UI for AI: Responsible content delivery (webinar)
With AI, users are taking control over content delivery through summarization, personalization, translation, and more. But what are the risks? In this webinar, Sarah O’Keefe, CEO of Scriptorium, and Fabrice Lacroix, CEO of Fluid Topics, explore strategies and share examples of UI for AI that empower users while protecting them—and your organization—from misinterpretation, incomplete information, and compliance breaches.
As somebody who works in structured content with metadata, taxonomy, and all those other fun things, we’re telling people, you have to do the work. You have to do the work upfront because once that ingestion step happens and the AI is ingesting not structured, not consistent, not governed, not accurate, not up-to-date content, then what chance does the AI have? The AI is not going to make your content magically more accurate. It’s not magic. I mean, it can do some magic looking things, but it is not magic. Your entropy always wins. Your content will always sort of degenerate, right? So you start for your best possible, and it goes down from there. So what’s the best possible thing that you can get into your database?
— Sarah O’Keefe
Resources
- Fluid Topics
- The true measure of success for AI initiatives (podcast)
- AI in the content lifecycle (white paper)
- Want to stay updated on key industry insights? Subscribe to our newsletter!
- Sarah O’Keefe (Scriptorium)
- Fabrice Lacroix (Fluid Topics)
- Patricia Grindereng (CIDM)
Transcript
Patricia Grindereng: Morning, good afternoon, or good evening. Welcome to a CIDM webinar. Today we’ve got “UI for AI: Responsible content delivery” with Fabrice Lacroix with Fluid Topics, Sarah O’Keefe with Scriptorium. Welcome.
Sarah O’Keefe: Thank you. Hey there.
Fabrice Lacroix: Thank you. Hello. Okay. Let me first get set. Make me a presenter. Share the screen. How does that work? Share screen. It’s always the same. Okay. Okay. You should see it now. We’re good. We’re good to start. Hey. Hi, Sarah.
SO: Hey, Fabrice. Good morning.
FL: Good, good. Good morning everyone.
SO: Welcome everyone. This should be fun.
FL: It should be. Tell me about today’s agenda. What are we talking about today?
SO: Right. So I get the blame or maybe the credit for this one, right? Because we talked about this. As we are working through interesting problems with content and AI, what has occurred to me is that we really need to sit down and think pretty carefully about UI, about the user interfaces and what it means to deliver content in an AI world. And so, it turns out that Fabrice and I have some different lenses on this question. And so, we’re going to walk through a couple of different ways of looking at this problem. I want to set the initial framework of what we’re talking about. The first thing is that AI and the use of AI transforms the publishing paradigm. What I mean by that is that we no longer have this concept of I, as the writer, create the content, and then I package it, whether it’s PDF or HTML or anything else, and then I deliver it to you, Fabrice, the passive recipient consumer of that content.
We’re going to see that AI changes that paradigm in ways that are important. And what happens is that when I … Again, now I’m the consumer hero, right? As the consumer hero, I have the opportunity to request or demand or transform content in ways that meet my requirements. And so, that ability to produce what I’m calling synthetic content puts me in control as a user, which then brings us to the question of, how do we ensure that that end user experience is good in this synthetic AI mediated way? So that’s kind of our agenda for today to talk about what that looks like and start the conversation with you all around how we’re going to do this well.
FL: Sounds good. So maybe a few words about you.
SO: Okay. So I run a company called Scriptorium. We’re based out of North Carolina in the US, and we are interested in enterprise content operations. How do you create product and technical content, enabling content, learning content in such a way that we can then deliver it downstream to the end users in whatever format they might demand. So most of our customers are very, very large and have enormous amounts of content that they are managing, reusing, mitigating, translating, refactoring, and delivering in lots of different ways. And in that context, we work with Fabrice and his company. So I’ll throw it to you.
FL: Thank you. Well, it looks like both of us dropped the jacket today if I check the picture. Way more casual. Yeah. So I’m the founder and CEO of company called Fluid Topics. We also have to have vendor. And we take care of trying every day to reinvent the way content is delivered and consumed by those end customers, end users, and how they can make better access, increase searchability, better readability of that content. And I think that part of that challenge is how to integrate AI in this process of making and transforming the way content that is properly written is consumed. So very, very, very hot topic for us as well. So let’s get started. So I think you have an interesting perspective here that I’d like you to share with us.
SO: So I simplified down the authoring process to something fairly unrecognizable, which is basically as an author, I create content, which then gets dropped into some sort of AI compatible storage. Now, I’m not saying this is specifically a vector database or specifically a large language model, but as authors in an AI world these days, our job is to create content and then that content goes to a place somewhere. Probably it’s an LLM, right? And this is what, of course, that right and that lovely green arrow, there are untold hours of pain and misery and suffering in that, right? But at a very, very high level, as an author, you’re going to create content. And this might also be mediated or supported by AI tools itself. So I’m not getting into any of that. I’m just saying authors are going to create stuff. Maybe the author is a magic AI.
Again, it goes into the vector database, the database, the repository, whatever that may be. Now, when we think about this from the consumer’s point of view, it looks a little different because in an AI world, as a consumer, what I’m going to do is I’m going to ask the AI a question, and you hear this referred to as conversational AI or conversational search or just using ChatGPT or Claude or any of the rest of them, right? But I ask a question, and I get an answer, which is mediated by this AI compatible database, whatever it is. And then I get a response, which I read, and then I say to myself, “Well, that wasn’t quite right.” And then I change the prompt, right? I ask it, “Tell me more about this,” and it tells me. And I say, “That’s too much information.” Or, “Include this information,” or, “What are your sources?” Or, “I prefer this in French.”
I would not in fact prefer this in French at all, but maybe it comes out in French. And then I say, “Can I have it in English?” Which would be better, at least for me, very much better. But you can customize as you go as the consumer, and you can keep refining that prompt until you get what you want out of the AI. Now, we hope it’s accurate, and we’ll get into some of those issues downstream. But this is basically what the process now looks like for the consumer, which means that when we think about this author and consumer, they are now almost literally at cross purposes to each other because as an author, I’m feeding content into the LLM. As a consumer, what I get out is just a starting point. And I am no longer, as a consumer, just I want to say passively accepting, and it’s not totally fair to say that, right?
FL: It is.
