab the appropriate information. The same thing is true on the content side. You have to have a certain level of maturity in your content development processes, in your content management, so that you can identify the right things to process and the right things to access.

Related links:

Want to know more about Sarah’s nifty little side project? Register for our upcoming webinar.

AI in the content lifecycle

Enterprise content strategy maturity model

LinkedIn:

Sarah O’Keefe

Bill Swallow

Transcript:

Introduction with ambient background music

Christine Cuellar: From Scriptorium, this is Content Operations, a show that delivers industry-leading insights for global organizations.

Bill Swallow: In the end, you have a unified experience so that people aren’t relearning how to engage with your content in every context you produce it.

Sarah O’Keefe: Change is perceived as being risky; you have to convince me that making the change is less risky than not making the change.

Alan Pringle: And at some point, you are going to have tools, technology, and processes that no longer support your needs, so if you think about that ahead of time, you’re going to be much better off.

End of introduction

Bill Swallow: I am Bill Swallow.

Sarah O’Keefe: And I’m Sarah O’Keefe.

BS: And today we’re going to talk about AI in content operations, or more specifically, a maturity model for AI.

SO: Everything needs a maturity model, even AI.

BS: Even me.

SO: I have no comment.

BS: My maturity model is written in crayon, what can I say? So okay, so we need a maturity model for AI as far as content operations are concerned, and probably in you know, many different degrees, but we’ll focus on content operations. So what might that look like?

SO: I’ve been thinking about this and what it looks like to employ AI as a tool to help you with content. And as I was thinking about what this looks like, you know, you always fall back on that standard five-step model where one is basically mass chaos, and five is the perfect world, generally. also, one is nearly always cheap, and five is nearly always expensive enterprise things. But, you know, let’s go a little beyond mass chaos versus governed, regulated, etc., and sort of sort of back up a little bit and talk about what this might look like. So level one in every maturity model typically is ad hoc. And what that means is that in this case, AI is being used sporadically by some people. It’s inconsistent. And I would say that when we look at AI and content specifically,

BS: Mm-hmm.

SO: This is going to be things like reprocessing your content using public-facing models. So I wrote a draft of something, I shove it into ChatGPT and I ask it to shorten it or tighten it up or identify areas that are problematic. Or I just say, hey, you know, write my article for me. The outcome that you’re gonna get on an ad hoc model is going to depend on an ad hoc level one.

BS: Mm-hmm.

SO: AI thing is going to depend on how good you are on the individual’s expertise and their level of interest. So if you want to just go in there and say, hey, I have a bio and it’s too long, and I’ve been asked to produce one that’s only 50 words for a particular conference, for example, then, you know, this is this is actually a really good example of ad hoc, right? 

BS: Mm-hmm.

SO: We have these long multi-paragraph bios and every conference I’ve been to has a different requirement for how that bio needs to be shaped. And the fastest way to success is to just shove it into a chatbot and say, give me a 50-word version. And and then read it and make sure it didn’t invent things or give you a PhD or anything like that, and then ship it off to the conference organizer. But this is much, much faster than rewriting it from scratch by hand. And also I think, it’s a good example of something where I have the extended version and I’m going to summarize down. And that usually works pretty well. So level one is ad hoc. It’s kind of sporadic. There’s no standard across the organization. It’s just me saying, this looks useful, or you’ve probably got some use cases in this space as well.

BS: Right. So it it kind of aligns with I guess level one of the the content maturity model that we talked about a while back, where level one is is simply content exists. Could be, you know, someone typing stuff up in Word or, you know, using a myriad of different tools, no style guide, just kind of getting content out there because people need it.

SO: Yep. So level two is tactical. And tactical is sort of like we’re using this tool to solve some specific problems. And what you’re going to see here is something like that Bill has invented some nifty time saving tool and he has shared it with other people. Or in a larger organization, maybe somebody invented a nifty validate or or something like that and they’ve rolled it out across maybe the department, probably not the entire organization. Aomething like AI support is being rolled out. Maybe the organization has created a chatbot internally for customers, right? So there’s a chatbot, it’s sitting on the company website, people can use it to get answers, but it’s really bad. And the reason it’s really bad is because nobody thought too carefully about the content going into the chatbot, because again, we’re tactical. So probably this looked like the AI team just raided the local SharePoint, grabbed a bunch of content, did not pay a whole lot of attention to the question of whether this content was up to date in release status. Those things don’t exist, right? It’s just, look, a bucket of PDFs. Cool. Let’s dump them into the AI and go for it.

BS: Mm-hmm.

SO: And tragically, in many cases, the techcomm team is sitting on rigorous, structured, vetted, approved content, and nobody remembered to go ask them, can we have your content? Or where is your official content source? Or how do I know what version belongs with which document?

BS: Right, because you know, in in their point of view there’s a PDF of it, so I don’t need to ask them.

SO: Yeah, it’s it’s just PDF. How hard could it be? So something like AI was support was rolled out, but nobody really thought about it. Maybe it’s at a departmental level, probably it’s not enterprise-wide. And nobody has really thought about connecting this AI thing to the assets inside the organization in a reasonable, rigorous, governed, organized kind of manner.

BS: And I suppose that’s where you get to the next tier.

SO: Right. So the next tier after tactical comes strategic, right? So we have an actual strategy. Now, one of the difficulties in talking about AI is that AI is a tool and it’s kind of like talking about electricity. You can apply it to lots of places and it’s more sensible in some places than others. But when we say what’s your AI strategy, like how do you use water? I mean, come on, and the answer is of course to drive the AI and you know destroy the environment. But there are things that you can do with AI that are useful for content. There are also things that you can do that are not. So if you have a strategic approach to this, a strategic approach to use of AI, backing up to the authors again rather than the delivery side, maybe this looks like a collection of prompts that have been built that are shared.

BS: Not with electricity.

SO: Maybe this looks like saying this is the workflow that you employ. These are the kinds of things that we do to actually test whether this thing is working. these are the metrics that we’re following. So there’s an actual overarching bigger picture that somebody’s thinking about that goes beyond, let me go shove this into chatbot of the day.

BS: Mm-hmm. Right, right.

SO: So there’s an actual strategy for the public-facing chatbots. Somebody has thought about the back end. The authors have useful AI tools that add to their you know their productivity. One of the things that I’m hearing a lot now, you know, low-hanging fruit, release notes. Nobody wants to write release notes. It’s a terrible drudge task. It’s and it needs to be done. Well,

BS: Mm-hmm.

SO: There’s now there are now a lot of solutions that look like look at the diff in the code, look at the delta from you know version one to version one dot one, find the diff in the code, find the changes that have been made, look at the JIRA tickets that have been addressed, that have been solved in release one dot one, and then consolidate that all into a set of release notes that say, here’s what’s been done. And that’s probably 90% of the work, and the last 10% of the work is read that and make sure it’s accurate. Right? Don’t please don’t skip that step. Like actually look at what the thing is generating. Now, what’s interesting to me about level three, this sort of more strategic approach, is that what you’re gonna start to see is that you have prerequisites for this. You can’t do this. 

BS: Yes.

SO: So release notes are good example. Let’s say that hypothetically, and this is gonna sound insane, but let’s say that hypothetically, you have software development and you have no source control.

BS: Hmm.

SO: Everybody’s screaming, right? Because this is nuts, and why would you ever do this? Okay. But hypothetically, you have no source control. Okay. How do you know what’s changed between version one and version one point one?

BS: It’s up here in my head.

SO: Excellent, great. Ha okay, cool. so I’m gonna need to connect the AI to your head so that we can pull those changes out of your head.

BS: That sounds fun.

SO: Yeah. Amazing. Right. So all of a sudden, because we want this automation, right? We want the ability to go in and extract release notes and do something with them. We have to have a certain level of maturity on the software development process so that we can grab the appropriate information. Now, the same thing is of course true on the content side. You have to have a certain level of maturity in your content development processes, in your content management, so that you can identify, you know, the right things to process and the right things to access. And why it is that, you know, we know that software has to be governed, but we’re not so sure about content is a mystery to me.

BS: I had never understood that.

SO: Yeah. So there we are. Okay, so that’s kind of like a level three. With there’s some sort of strategy emerging across the enterprise. There are some useful tools and they’re shared. This is kind of like in content when you start thinking about templates. We’re gonna have some templates and we’re gonna give them to people and they’re gonna use them and it’s gonna be great. All right, so level four is governed, managed. And so now good understanding of AI.

BS: Mm-hmm.

SO: It is being applied in a useful, intelligent manner, by which I mean don’t apply it to the wrong problem sets, right? Apply it to the things where it makes sense to apply it. Thinking about governance, thinking about metrics, thinking about success. And then your data sources and your content sources are being managed in such a way that the AI gets good input and can actually generate good output. So I’m actually not a big fan of the term human in the loop because human in the loop implies that the AI is doing all the work and then like the human eventually gets around to QAing it. You know what? We’re terrible at QA. You know who’s good at QA?

BS: Mm-hmm. AI.