SO: But if you think about it, if the author creates a PDF and I ship it to you, then you get a PDF, and that’s it. You’re happy. You’re not happy, but that’s what you get. Same thing with sort of traditional search, right? If I go into a portal with traditional search, it throws up a page of results and then I can click into those results, and I can say, “That’s not right.” And I can back up. But what I can’t do in the traditional non-AI model is change my output. I mean, sometimes there’s a little toggle or a little show me advanced stuff or we talk about progressive disclosure, little twisties, that kind of thing, but it’s very, very limited and critically, it is pre-planned by the author.
FL: It is.
SO: In an AI world, the consumer says, “I don’t like your output. I don’t want a PDF. I want HTML. You know what? I don’t like this HTML. Make me a podcast. Read it to me. Create a synthetic version. Scrape Sarah and Fabrice’s voice off this webinar and create a podcast in their voices with all of this.” Or if you’re reading the transcript of this, the transcript was auto-generated, right? So you read the transcript, and then you say, “I don’t have time for this. Pull out the three key points.” So the packaging and delivery process that as authors we used to consider to be the end point of the discussion is now the beginning because the consumer gets that, maybe it’s packaged and maybe it’s, again, dropped into an LLM, dropped into a vector database, and then the consumer says, “I don’t think so.” And they get to modify it and customize it to their specifications.
Now, the really troubling part about this as an author is we’ve been trying to do personalization for 30 years, and we’ve tried and tried and tried. And we’ve done some things, and sometimes you had multiple layers or beginner versus advanced or this and that. But again, it was all sort of prepackaged. We just said, “We’re going to give you three versions, and you can choose what level you are.” Or the system can say, “You’re a beginner. I’m going to give you everything.” But now we have this sort of extreme personalization mediated through the AI, which to me is a sort of be careful what you wish for because we asked for this.
We have been trying to do this, and we’ve been talking and talking and talking. And I put up puppies because I find this whole thing a little depressing. So here’s some puppies to help us through it, right? But we got personalization. It’s just that instead of us delivering personalization, the consumer is taking personalization, and it is a very, very uncomfortable feeling if you are the so called professional content creator. So Fabrice, I think it’s over to you to talk about some of the pieces that you’ve been thinking about in this context.
FL: You’re right.
SO: Oh, sorry. Yeah, one more.
FL: Go ahead. Go ahead. Yeah. Because you have this thing about the sandbox.
SO: This is the personalization issue, right? We still get to decide the size and shape of the sandbox and what we put in it, but what happens at that point is up to the end user. We have absolutely no control over how they assemble those grains of sand into their own personal sandcastle.
FL: Makes sense. Totally makes sense. I think that’s interesting because we have clearly … I mean, I fully agree with your perspective and how you frame this entire problem of users moving from a passive attitude to becoming active onto the content. I see that as well as an issue from a vendor perspective because it goes and with the UI we give them and the tools we give them to interact with that content. And maybe a way to control what they do inside the sandbox is by better designing the UI that gives access to that content, the window, the tools we give them to manipulate that content. And I think we fall short as vendors. And the perspective I have on that is if you look at AI, everything we have in mind is that’s what we all dream of. It’s got this J.A.R.V.I.S. in Ironman. It’s pervasive. It’s transparent, highly disruptive. You can talk to it.
And that’s the dream. That’s what we want. Something that helps us. And honestly, if you look at what we have, we have that today. I mean, it’s text. It’s done. It’s great. You still have to read a lot. I think we are still very far from providing the right UI and experience that people deserve for getting access to knowledge and better interacting with content. And if you look at what we do today, even ourselves, it feels like very much bolted on. It’s like whatever we do, we’re taking existing UIs, and I will show you screenshots. And we’ve put this AI thing on top without really rethinking the entire experience. And believe it or not, this is not an AI generated picture. I took it myself. That was in Florida near Orlando, and there’s a nice car. Let’s say that lobster is nice, so there are two nice things, but when put one on to the other, that becomes pretty ugly.
And I think that pretty well much describes what we’re doing today, this bolted on thing, which is we’ve taken our UIs, and this is screenshots from our own tech doc website. And we have a search result page here, and what we’ve been doing is just plugging onto its buttons interactions like, okay, if you want, we can summarize the result, or that very specific document that looks to be the right answer for you. Or by the way, there’s a chatbot, and here it is, bottom right corner. That’s where you expect it to be. But it’s just something layered on an existing UI without really rethinking the entire experience. And that’s the same for the reader. If you enter a document, we do the same. We’ve put some nice interactions like summarize or live translation of content by LLMs or machine translation engines. Now we can do this live thing.
So it’s super convenient, but again, it’s just added on top of an existing UI. And the problem becomes when you start having more ideas and say, “Oh, that would be nice because it’s a maintenance procedure, and then we could dynamically extract the parts listed in these 20 pages of procedures so that people know what part they need or what tool the needs or the pre-requirements for doing this.” Or if you’ve got a cut block, you want to put a button like convert the code into this language or run a test or explain this and blah, blah, do this, do that. It’s like the thing of five years ago or two years ago, it’s AI infused. It’s like we’re adding AI into the UI. It’s not infusion anymore. It’s saturation. It becomes clunky, unusable, and that’s where I think we’ve reached some limit into how we’re adding AI and helping the users.
And instead of helping the user, we are overloading them with things that become unmanageable, and we should be converging into something that it becomes more friendly, more useful. And if you look at our own tech doc pages, I’m criticizing what we do. Let us be clear here, that’s our own tech doc website. We have the classical search bar, and then we’re adding the chatbot on it. So why? And when you start opening the chatbot, you can ask questions. It runs perfectly. It’s accurate. We’ll solve that, but again, it’s still on top of. And if you click on one of those links here to enter a document, then you’ve got the document, but the chatbot is on top. So it’s a layer on top. So why? See what I mean? It’s like it’s not working for me. And it’s not just us. I mean, you can check other websites, and mostly people do that by, for the moment, adding chatbots everywhere on the website.
And I think that’s where some sort of a moment where we need to sit a bit and say, okay, everything we’ve done so far has been sort of a urgency mode, like we need to put it to show that we have it. But we have to now start rethinking the UI itself. And I’ve been working this past week with some people internally in the company to start imagining what it could be, what it should be, blending the AI into the UX, into those doc portals that we provide to our customers. And those are just examples, not telling that we are going to do it that way, but those are some mock-ups we did. And it’s like first, just one search zone, not two, not a chatbot, not a search bar.