SO: No, computers, not AI. AI is all about probability and whatever. It is actually not very good at QA. What’s good at QA is traditional software, right? One plus one is always two. In an AI, one plus one, sometimes it’s not two. So you manage that stuff and you put those guardrails up and you start putting up the guardrails that say, okay, when the AI kind of wanders off into the wilderness, we’re gonna like bring it back to reality. We’re gonna have, we’re gonna put it in a box, right? Make the AI think inside the box, and we’re gonna govern what that box is. AI is great at thinking outside the box. Unfortunately, that’s usually not what we want from technical content. So it needs to be in in the box and it needs to be consistent and needs to be managed and all the rest of it. 

BS: Mm-hmm.

SO: So we govern it, right? We go in there and we make sure that the processes and the tooling that’s being put in place and the automation that’s being put in place where we’re leveraging or using AI to do things is managed. And so the human in the loop thing. I don’t want the human in the loop to fix things on the back end. I want the human in the loop to fix things on the front end so that what goes in is better, so that there’s less work to do when it comes out. You know, fix it beforehand. Don’t remediate it afterwards. That’s a boatload of work and it is not fun. So fix it ahead of time.

BS: Right. Yeah. And likewise you probably wanna have, you know, some guardrails in there so that, you know, your AI, whatever it is, doesn’t go playing around with content that has been approved and released and is not slated for updating.

SO: Yeah, you know, don’t fix that. That one’s done. That one’s and you know, we’re not even talking here about what it means to be in a regulated industry or in a regulatory environment. there if you are shipping or sorry, if you are a large organization and you are doing things in Europe, then you are likely subject to the European, the EU AI Act.

BS: That’s a completely different beast.

SO: And you have to think about what that means for what you’re doing, because the fun gold rush wild, wild west strategy of just throw AI at everything is not gonna fly in Europe. Okay, so that’s governed, you know, hypothetically. And then level five is agentic, which is basically that everything, everything or a lot of it is running autonomously.

BS: Mm-hmm.

SO: You know, the layman’s explanation of what is agentic AI, the difference is that instead of saying I need to put a prompt into the chatbot, it does it itself because you’ve built out the systems that drive all of that happening. 

BS: It understands what needs to happen at what point in time.

SO: Well, let’s not say understands, but yes. I’m trying so hard. 

BS: Well, yeah, not understands, but there’s a workflow in place that the AI is following.

SO: And it’s so difficult. I think that, you know, there’s, as a side note, the why do we think, why do we impose personality on the chatbots? And the answer is I think that psychologically it’s very, very difficult to interact with something that play acts at human interaction. What a great idea! Good for you. I love your thinking, blah, blah, blah. So it makes you think you’re interacting with a human. And I don’t think that our brains are equipped to say, no, actually, this is a machine.

BS: It’s like a scary version of Teddy Ruxpin.

SO: It, well, it passes the Turing test. And so we just can’t separate if it feels like you’re interacting with a person, you know you’re not, but it feels as though you are, and feeling is always gonna win over knowledge. So

BS: Mm-hmm. Well yeah, the interaction is a lot more organic than you get from, you know, traditional tools.

SO: Or, you know, yeah. I mean, think about the difference between a search, typing in a search string, and you know, a conversational search, a conversational interface. It’s quite, quite troubling, actually. Yeah, so this is kind of the five-level model, right? From big mess ad hoc, some things are happening, th some things aren’t, up to it’s completely autonomous. Now, if it’s going to be the more autonomy you want, the better your inputs have to be, which circles us right back to, and therefore, you have to do the work on the content side, because if you don’t do the work on the content side, the AI is going to go off the rails in interesting, unexpected, and potentially disastrous ways.

BS: It will play with the mess you leave it.

SO: Yep. So that’s where we’re going with this. That’s the AI content ops maturity model as it stands today. I reserve the right to change it tomorrow.

BS: Today. Of course.So this model came out of I guess some little nifty side project you’ve been working on recently.

SO: I am working on a nifty little side project. We’re not quite ready to announce it. but I’ve got a a co-author and we’re working on a thing.

BS: Fair.

SO: I could say more but then I’d, you know, be in trouble.

BS: When might you be able to say more?

SO: I believe that we have a webinar coming July 22nd, where we will say some more things.

BS: Alrighty. Well we will learn more things then.

SO: I too will learn more things and probably we’ll have to we’ll probably we’ll have to change everything we’ve done up until this point because everything will change by then.

BS: Of course. Guess that’s a good place to leave this podcast. Thank you, Sarah.

SO: Thank you.

Want to know more about Sarah’s nifty little side project?

Register for our upcoming webinar.

The post From ad hoc to autonomous: The AI content ops maturity model appeared first on Scriptorium.

How do you really choose the right documentation tool? In this podcast episode, Sarah O’Keefe (Scriptorium) talks with Paweł Kowaluk and Michał Skowron (Guidewire Software) about building a successful tool selection process, the realities of docs as code, and what happens when the technology becomes the unpredictable variable.

Paweł Kowaluk: It’s funny how programming used to be deterministic, and it was the people who were messy. We always knew that people are going to be whimsical and maybe harder to rein in, but the technology is going to be predictable. Whereas now, technology is not predictable anymore, and you give it a prompt and you hope it’s going to do what you want. You adjust the system prompts and change the weight of things which are retrieved versus metadata, et cetera, and it doesn’t always work the way you expect it to.

Sarah O’Keefe: And now the people are being asked to be the deterministic layer, right? To be the QA on top of the AI.

Paweł Kowaluk: That’s actually very insightful. I like that. That is true. The human in the loop or whatever you call it, that’s supposed to be the voice of reason.

Related links:

Scriptorium: AI in the content lifecycle

Tech Writer Koduje podcast

Tech Writer Koduje: DITA as code – a modern approach to the classic standard

Tech Writer Koduje: Are people abandoning docs as code?

Tech Writer Koduje: A tech writing CCMS can also be a broken promise

LinkedIn:

Host: Sarah O’Keefe

Guest: Paweł Kowaluk

Guest: Michał Skowron

Tech Writer Koduje LinkedIn profile

Transcript:

Introduction with ambient background music

Christine Cuellar: From Scriptorium, this is Content Operations, a show that delivers industry-leading insights for global organizations.

Bill Swallow: In the end, you have a unified experience so that people aren’t relearning how to engage with your content in every context you produce it.

Sarah O’Keefe: Change is perceived as being risky; you have to convince me that making the change is less risky than not making the change.

Alan Pringle: And at some point, you are going to have tools, technology, and processes that no longer support your needs, so if you think about that ahead of time, you’re going to be much better off.

End of introduction

Sarah O’Keefe: Hey, everyone. I’m Sarah O’Keefe, and welcome to the podcast. In this episode, we are going to talk about tool selection with a couple of special guests. With me today are Paweł Kowaluk, who is a software architect at Guidewire Software, and Michał Skowron, who is a documentation tools developer, also at Guidewire. Both of them are based in Poland. Welcome.

Paweł Kowaluk: Hi.

Michał Skowron: Hello.

SO: I am glad to have you. For those of you on this podcast that speak Polish, you’re probably already aware that they have the one and only techcomm podcast in Polish that is available out there, and Michał and Paweł are also experts on doc process and tool selection, so that’s what we wanted to focus on today. So I will start and throw it to Michał and ask you the big picture question, which is what does a good tool selection process actually look like?

MS: For me, good selection tool process would be divided in three stages. The first one would be gathering requirements, looking what’s out there, defining what you want to basically achieve with this new tool. Then I would go to a pilot project where you can actually test the selected tool in the real world. Manufacturers and producers of software will tell you that it can do anything and it will promise that, “Okay, you can meet all your requirements easily and we can fix that, we can improve that, we can adjust that,” so everything can be done is usually what we hear, but then you want to test it in real world on a real project, so that will be a pilot project for you and your team.

And the third phase that depends on the outcome of the second phase, which is you either productize the selected solution or you just say, “Okay, that was a bad choice and we don’t need that.” Then we need to go back to the first stage and then say, “Okay, we need to select another tool,” and again, requirements, et cetera, et cetera. So for me, that’s the whole process, and the first stage would be probably the longest one because you need to make sure that you are meeting all your goals.

SO: So what’s the most common reason that a pilot doesn’t succeed, that you have to go back and say, “That didn’t work. We have to try something different”?

MS: It’s usually because you didn’t see everything when you were planning. For example, you have some projects that are very specific or you didn’t see all the problems or things that are coming your way. It’s hard to say exactly what the reason is, but it can be multiple reasons.

For example, using of, I don’t know, branching, let’s say, in a specific tool. When you have multiple versions of your product and you want to keep them separate when it comes to documentation, it can turn out that the feature says, “Okay, you can use branching and then you can do it easily,” and then you start using it and it turns out that it doesn’t work the way you expect it. This is actually a real life example because we had a system that… I’m not going to mention any names or anything like that, but there was a system and they promised us… That was years ago and it was a vendor that promised us that they’re going to introduce a feature called branching, and it turned out after they did that that it wasn’t what we expected. So it can turn out in many cases, in many ways, it can be the problem, but branching is just an example, but it can be many other things that can go wrong.

PK: Hey, if I can jump in here, I got a couple of examples. One is I could call it releasing strategy or versioning strategy overall, which is very hard to test in a pilot project. It’s very hard to scope for requirements because the little problems come out after a while, after a year of publishing, after two years of publishing. And another example which is related is reuse, and this one is down to formulating the requirements correctly. Because I think just saying, “We want to reuse something,” is not enough, because you have to say exactly what you want to reuse and how you want to reuse it and what you want the result to be.

So for example, if you say, “I want to reuse notes and warnings and things like that.” We sometimes call them admonition. So, “I want to reuse these notes in my docs, and if I update a note, I want it to update in every published version of the doc.” Then only if you have these details, like I want to update it once and then want it to automatically update and publish docs, then you will see it’s not working the way you expect it. Because if you just say, “I want to reuse notes,” every system can reuse notes. Even in docs as code, there’s scripts and macros that allow you to reuse notes.

MS: It can be also another thing that, for example, you compare the benefits with the actual cost of implementation, and it can turn out it’s not worth it because people are, for example, reluctant to use your new tool. The training, the cost of licensing, the cost of support is too big, and then you realize, okay, we want to achieve a goal like, I don’t know, reduce the time to market, and then it turns out it doesn’t work because people are struggling with using the product on a real project. And on paper, everything looks cool and you have all these features, you can use them, like Paweł mentioned, for example, reuse conditional formatting and things like that. And then it turns out it’s very hard to use, it doesn’t serve its purpose, and then you have to pay for every additional stuff and people don’t want to use it, so what’s the point?

SO: Yeah. And I think we find that the technical problems in general, if you’ve done your requirements work, then usually, the technical problems are solvable if the people engaged in the project want to solve them, and that’s where you run into the change management issues that you’re talking about, that if the team that is being asked to pilot, to test, to try things out is sufficiently disinterested in making it succeed, they will find a way to make it fail. And the reverse is true as well. If they want it to succeed, you can implement tools that are … Well, all tools are imperfect, but you can implement tools that are not perfect solutions and succeed if the team is behind you, and if they’re not, bad, bad things will happen.

PK: Oh yeah, that’s true. And I’ve been on projects where we did not do proper change management and I’ve been on projects where we really did it well. If you start early and you involve people, like get the biggest troublemakers, people who are the most opposed to any change, get them on the team, and if you can convince them, that means, one, you are making the right choice because you’re convincing people who are skeptical, and then two, you are set up for success. These people are going to be your biggest champions of the new solution.

MS: But it’s good that you mentioned it because I think it’s worth emphasizing that the goal of the pilot project is not to succeed. That’s not the actual goal. The goal is to verify the requirements against the real project, and so the failure is also a success to some extent. It’s not like you have to do everything to prove that the selected tool is the right one. No, you should be aware that the pilot can end up with your let’s call it failure, and then you realize that it’s either a bad choice or you don’t need that at all. For example, you don’t need that tool at all. It can be also the outcome of the pilot project. So there are many different outcomes, so don’t be fixated on the success path that is the only right way. It means, okay, the pilot project was a success because everybody agreed that that was the right tool that we want to use, so keep it in mind.

SO: Yeah. I think that there’s … I got into actually a debate with somebody about this. They were telling me that pilot project, proof of concept, and what was the third one? Prototype are not the same thing. Okay. Well, so a pilot project is, “We think this is the right answer and we’re going to try it small and then we’ll expand.” A prototype is, “We’re just going to try it and throw it away,” and a proof of concept is, “We think this is the right answer, but we’re not sure and we’re prepared to throw it away.” And I thought, “This is way too specific for me, but okay, sure.” But to your point, the project, whatever category it belongs in, has different purposes, and it’s important to be clear about what kind of a project is this? Is this essentially beta testing, this is step one, or is this more speculative? We’re just really, really not sure.

PK: I think it has to be falsifiable. Like they say in the scientific method, there has to be a failure criteria somewhere. This will fail if ABC happens, because if you don’t have that-

MS: Or a success criteria, right?

PK: Or a success criteria, but I’ll be more focused on the failure criteria because you can always show, “Yeah, we accomplished this, 30%,” but if your failure criteria is below 50% is fail, then you will say, “I failed this.” You fail five out of six and the project is a no-go.

SO: So I wanted to switch gears a little bit and talk about some specific tool chains and problem sets. I think it’s fair to say that the two of you are, I’m going to say mostly but not exclusively, focused on docs as code, so what does that look like? And there’s a lot of debate about docs as code versus structured content and they’re very much pitted against each other, but ultimately, what’s your perspective on the situations where docs as code is appropriate or maybe is not appropriate, and what do you look for in a project or in a problem set that matches the docs as code model?

PK: I think the reason people associate us with docs as code is the last eight years, we’ve been working at Guidewire and that’s the strategy we’ve chosen there, so I think we’ve been entrenched in this worldview. My previous job before Guidewire was a consultant, and I would go from a company to company and set up different systems for documentation. That was my specialty, systems for publishing and updating documentation. So yeah, circling back to your question of when it works, when it doesn’t, I guess docs as code is more about it works when the right mix of people are creating the docs and the mix of people includes software developers. So for example, at Guidewire, we have several dozen static websites which are maintained by software teams without any technical writers involved, and those are a variety of internal tools, tools which are almost external, and then tools which go out to customers, and all of these little websites integrate very well with our publication system.

And I think this is the main criterion for docs as code fitting is you are giving tools for writing to people where they work. So software developers work in code. You give them tools for writing in their codes so they don’t have to buy extra licenses, get training on anything external, use some alien process, alien to them, just because they follow SDLC, the software development lifecycle to update their docs. And it’s what they’ve been doing, and then it’s easier to convince or it’s easier to fit in a smaller team of technical writers, I don’t know, like 40 technical writers versus 2000 software developers. It’s kind of everyone contributing to the same documentation system. It’s easier for the tech writers to adapt and join the docs as code platform than the other way around.

MS: Just like Paweł mentioned, docs as code makes sense in certain environment. It’s not like we are tribal about it. We love this solution because it works for us. So after a few years at Guidewire, we realized that this is the way we want to go, because as I said, it makes sense and it just works. We tried different things before, and before I joined Guidewire, there were different solutions. We used, for example, CCMS and we had more a traditional approach to producing docs, let’s call it this way, and then we started building our own pipelines, our own solutions. We started integrating with what was there that other devs used for their work, not related to documentation, but we decided, “Hey, maybe we can use what’s out there and just plug into the same infrastructure.” So as I said, it’s not for everybody, it doesn’t work in every environment, so I wouldn’t say, “Okay, if you’re in a factory, you should use docs as code.” No, I wouldn’t say that. So maybe we positioned ourselves as docs as code proponents, let’s call it this way, but this is because we use it every day, we build it, and this just works for us. And also maybe I can mention that we decided to end this holy war by putting DITA into Git and CICD pipelines, so we have DITA as code, and now everybody’s happy.

PK: Oh, this is actually a great point because Sarah, also mentioned docs as code versus structured. Now we’re doing docs as code in a structured way, and where technical writers are working on the docs, they are free to use DITA and a lot of teams do that with all the reuse and all of that, but even if it’s something simple like some markdown files in a repo, we still impose a metadata structure which allows us to integrate the doc into our publishing pipeline, and the two key aspects of what we’re integrating with our authentication and search, and that needs metadata, right? It needs to know who has access to this content and it needs to know how to filter and direct the searches, and from that, our next project emerged. Like everyone else in this industry, we started working on AI solutions, and the AI solution is a third integration point where this structured approach to content also pays off.

SO: Yeah, I did want to touch on AI and so we should probably just jump to that. What are the implications that you’re seeing? What are the effects that you’re seeing on your current processes of the use of AI or the requirement to deliver to AI, and how do you see that working?

MS: Paweł, you want to start?

PK: Yeah. So the content remains the same that we’ve been working on for years, and the structure, the metadata is all the same. We could not change this overnight, but overnight, we had to introduce this AI which is going to look into the content, find the topics, the chunks we call them. It’s going to find the right chunks of documentation and generate answers based on those chunks. So what happened is we had to adjust the AI, but since we were so structured, it was pretty easy, and then we gained a new source of feedback from customers through the AI. Because when people started using our chatbot, we built a chatbot which is available on docs@guidewire.com, but it’s only available to customers and partners, so you need a login on the website.

When you go there, the AI answers questions, and then we monitor the exchanges that people have with the AI and we see what needs to improve as far as content ingestion, as far as content structure. We generate a lot of internal tickets to our tech writing team and identifying gaps, because we’re seeing now that this AI interaction is possible, here’s what people are asking of the content. And other people are asking, “How do I implement X, Y, and Z in the scenario where ABCD?” Which we didn’t know people were looking for before because they had no interaction with the docs. So they just come to the website, they either find or don’t find what they need, and we don’t know anything about that, and we could ask them, but we’re just going to ask a small group of people, whereas now, it’s as if we’re asking everyone.

So we’re seeing these new patterns and something emerges out of those patterns, like everyone’s asking for code samples in this specific area, so we go back to the team and we work on adding more code samples, or people are looking for illustrations of these particular workflows, so people create these illustrations. It’s amazing how rich of a source of feedback this has become.