And it should start with something that says type whatever you want. And you can see here this history zone where probably it’s like in ChatGPT, your past threads of discussion or your past things you did, not just the searches, but everything around the subject you wanted to do before so that you can get back to these discussions you had with the AI, with the UI. And whatever you type, probably it’s like imagine that the bot itself becomes the zone where you work. It’s not something that pops up, but that extends automatically you start typing. And maybe on the one side, you could have this dialogue that starts with something that’s more like an assistant and say, “Okay, considering what you typed, maybe you’re looking for this. I can advise you about that, that, and that,” and you can keep on typing here like a chatbot mode, discussion mode, while immediately you can have direct results.
And you see here, it’s not like you have a summary and then just reference links, but really having something more conversational and still direct access to documents because I still believe that experienced users, sometimes they want to get direct access to the doc. They don’t want a chatbot summarizing something. They want to see the entire procedure. So we have to balance this way of navigating, getting helped from the homepage initially. And you can keep on asking here and chatting, starting with your initial query, your problem, and keep asking. And imagine that here, the document list would refresh automatically as you start having this discussion, always being up-to-date to where you are in your discussion to provide you direct access to the content itself.
And when you go, I don’t know if you click on list or results, you want to go to something that’s more search result oriented, then the chatbot should follow you here, become an assistant and say, “Okay, you choose that document, but considering the discussion we had, that’s the answer you need.” Oops, that’s the answer you need. And you can keep on asking question, not just add the answer as you have today on many search engines, because you’re stuck. It’s just like, okay, I summarize the results, and that’s it. Here, it’s more like keep on having this discussion with the user, which where they could extend that zone, get more insight into the why and keep on having this assistant-like discussion. Or back to this homepage where you have this discussion, but links to documents. If you click and take access to a document, that’s where instead of laying over the discussion as we do today, probably your chatbot should become part of the UI and become an assistant more than a chatbot.
It’s not necessarily replacing the doc nor replying, but that’s where typically you would say, “Okay, show it to me in English. Show me the parts.” And that could become something from an assistant mode and not just put buttons here and there. See what I mean? So that’s probably the sort of things we have to start redesigning how we get out of this idea of a pure chatbot, which is something that is probably we have inherited from the ChatGPT thing, which again, is a very generic, is designed like this because it has to fit to any use case, any content, any user, any situation. So it remains very sort of middle of the road as opposed to a business situation we were in with the tech doc.
I mean, we are here to help customers achieve something, and we can drive them into being more efficient and mix and match the content itself as we do today with plugging that lesser chatbot more than a contextual assistance that follows you through everything you do from discovering to searching to reading. So I think those are just examples here I have try to make it more visual. Maybe there are things that are better to do or different to do, but again, it’s like we’re opening and we are discovering there’s a new road to a new path to be walked, which is rethinking entirely the sort of user experience we need to design when it comes to blending content and AI into helpful UI. So that’s a perspective I have at this stage on what we need to do as a vendor at least.
SO: Yeah. And I think it’s important to point out that your responsibility is for organizations that are setting up content delivery on their own websites with a content delivery portal, content delivery platform, as opposed to, there’s a whole other question around how do we feed into ChatGPT or the others and make sure that they perform? And largely, what we’re talking about here is the scenario where you are the content creator and owner and then have some influence over how that content is going to be presented on your corporate platforms, whatever those may be.
FL: Yeah. But you’ve got this situation where you mentioned where the content not necessarily goes into your own portal that can go outside of your portal as well, which means that anyway, you have to prepare your content for that, no?
SO: Yeah. So I wanted to step back again a little bit from the delivery piece, and you’ve sort of got the end state, right? We can get to this point, and it’ll be great. The number one problem that we’re facing right now in doing AI enabled anything is that the content that’s being scraped into the database is, to use a technical term, garbage. Not all of it. Not all-
FL: Not all of it, but…
SO: Not all customers, not all, but what’s happening very often is that the AI team is not connected to the content team. So the AI team is an offshoot of engineering or maybe IT, something like that, and they are not talking to the content team. What they’re actually doing is going to the organization’s PDF repository, which is almost certainly SharePoint, and they’re just scraping everything that’s there and dumping it into an LLM, an internal company developed system. And the problem with that is that you probably have files up there that are like Version 2 PDF and Version 3 PDF and Version 2 updated and Version 2 Final, Final, Final, Final updated. All of that gets just brought into the LLM, which has no concept of versioning or metadata or anything else. And so, you have this real problem and the solution, the canonical solution to addressing this is infamously humans in the loop, but what we really want to do is fix that ingestion point and get better content in.
And this is where, as somebody who works in structured content and with metadata and taxonomy and all those other fun things, we’re telling people, you have to do the work. You have to do the work upfront because once that ingestion step happens and they’re ingesting not structured, not consistent, not governed, not accurate, not up to date, then what chance does the AI have? The AI is not going to make your content magically more accurate. It’s not magic. I mean, it can do some magic looking things, but it is not magic. Your entropy always wins. Your content will always sort of degenerate, right? So you start for your best possible, and it goes down from there. So what’s the best possible thing that you can get into your database?
So I just want to put this in here to say these challenges that we have with good outcomes via AI start a long time ago with your actual content authoring and content debt because so many of us, I don’t know, all of our clients, I think, but most of them certainly, have content debt, and some of them are drowning in it. So that is a big, big concern. And if you’re sitting on the content side and saying, “Well, they didn’t ask me,” then yes, that’s 100% a thing that is happening. And my advice is to go find out who these people are and make friends with them.
FL: Makes sense. The content depth, I mean, even a company like us, we have it. So I can imagine global companies that go through merger acquisition and have legacy product, and I’ve been around since 50 years, how much that is. I mean, it must be crazy. So you think AI can help and what you see as using AI for enhancing your contents, like you have this idea that you can still use AI for tagging content, for example?
SO: Oh, absolutely. You can use AI tools to remediate content and make improvements. That’s a different problem set from what we’re talking about here, which is how do we use AI and how do we build end user interfaces for AI? But yes, you can absolutely use AI on the backend to find problems, start to remediate them, introduce consistency, terminology, taxonomy, all those kinds of things, but you have to really assess and remediate your content. You can’t just dump it into the AI and expect it to perform. So it’s sort of like, you have this bucket of content. You have to fix it, then you put it in the AI, then you maybe have a shot at delivering. And so, having seen your very concrete examples of where UI might be going and how you might integrate in a Fluid Topics or something similar, what I have here are some very, very high level, much bigger picture sort of thoughts about what an AI user experience might or should look like.