MS: And it also adds complexity to even testing all the solutions that we have right now, because with keyword search, I’m not saying it was easy because it wasn’t easy, but now we have another layer where you have this middleman, let’s call this chatbot, let’s say. It’s a middleman that can hallucinate, so you can either have a problem with your content or you can have a problem with the agent that responds. So before, when you had a keyword search and you were missing results, you were going usually to the content and see, “Okay, I’m missing this topic. It wasn’t indexed properly, or I just need to move some knobs in my search engine and just see the fuzziness or something like that.” And now you have this content, it’s being processed by this AI tool, and then it gives you some answers and you don’t know why the answer is wrong. Because it didn’t find the content? Because it’s not there? Because it’s not the right content? Because the content is right, but there is something that it missed? There are so many moving parts right now, more than before that it adds complexity to our job, and this is just one side of it.

The second thing is writers using AI for creating content, and we’re also exploring this path because everybody’s trying to incorporate AI solutions into their work. We code mostly on a daily basis and we use some tools that help us with coding that are AI-based, but we also want our writers to benefit from all this development in technology, and we’re exploring, let’s say Oxygen, XML, Positron. Just to clarify, we’re not sponsored anything. We’re just using it so I’m just mentioning the name specifically, but even if you don’t have any specific tool for writing that integrates directly with a AI tool, you can still use, let’s say, Copilot or any other solution that you like and you can just ask it to help you with the content, but it comes with a lot of caveats.

It’s not like you just throw a prompt and just get what you need. It involves a lot of fine-tuning, a lot of working on instructions, giving it context, giving it information that it needs to use for producing code or docs. Because I use it for both, for coding and for documenting, for internal purposes mostly, but it takes time to make it work the way you want it.

PK: Yeah. It’s funny how programming used to be deterministic and it was the people who are messy and the processes, and working with setting up the process for the doc tools, we always knew, people are going to be whimsical and maybe harder to reign in but the technology is going to be predictable. Whereas now, technology is not predictable anymore, and you give it a prompt and you hope it’s going to do what you want. You adjust the system prompts and change the weight of things which are retrieved versus metadata, et cetera, and it doesn’t always work the way you expect it to.

SO: And now the people are being asked to be the deterministic layer, right? To be the QA on top of the AI.

PK: That’s actually very insightful. I like that. That is true. The human in the loop or whatever you call it, that’s supposed to be the voice of reason.

SO: I don’t know about you, I’m not good at voice of reason. I’m much better at causing trouble. So what you’re describing is a fairly complex and time-consuming approach to this in that you’re saying we have this AI chatbot and we’re looking at the metrics and we’re adjusting accordingly and making changes, and then using the AI on the backend for some productivity kinds of things, but not as a replacement. We’ve seen in the US and in North America, we’ve seen a significant number of people losing their jobs because the idea is that, oh, the AI can just do it, and what you’re describing is not that at all. So are you seeing any of this sort of, “Oh, this will make us more efficient, and therefore, we need fewer people,” or is it a different perspective in your piece of the tech com world?

MS: I’m aware of all the layoffs and of all the bad things that are happening right now in the IT world, let’s call it, because it’s not only technical writers because it’s also developers and different jobs, but I think I’m still in this kind of a bubble right now where it’s not happening directly next to me so maybe I’m being too optimistic. But I keep saying maybe I’m going to eat my shirt after some time because I keep saying, “Show me one tech writer that doesn’t have a backlog that is not too long.” So we usually have too much work, and I hope that with the right approach and with a sensible management, these AI tools will be doing all the things that we don’t want to do or we don’t have time to do, and then it will give us time to do something more and something more significant.

I’m looking at my job and I’m not saying it’s perfect because of AI, because it has so many challenges right now and it gives you different problems, but I can see that I can speed up my cumbersome tasks very, very easily. And before, I had to … This is a simple example, and I’m doing a lot of infrastructure work and a lot of backend work, so many things go wrong, and very often, I had to debug difficult problems. It was taking me weeks sometimes to nail the actual place where it happened. Now, I can do it much faster. If I have to debug a problem, it takes me, I don’t know, an hour, sometimes even minutes, and I know that I would spend a day, two, or even a week if I didn’t have these tools. But these are good examples, but there are also a lot of bad examples, but I don’t think we have time for that.

SO: We’ll take the optimistic view.

PK: Yeah, I agree with Michal. The approach we have is these tools can give us tremendous productivity boosts, but not in the sense of getting rid of people. We can redirect people’s work to where it’s more meaningful. And what I mentioned about feedback, identifying these content gaps. Some of these content gaps are going to be very mechanistic and you solve them by generating a bunch of docs out of code, for example, and then you put those docs in the repository where the chatbot can find them, because they’re not going to require a lot of thinking. It’s kind of like API docs where you create the swagger spec and then you generate the dogs out of that. You just grab the source code and you generate some samples, and you use that to generate answers to people. Without people, this wouldn’t work because you need, like we said, the voice of reason, but yeah, people are still required in the process.

SO: I think that looking at it across the industry, if you take AI, take a chatbot, especially a public facing chatbot, and ask it for answers on literally anything, it will give you the average of what’s out there in the world because it’s math, so it gives you the average essentially. And so from a content creation point of view, I think the fact of the matter is that there is in fact a lot of below average content out there. Roughly half is below average, or perhaps exactly half.

If the information is bad enough, if the content being produced is not of high quality or ungrammatical, and we’ve all seen terrible, terrible documentation that was badly translated and is just incomprehensible, then the AI as a tool for creating content may be able to produce something that is better than terrible. It’s not going to replace a professional, well-trained, highly experienced and knowledgeable product documentation group, or it shouldn’t, but that’s not every company. Not every company has a really good group of tech writers that understand the product and are adding value as they produce the content that goes with that product. And so I suspect that what we’re going to see is a split with commoditization on one end, just auto generated, it’s not great but it’s adequate maybe, and then the higher end stuff, which needs to be done well.

PK: Well, there’s definitely documentation which only exists because it absolutely has to, and companies like that can easily generate just some kind of documentation and just use it, right? But in cases where … I think the worth, the value of a technical writer is not just writing and generating the content from below their fingertips. It’s more about the research and understanding, like you said, understanding the needs of the users and understanding how to meet them. You throw AI at a problem which sounds like a generic problem, it’s going to give you a generic answer, not necessarily one that is rooted in the organization you’re in. The list of products that you have, the way those products interact with one another, it’s going to miss all that.

It’s just going to give you … It’s like ordering a hamburger and you say, “So what’s in the burger?” And the AI is going to tell you, “Well, usually in a burger, it’s a patty and lettuce.” And it doesn’t mention that at your restaurant, you use pear and avocado. It doesn’t know that. You know that. You’re the chef back in the kitchen and you know why your burger is special.

SO: Yeah, I think that’s right. So I did want to touch briefly on the question of, because this is a rare opportunity to talk to some folks that are not based in the US or North America, what differences, if any, do you see in tech writing in the market? Now, you’re based in Poland but working for a, I think, US company, but what kinds of things do you see that are maybe different that especially the people listening to this in the US would not be aware of coming out of Eastern Europe?

MS: For me, always, it was the fact that tech com appeared much later than in the United States, in Poland of course, because I’m not talking about Europe in general because there are also differences between countries in Europe. For example, Germany is totally different from I would say the Polish market, because in Germany, you usually have factories and you have technical writing associated with hardware, let’s say.

SO: Yeah, heavy industry, machinery.

MS: Yeah, heavy industry, that’s right. And in Poland, it’s very often about software, maybe because we have a lot of companies that outsource to Poland when it comes to software development, R&D and stuff like that, so that’s the first thing. And don’t get me wrong, people were doing tech com way before it appeared in the mainstream, but sometimes they were not even aware that this is called technical writing or technical communication. Actually, it happened to me because I moved to techcomm from some kind of IT support job, and then after a year, I was like, “I’m wondering if this is anything regulated or there are some rules or it’s actually a profession.” Then I started digging and here I am.

But it turned out there are so many things, so I was also surprised. Okay, this is called that. These are standards. There are books. Well, actually, one of the first books that I read was your book, Technical Writing 101, which I’ll still recommend, although it’s been years since it was published the first time, but I think it still holds a lot of truth about techcomm. The core values of techcomm are still there. So going back to your question, the techcomm scene is relatively young, let’s call it this way, in Poland, which gives us a big opportunity to skip some stages, let’s say. So we don’t have the legacy, we don’t have some things that we used to do a certain way and we like it or we are just accustomed to it. We can just start fresh and just jump. In 2000s, we just jump into techcomm and we say, “Okay, let’s do it this way because now this is the way we do it.” And I think that people are flexible when it comes to looking at things a certain way. Not everybody because there are also people who don’t like changes.

But I think also what is unique about Polish techcomm scene is it’s relatively small, so if you do something outside your work, you don’t treat your job as only a means to earn money and just survive and you want to do something else, let’s say just write an article, like give a presentation, go to a meetup. After a year or two, you just keep meeting the same people, you know basically everyone who is in the circle, so it’s much easier to be visible if you do something outside your work. So I think that would be something unique about our techcomm scene.

PK: I think what might be a downside of the Polish techcomm scene is a lack of veteran experts, because we don’t have people who have been doing this for 40 years. I’ve been in the industry for 18 years.

SO: But you do understand that you two are the veteran experts.

MS: That’s what he’s trying to say, that-

PK: I’m getting to this. So yeah-

MS: Imagine that.

PK: Putting on my clown makeup every morning. What I mean is there’s nobody to look to who wrote the book on technical writing and has been there for 40 years. I’ve been here for 18 years doing this thing, and out of the people who are visible publicly and who contribute to the scene, I don’t know if there’s anyone who has more experience than me. Because I know there are people who have more experience but they’re just not visible. They don’t share their knowledge with anyone, and we miss that, so that’s why we look to the West, to people like you guys, like Scriptorium, and we read books which were created there and we see there’s definitely value, even though something is as old as the scene here or older.

MS: And people also have this tendency to look at things that were done in the past as like, “This is the old stuff. We don’t need that. This is the old way. Let’s do it this way because it’s better now, because we have all these tools and et cetera.” But I think it’s a big mistake, because what they say? History doesn’t repeat but it rhymes, something like that. So in order to do your job in the present, you just need to look at the past, and this way, you can also be ready for the future.

I think it’s worth looking at all this, let’s call it legacy stuff, all the experience that people with 30, 40 years of experience in the field have because it’s valuable. Because usually, when you look at it much deeper, it’s nothing new. It’s usually something that already happened but it’s dressed up as something else. So if you look closer, it’s usually something that let’s say you can understand what happened and then you can apply the same rules, maybe slightly change them or bend them. But my experience is the same happens in software development, in coding. People are inventing new things, and when you look closer, it turns out somebody already said that in the past. It was already done this way, but it’s dressed up as something else or named differently or is hyped more, or it was invented too early and now is the time to use it. So the history gives you a lot of perspective, a lot of things that you can use, so don’t dismiss it.

SO: Well, I think that’s a good point, and as we close this out, I want to ask you what you see in the past or where you’re gaining perspective on the introduction of AI. Whether it’s from a delivery point of view or a backend productivity point of view, where do you see that pattern previously, or do you?

PK: There are similarities, but they end. They’re not perfect analogies. So going digital was on example, and when I started, it was just on the brink of companies going from print to web, and that was a huge paradigm shift. And we’re seeing reverberations of this even today where teams are still thinking in books and chapters instead of thinking in pages and thinking about even every page is page one, which is, I don’t know, it’s older than me, the idea, but it still hasn’t been adapted to by teams. Now, the AI thing is probably a similar earthquake where it’s the AI reading your docs, giving you answers. You’re using AI to generate these docs, et cetera, et cetera. I don’t think we’re going to get 20 years of leeway to adapt to that and still see people doing things the old way. I think the world maybe moves faster nowadays, but I don’t know, we’ll see.

MS: It may be something similar to the invention of computer. So instead of writing let’s say manually, people have got this new device that gives you more power and you can do more things and then programming languages and everything, but as Paweł mentioned, the pace was much, much slower. So we had years of development to, let’s say slowly, maybe this is the bad word, let’s say slowly adjust to the change, and now it’s an earthquake. So every day, every week, there is something new and you need to keep up, keep up, so the pace of development I think is the biggest challenge. Because we tech writers survived many revolutions. I just started reading a book by Sharon Barton about the women in technical communication, and women who talk about their careers, they mentioned many, they start very early because there are a lot of examples from the States.

So they started working when they were not even treated equally with men, which was years ago, and they mentioned all those changes, all those revolutions, all those evolutions. And I’m saying, “Okay, this is what we are witnessing right now, but the pace is definitely faster, so we need to just keep up,” which is hard, of course.

PK: Well, I have another one which is a good one. I don’t want to not say it. The dot-com bubble. I think maybe we’re in a bubble again, maybe. You’re seeing the inflation of AI prices right now, and I think we’re going to end up in a position where you cannot do all these things with AI which we’re hoping to do with AI, so the surface of usage is going to shrink and there’s only going to be specialized uses and special case scenarios when you use AI because it’s going to be expensive, and a whole lot of AI applications are just going to disappear. This might be a parallel, but I don’t know, it’s always hard to predict, especially the future.

MS: I’m also predicting some corrections. Let’s call them corrections.

SO: And I three am also predicting some corrections. You touched earlier on the fuzzy. AI is very good at fuzzy problem solving, but it’s being thrown at things that are old problems that we know how to solve with scripting. And scripting from a computing power point of view is cheap or maybe free. You write your script and you run it, and every time you run it, you get the same predictable result. Now, sometimes you don’t want that. Sometimes you need that fuzzy pattern matching thing, but in the cases where that’s not what you want and we’re still using AI, that is part of the bubble, right? Everything looks like an AI-shaped problem. So yeah, I agree with you. I think that any predictions we make on specifics as to where this is going are guaranteed to be wrong, because I’ve tried this before and I’m always wrong. So I want to thank you both. This has been very entertaining and very interesting, and I hope that we will see you again and you’ll come back and tell us more about what you’re up to over there.

PK: That would be lovely. I would like that very much. Thank you.

MS: Yeah, sure. No problem. Thank you very much. That was fun.

SO: Yeah. Thank you both. We will see you soon.

Want to learn more? Download our book, Content Transformation.

The post Tool selection and the unpredictable variable appeared first on Scriptorium.

AI promises to transform content conversion, but what does it actually look like when you’re processing thousands of documents a day? In this episode, Sarah O’Keefe (Scriptorium) and Rich Dominelli (DCL) dig into the real-world challenges of using AI for large-scale structured content conversion.

Rich Dominelli: If you have millions of articles and you’re asking the AI, ‘What did we do for this project six months ago?” The AI has to find those articles, pull the relevant information out of those articles, summarize it, and hand it back to you. The best way of doing that is to give extra signals to the AI, structured relevant bits of information, front matter, back matter, publication date, keywords, abstract, that allows the AI to query the corpus and get the relevant chunks out of that corpus in a very quick manner. Then, it can summarize what those chunks are. So the AI almost becomes the user interface over that corpus. But to find that data in the first place, structured content is key. Structured content is key when you’re dealing with big indexes and the web, and it’s the same with AI.

Related links:

Defeating Nondeterminism in LLM Inference (white paper)

Data Conversion Laboratory (DCL)

Scriptorium, Machine experience (MX): Making content work for humans and machines (podcast)

LinkedIn:

Host: Sarah O’Keefe

Guest: Rich Dominelli

Transcript:

Disclaimer: This is a machine-generated transcript with edits.

Introduction with ambient background music

Christine Cuellar: From Scriptorium, this is Content Operations, a show that delivers industry-leading insights for global organizations.

Bill Swallow: In the end, you have a unified experience so that people aren’t relearning how to engage with your content in every context you produce it.

Sarah O’Keefe: Change is perceived as being risky; you have to convince me that making the change is less risky than not making the change.

Alan Pringle: And at some point, you are going to have tools, technology, and processes that no longer support your needs, so if you think about that ahead of time, you’re going to be much better off.

End of introduction

Sarah O’Keefe: Hey everyone, I’m Sarah O’Keefe and I’m here today with Rich Dominelli who is a Senior Developer and Architect at DCL. Rich, welcome.

Rich Domineli: Hi, thank you for having me.

SO: Glad to have you. We were talking before we hit the record button, and you described yourself as a perhaps hopeful AI evangelist.

RD: Yeah, I am well and thoroughly immersed in the AI game at DCL and using it and plus I play with AI assistants at home. I’m enthusiastic about the future of AI, sometimes disappointed about the present.

SO: So DCL, as I think many of our listeners know, is focused on conversion at scale, which to me makes a great use case for AI because ultimately conversion is about edge cases and about inconsistency, right? If everything was 100% consistent, conversion would be pretty easy.

RD: Yeah, no, DCL does a lot of structured content generation out of unstructured data, and the creativity, especially in the academic space, of what that unstructured data looks like is sometimes nightmarish. So the AI lets us, does a lot of the heavy lifting for us when it comes to looking for particular items, identifying concrete data points within the documents, pulling things like authors and affiliation, front matter type information, and back matter type information out of the documents and in automated fashion. It can be painful from time to time, but it’s definitely helped.

SO: Yeah, so this is, think, you know, the reality of working with AI and working with it in a production environment in order to address all these weird edge cases and what’s going on. So tell us a little bit about how you’re using AI in, you know, these conversion use cases. What does it look like to go in there and start applying some of these tools that we have?

RD: So, I mean, typically our flows work in a way where we’re coming in with a PDF or a Word document or some other unstructured format. We take it, we reformat it into a version that’s more AI-friendly, like Markdown, for example. And that’s usually the first step we’re doing when we’re looking for information to pull out of it like front matter. It’s a very common use case.

If you look at academic papers, the front matter, the authors and the affiliations that are on that paper can be formatted in more ways than I could list out during the course of this podcast. It’s kind of crazy. So what we’ve started doing, and we’ve been doing this for a couple of years now, is we’re using the AI, we’re handing it the Markdown document, and we’re saying we need to list authors and affiliations, please extract it for us. 

Now, naively, when we started that process, we assumed that the AI would give us a consistent list of authors and affiliations. And sometimes it does. But every time you do that call, you’ll get it in a different format. So then you have to start tightening things down. So OK, give me a list of authors and affiliations. I want it to be structured exactly like this. And typically, we have a JSON structure that we’re presenting to the AI, along with our prompt, and saying, give it to us. Well, okay, and that gets you a good chunk of the way there. And that was very exciting when we had that working consistently, we were getting things out of the system on a consistent basis. Awesome. But then you start looking at the results, and every once in a while, you get an author that was missed, or there would be too many authors on that paper. 