So one of these is … Oh, sorry, go back for one, just for a second. This one, what I’m basically saying here is that if you output the official content, the stuff that got authored by the organization and is their official doc, then it gets a logo. And you’ll notice we put both Scriptorium and the Fluid Topics logos on the stuff on the left because that is official, vetted, approved, reviewed, packaged. Maybe it’s a PDF, maybe it’s something else, but that is the content. The thing on the right, and interestingly, I mostly see this for images. It’s pretty common when people are in news coverage, you will see an AI generated label on images most often when they’re talking about deep fakes. They slap an AI generated on it to make sure that people don’t think it’s the real thing. But I think that if we are generating a synthetic summary or an AI or a machine translation, it should say this was machine translation. This is AI generated.
So I’m just saying that when you generate that summary or that thing that is AI driven as opposed to being the official thing, it should have a label on it so that people know. And I think the word that we’re going to see increasingly here is provenance, which usually is like an art history term. What is the provenance? Is this a forgery? Is this a real painting? Was it Rembrandt or not? Was it Monet or one of his students? That kind of thing. And so the provenance of the content, am I as the content, the corporation, the big company who is producing this content standing behind this with a logo and saying, “This is mine.” Or am I saying, “Well, you got it out of my chatbot, and I’m going to put a disclaimer on it just in case because I can’t be sure.”
So that’s one. The next one I have here is a sort of, let’s see what the next one is. Oh, right. So the synthetic and the original side by side. What I’m kind of envisioning here, we saw this years and years ago actually on Microsoft, on MSDN. When they started doing machine translation of technical articles, KB articles, what they would do is they would put up the translation and say, “Here is your translation in Italian or Turkish or Chinese or whatever, but this is a machine translation. Here is the original in usually English.” And so, what they would do is they would put those two things side by side and say, “You asked me to refactor this content, and we did. We translated it, but here’s the source.”
So in your example where you’re showing the output from the LLM, it would say, “Okay, here’s what you asked for, but then also here’s where it came from.” This is the original document where you said, “Please summarize,” so that I can look at that summary and say, “Well, that’s not enough,” and go back to the source. A different version of that is kind of this next one where we’re moving away from displaying the source document, but rather we’re saying, “Here are the citations. Here are the links. Here’s where I’m getting my information.” Now, the problem I’ve run into with this one with public facing models is that sometimes the links are invented. You start asking questions and then it turns out the links don’t exist.
The other day it actually invented an entire competitor for me. It was telling me all about this company that did content ops consulting that I had never heard of, and I dug and dug and dug. And well, it’s a loose consortium. It’s run by these people. I know the people that it referenced. They are not in fact running this non-existent company. I actually ran into one of them at an event in person and congratulated them on their new consulting operation. They were very, very puzzled. So then I emailed them the actual output from the LLM where it said, “So-and-so is now running this,” which he is in fact not.
FL: No.
SO: So links, here’s the AI output. Here’s where I sourced it. These are my citations. I think we’ve all heard about the problems in the legal world where these citations turn out to be invented, which is a huge problem. But I’m thinking more, again, in a controlled environment such as what you have, you can say, “Okay, when you surface new content, when you synthesize content, tell people where it came from, what were the primary documents that you’re pulling this from?” So that then I can read the AI output and say, “Well, let me go back and read the original and see if I got … Maybe there’s a piece, a little nuance that’s missing from what I wanted.” And then the next step is essentially this, but this is my dream, this last one. What I would like to see is the ability to highlight in the output a particular sentence and have it tell me where it came from.
Now, that’s not exactly how an LLM works, right? It’s not necessarily pulling a sentence or a chunk directly, but I have found that sometimes you can recognize wording. I’ll be reading something that it generated, and I say, “That sounds awfully familiar.” And I can find it. I can find the source document where that particular sentence or that particular phrase occurs, and it got surfaced in the LLM for whatever reason. But I would really love to be able to see accountability. You highlight it. You get traceability. You can say, “It came from over here. I got it out of this document.” And then I can go read that. Let’s say there’s a step here. Let’s say I made a procedure. I highlight the step. I go back to the original. I look at it, and I say, “Well, that’s not actually what that step said,” or maybe it is. The thing that keeps me awake at night is what if it’s wrong? What if it’s wrong? And it’s high stakes content.
FL: I can think of things. It’s pretty sophisticated what you described here. You have this AI output. And when you… without doing nothing, it’s highlighting parts of that. And dynamically, you see on the side the sheer text, the real text that was used for that, even though it’s not the exact phrase, but means that you have to reconcile the content with the AI outputs in dynamic way.
SO: But you can do it, right? You’re super smart.
FL: These phrases come from that entire paragraph, from that document, even though it’s not the exact same. I love that. Then you have full accountability, and you can check rapidly. Wow. Okay.
SO: But you can do it, right? You’re smart.
FL: I don’t know. Challenge accepted.
SO: Maybe next week.
FL: Maybe next week or month. No, no, but I see the point. That’s pretty cool. Whatever, because you want to…
SO: To be clear…
FL: … look like that.
SO: Yes. To be clear, I built this mock up in Canva, and your graphic designers helped out with some of it. There is zero code behind this, right? This is just, I made a pretty picture, which is super easy to do. And then I can just say, “Hey, Fabrice.”
FL: Do me that.
SO: You can. It’ll be great.
FL: In two years. No, but I see the point. And I think that’s pretty interesting to have this idea that you can almost just without doing nothing move over some AI output and dynamically on the side, you can show and as you move the [inaudible 00:38:11], suddenly you have another paragraph from another document, and you can rapidly check. I like it. I like this idea that you don’t … Because you know as RAG works, usually even if you do RAG, which is the more controlled way, you send over to the LLM before that generates this AI, put something like five, 10, 20 pages, 40 pages of content, and the LLM reads those 40 pages, 20 pages of content which comes from different topics, usually different chunks on your content.
And you’re right. You don’t know exactly how the LLM will read those 20 pages and rewrite something based on that and being able to trace back every assertion generated by the LLM to the pieces of content that probably were used to get to that assertion of the LLM would make sense. And then you would have full both provenance and accountability and traceability of the LLM output. Makes sense. I like it.
SO: And to be clear, you don’t need this for all your content or for all content types. The place where this type of thing is going to be necessary is in environments that are regulated, that are high risk, that have impact on health and safety, because that’s where this type of thing is going to matter. If your content is about a product that is less dangerous or less potentially dangerous, then you don’t have to be that careful.
FL: Yeah. That reminds me, that rings a bell to me because as you said, when you start entering situations where safety matters or precision or accuracy matters a lot, and you challenge me about other UI ideas, and that led me to this thing, which is if you take something like I took something from iFixit here to make it. But imagine that you have a complex maintenance procedure on a jet engine or something like that, that is about 40, 50, 100 pages, something to be done in a very precise order with all the safety warnings and everything and the checks and all that. So it’s oversimplified because it’s an iFixit, how to change a screen from an iPhone, and it’s already very much stepped and well-framed and illustrated. But we know that many of the documents that we see are not that well documented or structured, that iFixit, I must say.
I think that’s almost a… You see what I mean. But the cost of turning 40 pages maintenance procedure into something as documented as iFixit would cost probably 10 times the cost because you have pictures of everything done. And probably as well, what you see here is made for people that are not knowledgeable. And when you write content for professionals, probably most of the details wouldn’t be relevant to professionals because they know how to remove that part or remove the cover or whatever. So you try to focus on the things that are more specific. But still, I think that’s interesting because you can easily mainly make it your own based on your product and where you have those long procedures. How can we help customers, technicians, do this without making sure that they got the exact information as it has been written?
So because we’ve done that in our web portal. We’ve got a maintenance guide, and we have actually this AI infused, AI thing where we had a button which says extract parts and tools. And then we let the LLM read the procedure and dynamically generate the list of the parts and all the tools that you need for executing that, which means that you don’t need to maintain this list yourself as a writer because they are dynamically generated, which is easier. So we get that, but I feel that at some point we started thinking about how can we turn something that is more like a long procedure, long text to read into something that is more step by step and guided.
And so that we make sure that, for example, you could imagine that when you go from, “Okay, drive me step by step through this entire procedure.” When you click launch, instead of having the user, the reader scroll through, which is always dangerous because if you scroll too fast, you can miss something, and it’s more [inaudible 00:43:07] and you could start, for example, by, okay, guys, first you need to validate that you’ve read the safety thing, which is maybe three pages above or come onto three procedures that people wouldn’t see. And can we bring that into the UI and force people to say, “Read that, be sure that you have secured yourself or turned off the device or whatever.”
And then maybe let’s say now you can check the tools. If you want, I can extract the tools and make them make sure that the technician has said, “Okay, I got everything. I got everything in my … The parts and the tools, I prepared my toolbox. I can move on to starting, or I have taken from the warehouse all the parts that I need, put that in my truck so that I have to drive back to the warehouse twice, three times before I’ve forgotten something.” And then maybe you can, as I said, move to some thought of UIs that go step by step automatically, even though the document has not been clearly designed as such, which is more text-based and continuous reading, and then you can get people through all of this.
So I think that’s many ways we can imagine, and I don’t see AI being a replacement of everything. Back to what you mentioned initially on this sandbox, because in many domains, AVD are lightly regulated. Any transformation of the content is a risk. I mean, if a procedure has been written in a specific way, every word counts. Every word matters. Every step matters. And if you start allowing users to change that content or put this AI thing that allow them to redo whatever they want with the content and then use that content that is rewritten [inaudible 00:44:59] risk. And I think that another way of looking at it is, okay, can we use AI in ways that help create those new ways of navigating content that are helpful to the users for being productive without creating this danger and this gray zone where they start being at risk or the company starts being at risk as well.
SO: And I think these are really interesting because the parts list, I mean, as you were looking at this, you were saying, “Well, we don’t want the author to have to create the parts list.” So instead we scan the document and generate it, which I absolutely agree with, except as you think about this, maybe the answer is that the author has that parts list generator, right? So as their last step as they’re authoring, or even as a supporting tool that says, “Hey, you added a step, and it looks like there’s another part in there,” that the backend authoring system would add the parts list and for that matter would generate the summary, right? So then I validate those, and then I ship it. Now, if I push summarize on the front end, it doesn’t go to the LLM and summarize over and over and over again, right?
Because if it’s just summarize that page or summarize the parts list for this procedure, that’s static information until the procedure gets updated. And so therefore, we should do it once on the backend and publish it and make it available. So we’re still AI enabled. We’re just not doing it on the fly per user, but rather on the backend per document such that that intelligence is then available just right there, but it’s been pre-produced essentially.
FL: But it’s controlled. And then it’s controlled as well. So you can provide the same features, but pre-executed, validated in some ways, and then made available to the customer that way. That’s interesting. I like the idea. I need to think about it, which brings us to this complex zone of liability, risk, edits management. And I think you had something to share about that as well.
SO: Yeah. This is a little bit sideways from everything that we’ve been talking about, but I think that from a risk point of view, risk and regulatory, the systems and the accountability and the guardrails that we should be building for AI-driven systems need to match the risk of the product, of the content, of the thing that we are documenting, right? So clearly the higher the risk of the product operating and of the product being operated incorrectly, the more careful we need to be and the more guardrails we need and the more risk mitigation we need to do. I always use video games as the example of the thing that is lower risk, but in fact, there are a couple of issues there.
And the big one is that because video games are, the content effectively is the product, right? I mean, you could look at the game and the code, but also the story and the interactions and the voice acting. Because the game is the product and people pay to get content as part of that product, they are very, very interested in that content as art, not as just a thing that you, the video game producer, produced as fast and as cheaply as possible using AI. So there’s been a lot of pushback on AI in video game content. From our point of view, mostly on the corporate side, it’s low risk, right? If it’s wrong, nobody … Well, okay, your character dies, right? But it’s good. But from a commercial point of view, there’s business risk that people will reject your product because they don’t like what you’ve done to the content, whether it’s voice, audio or video for that matter, or text or anything else. But I want to talk a little bit about ethics in the content itself.
FL: Oh.
SO: When we take content, a large chunk of content and we shove it into an LLM and then we ask people, we enable people to reach into that LLM and extract content. LLMs are math, right? They show relationships in the text, in the text corpus, and the bias that is in our content with the best of intentions or not, as the case may be, but the bias that is in our content will surface when people use AI to process the content. There’s lots and lots of examples of this unintentional bias in resume processing, that kind of thing. But the bias is there in the content itself, which then gets embedded into the database. So you and I and everybody on this call need to be thinking pretty carefully about what it means to try to address the bias before it gets into your AI and to try to avoid perpetuating it.