We had one test paper, which I loved, which had 600 collaborative authors in it. And the AI would just choke after about 280-ish. So then you have to start dealing with things like paging through the data and formatting the data. And then you have to figure out, well, did it miss anything? You have 600 authors. Good luck. So now you have to take what the AI did and compare it against your own representation of it and write a program to do that comparison to say, OK, is it good? Is it good?

You have to take a step back and you look at it and you say, okay, we have the information that’s in the non-structured format. We’re handing it to the AI. The AI is gonna give us a structured version of it and we need to validate it. Well, the first validation is very easy. Does that structured version match the schema that we gave it? Yes or no, that’s easy. Well, then you have to say, okay, is everybody there? Well, is there anybody added? Because the nice thing about AI is they occasionally get very creative. Even if you have that temperature dial turned all the way down to zero, it will pull names out of thin air and then come back to you with some random name and stick it in the middle of the data where it’s not obvious, of course, and then hand it back to you. So then you have to start saying, are all the names that appear in this list actually in the document? Are the counts matching? And if it’s not, you go back to the AI and you ask it again, and usually you’ll get a better answer the second or sometimes the third or fourth time. 

But you need to be able to catch that, especially if you’re doing this at scale, because if you’re doing a few, it’s easy, you can eyeball it. If you’re doing 1,000 of these a day, you can eyeball all of them. You can say, you can ask the AI, OK, give me a confidence level, but if you can’t trust it in the first place about what it’s returning, yeah, I’m very confident about what I’m giving you right now. It’s really the truth, I promise you this time. I don’t know how trustworthy that would be. So you have to write tools to validate what the AI is producing, or you have to use the AI to validate what it’s producing. So coming in the first time, obviously, we did the count, we did the schema validation. We then said, okay, we’re going to check to make sure all the names appear in the document, we’re going to have landmarks in the document that we can refer back to. So if you start with Microsoft Word and you have track changes on, you can have paragraph IDs that are supplied. So you can make sure that you can find all of the authors in that list and they all have a paragraph ID and you can have your landmarks and that’s great. Or you can even hand the results to a separate AI call and say, proofread this. Is this accurate? Is this the best answer that could be for each of these? I know we’ll come back with an answer. And you can use that as a signal to gauge accuracy and to gauge repeatability and make sure it’s correct.

SO: So you’re, let’s see, generating an AI, not a test bed, but an AI environment that’s doing this conversion or that’s processing the files for you for conversion. And then you have to go in and do all this validation to make sure that the output that you’re getting is actually correct. As compared to, I’m gonna say old-fashioned, but you know, as compared to scripting, deterministic, pretty straightforward, if A then B kinds of scripting. What are the differences between that and AI-driven conversion in testing and validation? What are the test plans? How are they different conceptually?

RD: So from our perspective, the frustrating thing sometimes is the AI is completely non-deterministic. 

SO: Mm-hmm.

RD: It can give you a name formatted one way today, and then tomorrow, its formatting might be subtly different, where in the paper it has “Richard Dominelli, Junior.” The AI may decide, well, that comma probably shouldn’t be there, or junior should be followed by a period, and it wasn’t in the paper originally. And you can try prompting around that and tell it to prompt around that and make sure that it’s accurate. But it doesn’t always follow your instructions exactly when that’s the case.

SO: And why is that? Why is it non-deterministic?

RD: Because AIs are built on a neural network, the neural network itself has fuzzy fields within that, mostly due to floating-point arithmetic. So when you’re looking at it and it’s that weight on that particular key might be out to like 16 digits of a number and it might shift it slightly one way or the other. There is a fantastic paper from, I wanna say it’s anthropic, that goes through the different reasons why AIs are non-deterministic. It goes through repeatedly querying for the AI and who Richard Hyman is and getting back a different answer every single time. They’re all correct. However, they’re all slightly different. The other thing that will lean into that is if the AI is being heavily used, the memory and model weights will shift ever so slightly and you’ll get a different result.

So you’ll end up having an issue where today I’m getting accurately this way and it’s relatively consistent, not perfectly, but close enough. And then tomorrow, it may just give you a dumpster fire of random information and you need to be able to detect that. Okay, the other challenge we hit fairly early on is more and more people are aggressively using AI right now. So we’re actually starting to hit issues where the LLM providers are overwhelmed. So you have to be able to code in sale over because you’ll literally get too many, you’ll get 429 errors, which are basically, I’m too busy. I can’t deal with your request right now. Call me back. And you’ll have to go back and repeatedly query to get around that. I am hoping at some day in the near future, we’ll be able to have in-house AI at scale and have these wonderful models that are so intelligent that we can run on our local hardware. And so I won’t have to deal with that, but right now, that’s not the case.

SO: So given all of this, I mean, I’ve asked you the leading question about the issues and the negatives, but what then makes an AI-driven conversion appealing versus a sort of scripted, deterministic, if I plug in AI, I will always get B output?

RD: So part of it is the type of data we’re dealing with. We’re dealing with unstructured information and the unstructured, the creativity of the unstructured information is rather astonishing. You’ll have people format things, know, we’ll get papers in where the entire paper is placed in different cells of the table. It’s not tabular information at all. They just, you know, we wanted this particular section to be in this cell and this particular section to be in this cell and this particular section. And the AI, I don’t want to say is immune to that, but it’s a lot more forgiving than having to write those reg ex or traditional programming or word interrupt things to try to extract that information, because the AI can address it in a much more fuzzy fashion. I know approximately what an author’s name looks like. I know approximately what a reference looks like. Even though today they decided to do it in Comic Sans or with Wingdings fonts, I can still read that and move on. So that’s really the wonderful aspect of it, is it gets around a lot of that fuzzy logic coding. You’re not dealing with having to address each of these nuances in a generic switch or state machine to try to figure out, OK, this paper should be classified this way and this approach used. Instead, the AI does a lot of that heavy lifting for you.

SO: Okay, so it gives us that sort of fuzzier, more, I’m gonna say more flexible, I know if that’s exactly the right word. And then the outcome, what you’re describing is you’re ingesting unstructured word, PDF, those kinds of things, and turning them into structured content, presumably fundamentally XML of some sort, but also some other downstream formats. So I wanted to switch gears a little bit. There’s been a lot of conversation about using structured content as an input for AI. So this, guess, is the scenario where you’ve already ingested the unstructured content, have remediated it in various ways. We now have structured content, and we’re gonna take that and feed it into, I guess, AI part two, right? So we’re past conversion. And there’s a lot of people saying, you should feed structured content into AI, it will make the AI better. And so my question for you is, you know, is that the case, and also maybe why and what goes into structured content that makes it produce better AI outcomes, potentially, assuming that it does.

RD: So there’s a bunch of guides out there. There are two pieces of conversation. First, there’s a bunch of guides out there for prompting AIs where they suggest using XML or simplified XML tagging to give the AI signals about your prompt that aren’t verbally expressible. So here is my question. Here is an example. Here’s how I want my output to look like. And you can put tags around that when you’re actually prompting the AI and the AI will know that those signals mean that it should pay attention to it. Okay, so that putting that aside, what I think you’re really asking though, is how does structured content, structured documents, the JATs and the DITAs and the S1000Ds and how does that help the world of AI? And to answer that question, we have to go through two things.

One, we have to go through retrieval augmented generation and context rot. So let’s talk about context rot first, because that’s a really interesting topic and people don’t talk about it enough. You have these large language models that are coming out right now and they’re advertising this sticker shock value of, can ingest a million tokens and it has this tremendous memory so you can stick the entire encyclopedia botanica in it, and it will be able to ingest it and regurgitate it. There’s a whole lot of academic work out there that basically says that, hold on a second, practically speaking, once you exceed a certain size, even though they can technically hold that million tokens of data in memory, they’re not gonna be answering as accurately as a smaller model.

The most common example or the most easy test for that is needle in the haystack test, where you take a document, you stick a random fact in the middle of it, and you hand the AI the document, and then you ask them for that random fact. Nine times out of 10, it will answer incorrectly. An even easier test is there’s a website which I actually like called A Thousand Names. And all this website is is a thousand randomly generated human names. The thousand randomly generated human names. You take that, you give it to the AI, say, how many names are there? And more often than not, you’ll get, well, when you do 100, you’ll get an accurate answer. 200, accurate answer. 300, things start to break down. You might get 300, or you might get 280, 320. You might get a random answer. 

And then it gets progressively worse as it gets bigger and bigger. So if you’re working in the context world, content world, you’re looking at ingesting documents into a corpus of some sort. You’re making these structured documents in such a way for the sole purpose of making them retrievable. You want the AI to be able to retrieve those documents and the relevant documents from the corpus so that I can answer the question. A, because your corpus is probably bigger than that million tokens. And B, because the less data you send the AI, the more accurate the answer is. So the better way of thinking.

SO: And so a token is roughly a character, right?

RD: No, a token is actually roughly a word. It’s less than a word. It’s kind of a lot, but it’s still not like a PubMed-sized corpus or anything like that. It’s roughly the size of the New and Old Testament of the Bible, roughly a million words. So just give everybody that mental picture. But that’s just one book. 

SO: Roughly a word. So a million tokens is kind of a lot. It’s a lot of words.

RD: So if you have millions of articles, or and you’re asking the AI, you know, what did we do for this project six months ago that involved JAPs in this solution? And the AI has to say, okay, it has to find those articles, and then it has to find the relevant information out of those articles to be able to summarize it and hand it back to you. And the best way of doing that, and the best way we know how to do that is to giving extra signals to the AI, giving those structured relevant bits of information, front matter, back matter, publication date, keywords, abstract, that allows the AI to query the corpus and get the relevant chunks out of that corpus in a very quick manner. And then summarize what those chunks are. So the AI almost becomes the user interface over that corpus, because it’s going to summarize the data. But to find that data in the first place, structured content is key because for the same reason, structured content is key when you’re dealing with big indexes and web, same with AI.

SO: So then structured content is potentially helpful. And I guess then circling back, let’s say I’m sitting on a pile of content of varying degrees of structured or unstructured, varying degrees of quality or lack thereof. What kinds of things should be happening before that content gets ingested into some sort of an LLM or some sort of a corpus to be used in AI-generated outputs?

RD: So these are the same type of things you would do to make them easily retrievable ahead of time. So the standard approach that was being espoused about two years ago, a year and half ago, was something called Naive RAG. You can just take your PDFs and throw them at the AI, and the AI will ingest them into a vector database, and it will do semantic similarity and find the documents that you care about, not the best approach when you start talking about large amounts of documents. And there are issues with semantic similarities, where the AI will have a hard time distinguishing negative cases, will have a hard time peeling out the best documents, and that type of thing. So the best approach to take is you want to take those documents, you want to turn them into structured information in such a way that it’s easy for the AI to ingest. So typically that involves chunking it, into topic-level pieces or semantic chunking, coming up with summaries to make them easy for the AI to find, and whatever other information you may want to chase out of those. 

So, for example, if I’m handing a PDF to an AI and saying, I want to be able to search this PDF later, well, six months from now, if I get a new version of that PDF and I want to search it, search against the two of them, I really want my answers coming out of the second PDF. That’s metadata, that’s structured information that doesn’t appear in the text of the PDF or may not appear in the text of the PDF. You wanna be able to do things like versioning, you wanna be able to do things like dates, you wanna be able to give these signals to the AI to be able to pull that information back quickly. And that’s really where structured content comes in. So for the purposes of preparing your own corpus, you want to convert them into an easy to ingest format, which typically means Markdown or XML or something that the AI can deal with. You want to give it whatever other signals you can so that it’s easy to find. And then you want to hand it to something that first does chunking and then text embedding, which is basically turning the information into numbers so that you can do those cosine similarity searches. And then you want everything handed off to some kind of object store like a hybrid brand database or the hybrid factor database or graph database so that they’re easy to pull out.

SO: Awesome. So you started this off talking about being the hopeful evangelist, and now having gone through all of this, it sounds as though you’re really thinking about these issues and dealing with them at scale. What are some of the top things that you’re thinking about going forward, whether hopeful or not, the good, the bad, and the ugly?

RD: So one of the interesting aspects of my job is I get to do a lot of interactions with AI from an R &D perspective and do some in-house programming and do some in-house tool use. And what we’re finding is developing our own internal mechanisms for AI to call third-party tools, to be able to call Crossref or Grovid or some of these reference facilities out there through like model context protocol or through API calls so we can execute those calls and get that information back and do validation before it hands back the results is a very interesting topic for us because that would let us do things like any AI have it do the first few rounds of validation before it ever comes back to us without having it go to the next step, do a validation step and then the next step and then possibly do a round trip. It would be a much faster interaction. We use right now, of course, like most of the world, we’re using a lot of AI coding tools to tighten up our code bases to make sure things are working well, to basically act as a force multiplier when we’re doing development on projects, which is phenomenal.

I can’t say enough good things about Cloud Code, you know, because it’s really become an essential tool in my day-to-day life. But I’m also seeing a lot of people out there using these tools to help analyze their own and improve their own workflow and that day-to-day work. We talked with one of our customers recently, and they use cloud code, even though the person giving the demo was not a developer; they use cloud code to answer the RFP. And Cloud Code does a tool use call against their document corpus, answers the RFP correctly, and what used to take two or three days of slogging through documents and finding things are now being done in an hour by one person instead of having multiple people working on this project. So it’s great to start seeing that type of stuff in the enterprise just blossom because it’s really exciting.

SO: Well, Rich, I really appreciate your insights on this. I learned a few things and I think that it’s great to hear from people who are actually using this stuff, you know, in a production world, in a high stakes world where you’re actually, you know, need to get the content right, get the information right as opposed to just, you know, that we’ll play around with it and not worry about it too much. So thank you, and we’ll look forward to hearing more from you and what you’re doing at DCL.

RD: Sounds great. Thanks for having me.

Conclusion with ambient background music

CC: Thank you for listening to Content Operations by Scriptorium. For more information, visit Scriptorium.com or check the show notes for relevant links.

Want to learn more? Download our book, Content Transformation.

The post Taming AI: Using AI for content conversion at scale appeared first on Scriptorium.

Your website may look great to humans, but can machines understand it? In this episode, Sarah O’Keefe (Scriptorium) and Tom Cranstoun (Digital Domain Technologies) explore the emerging discipline of machine experience (MX). Sarah and Tom discuss what AI agents actually encounter when they visit your web pages, why microdata and metadata are critical, and what content creators must do to ensure content is consumable for both human and machine audiences.

Tom Cranstoun: Humans are looking for pictures, they’re looking for text, and they can infer. You may think, “Well, we’ve already added information on the page,” but by putting it in as microdata, it doesn’t appear on the page for the humans. It appears on the page for the machine. I think that that’s a critical distinction. We are trying to design for both. We don’t want to overload a human with information, but we do want to give the machine as much information as it can take.

Related links:

The Gathering

Digital Domain Technologies

MX books

The Scriptorium Content Ops manifesto

LinkedIn:

Host: Sarah O’Keefe

Guest: Tom Cranstoun

Transcript:

Disclaimer: This is a machine-generated transcript with edits.

Introduction with ambient background music

Christine Cuellar: From Scriptorium, this is Content Operations, a show that delivers industry-leading insights for global organizations.

Bill Swallow: In the end, you have a unified experience so that people aren’t relearning how to engage with your content in every context you produce it.

Sarah O’Keefe: Change is perceived as being risky; you have to convince me that making the change is less risky than not making the change.

Alan Pringle: And at some point, you are going to have tools, technology, and processes that no longer support your needs, so if you think about that ahead of time, you’re going to be much better off.

End of introduction

Sarah O’Keefe: Hey, everyone. I’m Sarah O’Keefe. Today, our guest is Tom Cranstoun, who is founder of a machine experience, or MX community, called The Gathering. He has a couple of books on MX and is currently a consultant operating as Digital Domain Technologies. Tom, after 53 years in the business, some experience with AEM at very, very large companies, including a huge project at Nissan, has turned his attention to the question of how machines, which is to say AI agents, interoperate with the current public-facing web. And so today, Tom, I’m delighted to have you on to talk with you about machine experience, or MX, and what this all means as we move forward in this brave new AI world. So welcome.

Tom Cranstoun: Thank you, Sarah. I’m very pleased to be with you today.

SO: I am delighted to have you. So I guess we’ll start with the extreme basics here, which is what is machine experience, or MX?

TC: Yeah. MX, well, to my definition, machine experience is like user experience, but it’s for machines. Machines cannot ask a friend for help if something goes wrong when they’re browsing a website. They can’t turn to a partner and say, “What do you think this means?” They can’t retry a failing form input because they will just go through the same mechanical patterns to try and carry on throughout the web journeys. Therefore, machine experience is thinking about what elements one must put on a webpage to help a machine understand and action the final goal of the webpage, whether that be a CTA that lets you purchase something, or an information document that lets you know about a government policy, or a charity good, whatever the author of the page is trying to get across to the audience.

SO: And so at a high level, what does it look like to build out machine experience? What are some examples of things that you need to put onto a webpage to accommodate the machine that’s reading it?

TC: Well, the very first level is the disabilities angle, things like the Americans with Disabilities Act, that kind of WCAG, W-C-A-G, the accessibility work. The more accessibility information is on the page, the more the machine can understand the background of the page. So machine experience and accessibility are pretty much at the top level, the same sort of thing. If you put in JSON-LD, microdata, and you enrich your pages with the things that Americans with Disabilities Act would like, you’re actually helping a machine understand the page. So that is the top-level constraint. When you go below that level, you need to give the machine lots of information about your product, not just the thing that a human wants when it’s glancing at the page now, and as you go through the journeys, things will be added on. Humans can only take in two or three items at a time, so we design pages to reveal what is happening. You go to a catalog, to a product, to a variation, to a purchase, four different steps. Each step introduces different pricing and concepts. It’s best to feed the machine on the page that the machine lands on with all of the information that it needs. This may not necessarily be surfaced to the human reading the page, but it’s there for the machine. This helps the machine when it arrives at your webpage.