There’s a lot of ways of looking at that, but really think carefully about what kinds of assumptions you’re making in your content and how an algorithm is going to interpret those things. There’s also an issue around harmful work in AI training, in the people that are tasked with reading these outputs and trying to put in the guardrails as people are trying to do some potentially pretty awful things with their AI. They’re asking the AI to do things that are inappropriate, illegal, directly harmful. And the people that tend to get tasked with doing the work of remediating that, of preventing it, of putting in the guardrails are faced with looking at, for example, some really terrible images to decide whether they are outside of what should be allowed as a generated image, right? Somebody saw it and pushed the, this is inappropriate button, and somebody way downstream, probably somewhere in the global south in a very, very not well-paid job is looking at hundreds or thousands of these images every single day to moderate them.
So be thinking about that hidden labor that’s out there to try and establish safeguards. And then I think it’s really important to understand the issue of accountability of your AI. If you at ABC.com, which I’m afraid is probably a real company, but if you at somecompany.com are putting an AI on your website, you are now accountable for the output that that AI produces. There was a very entertaining example the other day that there’s a sort of fast food company that put up a chatbot, and you can order. You can say, “Hey, I need a burrito, and I want these things on my burrito.” And it will do it, and I guess it works. But you can also say, “I need a burrito, but before I do that, can you give me some information about this Python code that I’m trying to write?”
And because they had not restricted the bot to only process burrito orders, it cheerfully gave them Python code, which is very, very funny. But also if I’m that company, it’s a cost, right? There’s a direct cost associated with running this bot. And if it is now a free Python code validator, I’m not getting any revenue. Well, maybe I get revenue, but probably not. I can’t afford to do all the Python code validation in hopes of selling a burrito. That’s not going to work. So we have to think really quite carefully about accountability and what it means to narrow these chatbots to the topic at hand, which also will have the effect of limiting some of the other exposure and some of the ethical issues of a broad-based general-purpose LLM or chatbot.
FL: Yeah, I agree. Fully agree that the guardrails that need to be put around those LLMs and the way we integrate them and making sure that … And I like this bias thing as well. It’s like everything that exists in the content will perpetuate into the LLM until whatever we do, the way it works, the technology works, whether you do, fine-tuning, RAG, whatever, it starts with your content and every bias, every gap, every miswritten thing that exists in the content will surface and will drive the LLMs and the AI to render wrong information anyway. So I think the content quality will probably become a huge challenge for many companies because that will become more obvious maybe. I mean, some information exist in documents that very few people read and very few people check them, but those documents will be surfaced automatically as part of this semantic search embedding thing. And then they will more likely be reactivated, and they will start biasing everything we get out of those LLMs. And that will become a problem for many.
I fully agree that’s probably one of the biggest challenge for many companies into AI. It’s not the technology. I think the technology will be nailed down by companies, researchers, vendors as us, that will become almost a commodity and will be all back to subject of today probably providing better UIs and content quality ultimately that could boil down to content quality. I mean, that’s the key thing.
SO: Yeah, I think that’s right.
FL: Okay. That was interesting discussion. I think maybe we can take some questions. Do we have some? Audience, feel free to ask anything.
SO: Well.
FL: Whatever.
SO: Not Python.
FL: Don’t about burritos, but you can ask about…
SO: You ask me about Python, I’m going to send you to the Chipotle bot.
FL: I’m sure you will reply about Python code, so will I. Oh my God, maybe not. Do we have questions now?
SO: So Fabrice, these examples, while people are typing, these examples that you are showing, I’m not going to ask you to publish your roadmap, but what’s your thinking in terms of this type of integration? And are we talking months or years or decades?
FL: No, no. Clearly it’s about months to maximum one year because I think that, well, you know me from quite a while now, it’s like, got a sense of that will rapidly become … It’s not like our customers come to us and say, “We got a problem,” but we can feel through the feedback we got from the UI integration that they start saying, “Ah, but would it be better if … ” And probably they don’t ask enough, you see what I mean? For the moment, they’re still thinking inside the box, not outside of the box. So it’s more like, can we make this better? Can we make that better? But the can we make this better, in fact, reflects on a feeling they got, something itching them somewhere, that it’s not the right way to do it without them knowing exactly what they want.
It’s like white page, blank page thing, and I’m pretty sure that the moment you start presenting, as I did few mock-ups like that, they will start saying, “Okay, we need this. We need that. We want it now.” So that’s why you ask when that will that be? I think that’s something that we need to deliver as part of the entire product within the next few months, I’m pretty sure, I think. Yeah.
SO: Yeah. And we’re seeing a lot of interest in this, but as I said, as a consultant, I just make pretty pictures and say, “Can I have one please?” And there’s more underneath it. But for me, the problem that we’re hearing over and over and over and over again is that the wrong information or the not up-to-date information is being ingested, and unwinding that is going to be highly, highly problematic. There’s also, I think, a distinct lack of understanding, in some cases, willful lack of understanding of what the limits are of what AI can do, right? That you still have to put in the work to make the content better.
And of course, we can use largely generative AI on the backend for certain kinds of productivity and automation things, but you just cannot create something out of nothing. And then I hear, “Oh, well, we’ll just use the product specifications and render the documentation off the product spec.” Well, that’s amazing if you have a product spec that’s … Well, first of all, if you have a product spec, let’s start there, but is it accurate? Is it up to date? No, it’s not. You know it’s not. Come on. Oh, we’ll just generate it out of Jira. No. Jira, no, that’s just a collection of people thinking out loud.
And so, I mean, yeah, if your product spec is pristine and if your governance is really good and if your Jira is really clean, but now we’re just pushing that moment of the tech writer or the content creator taking the content or taking the starting point and creating knowledge, creating content. We’re just pushing it upstream. So at that point, you’re going to need tech writers or writers, content creators, way in the back on the product spec. We’re just deferring or we’re accelerating the moment where you recognize that you don’t have docs.
FL: We got a question for Sarah. So Sarah says, it’s very interesting, rather than people walking through a guide, isn’t the next logical leap that the product itself becomes comes AI driven. I mean, the product asks what the user wants to accomplish and then helps them do it. It’s like the product becomes the UI, in fact. It’s like you don’t need the UI. The product is the UI. It’s the product plus the UI of the product for how to fix the product. You know what I mean? I think that’s valid. That’s interesting. It means that the product has to be connected. That would increase the cost of the product because the product has to embed a computer or something that is displaying something or answering to the question. So it depends on the type of product.
SO: It depends on the product, but if you think about software broadly and like a software UI, a user interface, and I go and I click around and I do things. And then compare that to a command line, tell it to do things on the command line. Ultimately, if the product is AI-driven, what you’re really doing is just going to the command line. Now, it can be sophisticated, and you can do some cool things around that. But what you’re essentially saying, and I’ve got a couple of products on my computer right now that do this where, for example, I have a tool that does reporting, and I can go in there. And I can build a report and pick a data source and do this and do that and do the other thing. But the other thing I can do is say, “Hey, make me a report that does this.”
And I write what I want, and it extracts meaning from what I wrote and generates a report or a draft report that I can then continue to modify. But to your point, the implication is that the bottom line is this is exactly the same as the problem we have with structured content and automating delivery into PDF or HTML. The intent has to be in the document or the software or the experience, the interaction. And if you can capture that intent precisely, then you can tell the machine to do a thing.
But the tricky part is how do you capture … When I write a query or a request that says, “Make me a report that gives me A, B, and C,” I almost certainly don’t use exactly the language that the tool is expecting, so it has to interpret that. And what was I actually asking for? And that’s where you get the friction, and you get the problems. And that’s why we have UIs because it is in fact easier to look at a page with all your options available to you and click the various things because now by clicking, I am pre-interpreting my intent, right? Because the software says you have four options, which one do you want? Instead of me typing in a bunch of stuff and the software is saying, “Well, there are four options, so what is she actually asking for?”
FL: I like the way you depicted it. We move from, particularly for software, we move from common line interface, which were very nerdy to graphical interfaces, to simplify the setup, structure everything so that people could see it. And then maybe there’s a next gen UI for software configuration where you just say, “I’d like to do this.” And it’s like even the UI, the graphical interface would be a fallback to that if not disappearing at some point.
SO: But how do I know? How do I know what I can do?
FL: To do for software, that’s for sure. Yeah.
SO: Yeah.
FL: That’s an interesting, that’s a valid point. It’s like at least looking at it for software. And what you’re mentioning is right, I saw that I read an article this morning where we see that with BI, business intelligence software that have moved from on prem to SaaS to cloud, and now, it’s like they’re wondering whether or not … That’s exactly the example you mentioned. It’s like people just want to generate a report on that or how did these things evolve over the past six months, the sales per region or whatever? And people don’t have to learn SQL or whatever query language, whether it’s technical or made more simple through drag and drop or whatever. So it’s like talk to the BI system, and the front end of the BI system becomes just talk to me.
SO: But it’s a fire hose, right? Because you could do anything. And so, one of the interesting things about this is that they also gave me something like 40 default reports. So you can just click a report, and it’s just there because they pre-built it. So they gave me a bunch of templates. They gave me a UI, and they gave me an AI option. And I think that’s really the critical thing is if I don’t know what I can actually do with my reporting, I’ll never find the right stuff. And this is true for content as well. A UI is arguably a way to help people surface the information that they don’t know that they need or that they don’t know that they’re looking for.
FL: I’ve got another question from Kyle. Where do you think generative AI’s influence can be detrimental in technical documentation creation? So here we are on the backend side, which is … I think it’s on your side, Sarah, maybe you’re better than me to reply to that one. What do you see? And I guess you have this discussion with your customers, the companies you’re helping, and maybe the tech doc team that say, “How much should we use GenAI for generating content or writing content or rewriting content?” What’s your view on that? I can start because I have a very naive one.
SO: Well, what’s yours? Yeah.
FL: Everything is a pendulum, swinging pendulum. For the moment, everything will go far. People will say it’s replacing everything. It’s going to be magical. Then the pendulum will swing back and say, “No, it’s not working.” And then at some point, the pendulum will go to someplace where it’s useful. It’s not detrimental maybe, or maybe it is, I don’t know. So maybe to you, where are we today and where do you think the pendulum will stabilize, and how can GenAI be used and will it be detrimental or not?
SO: So first of all, I’ve spent my entire career on automation. How do we automate things? How do we get rid of busy work in order to automate? How do we use templates so that everything will look consistent? How do we do structured content with multichannel magic output to five or six or 17 different places? How do we do reuse to audit? It’s all automation, right? So when you look at AI, which is also effectively automation, the big thing that we have to keep in mind, which comes up over and over again, is that GenAI and AI in general is pattern based and probabilistic. It’s based on probability. One plus one is not always two. Whereas, we also have what I’ll call traditional coding, which is deterministic where one plus one, or if you do A plus B, you will always get C, right? That’s literally the code says A plus B equals C.
And so, the thing that you have to decide is what is the appropriate place to use patterns and what is the appropriate place to use deterministic code? Summarization is a great example of this. If I have a longish topic and I want to have a summary of that topic, generally something like AI will do a better job of that. You can’t go in there and say, pick off the first sentence of every paragraph, and that’s my summary. That just doesn’t work, right? But a more fuzzy approach to it works quite well. Now the question is, where and how does this become detrimental? The AI is only as good as the content coming in, as the content that’s feeding it. And that ultimately is where this thing breaks down because in general, the content coming in is not perfect, and sometimes it’s not even good enough.
And so, when you take not very good content and then you build an entire pattern-based thing on top of not very good content, some really, really bad stuff is going to happen. So yes, we can automate, and we can use it for productivity. And we can do things like generate a first draft of my output or generate a framework or an outline, and then I’ll fill that in. And if your content is ultimately very similar to an existing document, then it should work most of the time if your documents are pretty clean, right?
It should work. But the thing that makes it problematic is that if we focus only on velocity, on how fast we can get the content out and not on is this actually the right content, we run into big problems because now we’re producing content at scale, and it’s not necessarily correct. So I sort of go all the way back to the beginning of how good is your product spec, right? If your product spec and your PLM, your product lifecycle management systems are pristine, then yeah, you can pull a lot of stuff out of there and do it really well. Now, arguably, not necessarily GenAI. If I’m doing a data sheet and I need to pull all those product specs out of my PIM or my PLM, that’s probably a script, not an AI, right?