SO: So I’m really enjoying this concept that a properly organized page with proper accessibility WCAG or ADA compliance and support then results in the machine being better able to parse the page for essentially the same reason, right? It’s properly structured, it’s predictable. The things that are labeled are labeled correctly. I don’t know that we should be driving accessibility in order to enable AI, but on the other hand, if it gets us more accessible pages, then let’s certainly do that. Can you give some examples of what happens when pages are not machine-compatible? What are the kinds of problems that people run… Or not people. What are the kinds of problems that the AIs run into when they try to parse a page that has not been labeled properly or encoded properly?

TC: Yeah, I collect these examples from real life. Whenever I use the web as a normal person, I say, “Well, how would a machine interpret this?” Recently, I was looking for a holiday, and I asked an LLM to give me a list of five companies that offer cruises up the Mekong Delta. The machine came back with one offer at $200,000 for a week’s holiday, and the rest of them were $2,000 for a week’s holiday. What had happened there was that the machine had found a European website. Now, the Europeans changed the comma and the dot in monetary labels differently from what the Anglo-Americans do. We use a comma separator between thousands and a full stop between fractions. The Europeans actually put the full stop as the thousand separator and a comma between the fractions. This meant that when the LLM built a table of prices for holidays, it didn’t understand the distinction, and it tripped up. The agent hadn’t been instructed to compare prices and make sure that they were all within the same range and were reasonable. It just produced them as a matter of a fact. “Here’s a holiday for you. One of them is $200,000. The rest of them are 2,000.” There was no knowledge, no information that could tell the agents what was happening. If those pages had been decorated with currency and they had microdata with the… microdata always says that you should use commas as a separator and full stops as the fractional separator. If these things had been in the page, the machine wouldn’t have flipped up. Now, a human could have read a page and seen the locale values shown on the page, and both people would be able to understand what was going on. So that’s a typical trip-up from an undecorated page.

SO: And so essentially, the presentational component that says, because I’m serving this page to somebody in, for example, Germany, they are expecting a comma separator between the full Euro amount and the cents, the Euro cents. But that comma is essentially formatting, as opposed to data, and so here we are.

TC: Yes, correct. And the microdata has got the thing in a proper machine-readable way. The other things that we always get problems with in the world are English and American date formats. We swap the month and year around when doing short form. The machine-readable version uses ISO dates, and ISO dates put in as a microdata tells the machine categorically. It doesn’t matter what the locale is, this is the date and time.

SO: Yeah. And so as the expression of the date, whether April 1st is 1-4 or 4-1 is essentially a formatting problem.

TC: Correct. And these are not visibility problems. These are machine experience problems. So it’s layering up. You start with fixing the disability by doing machine experience, and then you fix the locality and the community values, the human factors, display factors.

SO: And so I think we’re all familiar with the concept of a customer journey, but you’re now talking about a machine or an MX journey. What does that look like? I mean, how is the machine processing of a website? How do you explore that journey and what it looks like?

TC: The machines will not discover your website, come in through your landing page, and then look for offers or products. A machine will have an idea of where it wants to go and will land straight in at a page. It will arrive five pages into your journey, and read the webpage as it is. The owner of the website has lost all of the signals about what the dwell time was on each page, how’s the reader arrived at the end location. Did they go sideways and look at other things? Those things don’t happen with machines. They go straight in, see if they can get what they can. If they can get what they can, they will action it. If they can’t, they will move on, and go to another page or another person’s website and do exactly the same to them.

So when a machine arrives at your webpage, it will not be giving you any referral details. It will not tell you what the journey it is, and it won’t tell you what else it’s interested in. You’ll just get a cold caller who will arrive and disappear. I call them invisible users. They’re invisible to your analytics, they’re invisible to your tracking, and they’re invisible to your future. You cannot tickle them and say, “Hey, you left something in the basket.” You cannot use those parts of the journey. A machine comes in and goes, gets what it wants or it doesn’t. So you must give it, front load it as much information as possible on any and every page that a machine may land on. 

SO: So then coming at this from the perspective of structured content people, because a lot of what you’re talking about, I mean, is web experience, like how does what we view as the end state result of the content that we’re creating. So if I have an enormous DITA CCMS full of stuff and then I output it to some semblance of a website, your focus is on what needs to be on that website so that it is describing itself in such a way that the machine, that an AI or a crawler can go in there and pick up what it needs to and process it accurately and not offer you a vacation for $200,000. I assume you did not pick that one. So what are the opportunities? When you look at MX and then also DITA as a backend, what kinds of opportunities do you see there to map those things across and take advantage of some of the structure that perhaps is already in the XML and/or structured content systems?

TC: Yeah, I see the backend is full of good content operation stuff. Everybody has got details about pricing and dates and frequency, and there’s lots of backend information, which often doesn’t make it into the front end for people. Humans are looking for pictures, and they’re looking for text, and they can infer. They can infer if two prices are on a page and it says, “Was $200, Now $180.” A human understands that. A machine, well, depends on the quality of the machine, whether it can read and infer those things. So the backend information has to be made more visible and in a redundant manner. You may think, well, we’ve done this on the page before. We’re doing this on the page after. But by putting it in as microdata, it doesn’t appear on the page for the humans, but it appears on the page for the machine. And I think that that’s a critical distinction. We are trying to design for both. We don’t want to overload a human with information, but we do want to give the machine as much information as it can take. We don’t necessarily have to surface all of the information within the page, but we have to carry it with the page. So a page taken in isolation contains the entire story, not just the fraction that a human is looking at, which does mean that a lot more pushing off the backend from data to the front end. And some people will think that’s a waste of time, but I don’t think so. I think giving that extra material to the machine is what makes the journey successful for the machine.

SO: I’ve been to a lot of conferences in the past couple of weeks, and the conversation around what is needed for successful LLM processing, or crawling, or ingestion, or agents for that matter and what is already provided in a metadata-rich structured content system is sort of, well, we have all of this. Now, what do we do with it, and where do we put it, and how do we make sure that this all works? So it seems like this discussion around machine experience is going to help to maybe close that gap and connect the pieces such that we can do this successfully.

And so as we move into this, I know that you have some material out there, but also a community. Can you talk a little bit about the MX community, and what you’re looking for there, and what it’s called? We will put all of the links in the show notes. But what does it look like to participate in that community, and what sort of participants are you looking for?

TC: Yeah, we are looking for content creators. We are looking for business owners. We are looking for technical writers. It’s called The Gathering, gathering being a Scottish term for the gathering of the clans. We all get together to do something that’s good for the combined grouping. And then after we’ve created whatever we’re going to create, we go away and do our own things. Now The Gathering is tg.community. That’s https//tg.community. We are building a set of community-led standards to try and make it easier for machines to understand documents. The Gathering is not just interested in HTML. We’re talking about documents of all types, and we’re talking about keeping the metadata that you have in the backend of the content creation systems, whether that be data or other content creation systems, and passing it through into the end documents.

You have metadata in PowerPoint slides. You have metadata in Word documents. You have metadata in JPEGs. These, too, deserve the machine experience. If you can tell the machine details about an image inside a JPEG, then the machine doesn’t have to try and scan and interpret the image to find out what it is. It makes things so much better. And The Gathering is a community that is trying to build these as open community-led standards. One of the first things that I am proposing for the community, which was just launched on the 2nd of April, 2026, by the way, it’s very young, and we hope to build at the speed of LLMs. We need to work fast.

The key point and the key thing that helps LLMs understand your website, there’s a thing called llms.txt, which people don’t really understand and machines don’t really use. It’s a standard for describing your website in a way that a machine can help to understand, know what’s going on without reading your site map. It is not used by the machines because, one, it’s not served as HTML, and, two, it’s not in your site map. Therefore, the crawlers that build your training material do not pick it up and do not ingest it. I am suggesting, and I have it in my books, I talk about this, if you wrap the llms.txt in HTML and serve it as HTML and put it in your site map, then you will get a better response from the training stage and from the inference stage. So you are seeding the machines with the information about your website, something that is currently missing from the world, and that’s step one. There are five steps that you’ve got to go through before you can do a successful e-commerce position. And that is feed the machine, get noticed, be descriptive, be MX-aware and be citable, and MX lets all of those things happen.

SO: Perfect. Well, Tom, I know that there’s a lot to discuss here, and we could go on for a very long time, but I hope this gives people a little bit of an introduction to this idea and an opportunity, if they’re interested to reach out to you and to the community that you have. And there’s also a book or three. Any closing thoughts that you want to pass on before we close this out?

TC: My personal opinion is that I think that we should treat the machines as first-class citizens and not block them from our content and to create content that works for them. The more that we do for them, the more they will do for us. And if we start treating them as an afterthought, it’s not going to be such a good web as we could build.

SO: Okay. Well, thank you so much. I’m glad we had an opportunity to talk.

And we will, again, put the links to the various resources that Tom mentioned, including the community. There’s some RFC, some standards drafts and a manifesto and a book. We will put all of that in the show notes. So Tom, thank you again for being here, and I look forward to hearing more on this effort.

TC: Thank you very much, Sarah.

Conclusion with ambient background music

CC: Thank you for listening to Content Operations by Scriptorium. For more information, visit Scriptorium.com or check the show notes for relevant links.

Want to learn more? Download our book, Content Transformation.

The post Machine experience (MX): Making content work for humans and machines appeared first on Scriptorium.