You only need AI if it’s fuzzy and weird. And then the other thing I’ll say is that, and I would encourage you all to try this yourselves. Go into your chatbot of choice, one of the public facing ones, and ask it a question about something that you don’t know a lot about, like particle physics or I don’t know, care and feeding of dairy goats, right? Ask it a question about something that you’re just really not an expert on. You’re going to read it, and you’re saying, “Oh, this is not bad.” Then I want you to go in and ask it a question about something that you are an expert on, knitting, quilting, woodworking, piloting private planes, I don’t know, whatever your favorite hobby is or a work related thing, but ask it a question about something that you know a lot about.
And then, it’ll give you an answer, and it’ll say, “Would you like me to give you more about this,” or would you like me to something, right? What I’ve seen in doing this is that the initial answer is usually pretty good. That first summary is usually pretty decent. And then as I start asking more questions and drilling deeper into the details of the thing I’m asking about, it gets wrong, and it’s more and more wrong the more detailed you get. So I was talking to my hairstylist about this, and she said she’ll go in there and ask it for various kinds of techniques. And initially it looks okay, but then when you ask for more details, it’s wrong. And wrong in ways that would cause bad things to happen to your hair, right? So we don’t want that.
Really understand that at the surface level, it’s probably going to be pretty good. And then as you drill down, it’s going to get problematic. And that’s the thing that worries me about it being detrimental because technical documentation, enabling content at the end of the day is about the edge cases. If you type in a name that has an apostrophe in it, this system will not work. Needs to be in the documentation for a surprising number of tools out there in this year 2026, right? That’s the kind of thing that, that’s what adds value to documentation, to content, to say, “Oh, by the way, you can’t put this character in.”
FL: Okay, I get it. I think, by the way, I recognize the problem with the LLM and the, how to take care of my hair. Maybe I trusted too much in the LLM and AI. No, but what you described is what you said, probably I can relate to it quite easily when it comes to will AI replace developers? That’s the same question as will KodeKloud replace the developers? Will GenAI replace the tech writers? I think that’s the same. For me, as a vendor and we try it internally, the answer is clearly no. We use AI for enhancing the productivity of the developers. Typically, doing chore things and things that are boring needs to be done like writing the tests. For every line of code or every module we write, we have to write all the tests. It’s pretty good because you give the code to KodeKloud and say, “Write the test.” And then they can write thousands of tests that you then integrate into your CI.
And that’s pretty good. And usually it’s less biased than people, so it’s writing more comprehensive tests. Or you want to refactor the code, or you want to upgrade to the new version of the library. So it’s going through the code and say, “Okay, we need to change those parameters.” So it’s not replacing. It’s doing the boring work like generating the summary, or as we talked about earlier, generating the parts catalog, the parts list. So you don’t have to read through the entire 20 pages yourself and say, “Oh, this part, copy, paste, copy, paste,” and maintain the table by yourself manually. So I agree with you that that’s the same. All the tests we’ve run through with KodeKloud for the moment, there’s any chance, barely a chance it will replace the developers. It’s just making them more productive and letting them focus on the more valuable things part of their job.
They use it sometime. They use it to help write the code, check the code, check the merge requests, but for the moment, all the tests we’ve done, which is writing the proper code, it’s not that productive. The productivity gap is not there yet, and there’s too much risk of generating non-optimal code the same way that if you use GenAI for just you send the specs and you say, “Write the doc,” There’s a huge chance that the written documentation, there are missing parts. It’s becoming too verbose or boring to read. So I guess that no, it’s not replacing.
So for me, it’s not like whether the question should be, is GenAI detrimental to tech writing, is it detrimental to developers? I say no, embrace it as a tool for making your job more fun, to do the part of your job that you don’t like and that can be easily automated through AI, but not be blindly trustful about the output of AI, just put it there to write the doc from the specs. I don’t believe in that. And regardless of the quality even in five years, because there’s a human factor in it that needs to stay there.
SO: The fundamental problem, the fundamental reason that GenAI is such a huge threat to technical writers in general and content creators is because the people paying the bills in many cases do not understand that there is value being added by the humans as they’re writing content. They see it as, like you said, “Oh, well, we wrote the code, and now all we have to do is write the tests and generate the documentation.”
And if your code is perfectly commented, which I’m sure it is 100% of the time, right? Yeah. You can generate your API docs to a pretty decent level of specificity, but why do I use this API, and where do I use it? And what are some use cases? And that context that the humans are adding and that understanding is what is being not captured in many cases. And so, I would say it’s detrimental because once again, and this is like the story of our life, right, people do not understand what it is that a technical content creator actually does.
FL: Yep. I agree. And that’s exactly that sort of question remark from the audience says, interestingly, companies are still evaluating tech writers primarily on the writing ability. So what does that tell us about how far behind talent strategy is from a professional reality? I think that’s right.
SO: That’s exactly it.
FL: You can use that for writing. For me, probably there’s a bit of a shift there that’s the skill or the primary value of the tech writing team shouldn’t be about the English writing styling capability, but more about the truthfulness and the comprehensiveness. Finding the gaps, spotting the gaps into what’s written, as you’d said. Why do you need this API? What are the use cases? See what I mean? Then you can write good English, but you can always use LLMs to distill your English and make it better. That’s true that probably the skills required and the way to evaluate tech writers should evolve, but AI won’t replace the tech writing team tech writers because someone has to be the guardian and accountable for the truthfulness of the information. And AI can’t be that guardian.
SO: Right. And you can produce infinite amounts of content out of the AI, but you do not have infinite humans to review it, which means you have to worry about getting it right as it comes out. You can’t just throw bodies at the problem to fix it after it’s generated. We have to fix the input problem.
FL: Yep. Okay. I think we have exhausted the questions. It’s been an hour, 20 minutes. It has been a pleasure, as always.
SO: Thank you.
FL: Trish, any last input, comment, remark?
SO: That was very fun.
PG: Sarah, I was thinking of your comment about making the argument to the people who write the checks, and that’s so important, right? They’re the ones we have to convince. I mean, we all know the reality, but it’s the perception that we have to work on, but this has been an amazing talk. Very, very informative, great, great as always. So just a quick reminder of recording we send out to all the attendees, as well as our past webinars page of the CIDM website. So with that, any last thoughts for you guys?
FL: Nope. Has been a pleasure to have this discussion.
SO: Thank you. Thank you for Fabrice, and thank you, Trish.
FL: Thank you, Trish and Sarah. So have a good day then, and see you next time. Bye-bye.
PG: Until the next time.
SO: Bye. Bye everybody.
FL: Bye.
Stay updated on future webinars by subscribing to our monthly newsletter.
About the author
