How a Construction CEO Built an AI Product: Henry Chart of CFS

Charlie Cowan: Let me share my screen. Screen sharing. Okay, good stuff. Move this out of the way. Well, good afternoon, good morning, wherever you are dialing in from, and welcome to this edition of the Kowalah Wednesday webinar. If this is your first Kowalah Wednesday webinar, we get together every Wednesday, same time, and spend an hour talking about something that is top of mind in the world of AI. So it might be something that we are building or using internally. It might be something that we're seeing out there in the industry or with clients. And today we'll do some introductions in a moment, but we've got one of our special spotlight webinars. And we have a special guest, Henry Chart, who will be walking us through. Now, a little bit of housekeeping. Like I said, we're going to spend an hour together. If you're watching live, then you're in a Zoom event.

Charlie Cowan: And, depending on what device you're on, up at the top or the bottom of your screen, you will see that you've got a Q&A panel, so you're able to ask questions in there. And then there's also a chat panel, and I'll do my best to keep an eye on that so you can ask questions as we go through. We are going to follow our normal format, so we're going to spend probably 40 minutes or so talking about our main topic of the day, which I'll introduce in a moment. And then we have a little recurring segment where we talk about AI in the news. And there's a few little updates that have happened in the last week, or big updates, depending on how you look at them. So we'll touch on those. And then there is an opportunity for those of you that are watching live to ask any questions.

Charlie Cowan: What I would say is that the discussion that we're going to get into with Henry, I think we're going to go deep on a number of topics. And if you've got questions that come up during that, then don't feel that you need to wait to ask your questions. So just use that Q&A panel or the chat panel, and I'll try and keep an eye on it. And also my colleague Caitlin, a human, is on the call as well, and she'll keep an eye on that and let me know if I'm getting a bit distracted. So with that, let me introduce our main topic of the day, where we're going to be speaking with Henry.

Charlie Cowan: I'll ask Henry to introduce himself in a moment, but we're going to tell the story about how Henry has gone as a CEO of an organization, a, as you might say, fairly traditional manufacturing organization, and has jumped feet first into the world of AI and is transforming the business into an AI platform. So with that, Henry, it'd be great for you to introduce yourself and maybe if you feel like it, we can go straight into just talk a little bit about CFS and the company and a little bit of background there.

Henry Chart: Yeah, sure. Thanks, Charlie. What an introduction. So yeah, I'm Henry, so I run CFS. We've been around 25 years. My father originally started the company as he worked for kind of a German competitor which operates in the UK. But what we do is we design and manufacture construction products for the UK and Europe. We work across major residential buildings, commercial buildings, and major infrastructure projects. That's the core business, the core business, as I mentioned, has been around 25 years, and the new venture, which is probably what we're going to be spending a bit of time talking about today, is Clarity, and Clarity was born out of, I guess, frustrations on the way that the industry designs products today. Construction is very, very complex. There is a ton of people involved, which makes the whole process quite manual and hard. But we've kind of been focused on how we can try and drive and improve that process. And we're going to spend a bit of time talking about that with Clarity.

Charlie Cowan: Yeah, fantastic. And so a bit of your background. So prior to sort of taking on the lead, the leadership of CFS, were you busy as an AI engineer? Were you coding? What was your background?

Henry Chart: A slightly strange route into the family business. So prior to coming to CFS, I've been at CFS now for five and a half years, but before that, I worked as a data science consultant, so we were working for, like, M&S, Holland & Barrett, major retail companies. And I was, at the time, writing code, which now, when I look back, if only we had some of the things available to us now, that job would have been much more efficient, or I might have not even had a job. But effectively what we were doing there was trying to improve their strategy through using large data and using analytics and certain processes and tools that the company had as IP on how to improve things, which would stretch from things like pricing of goods, offers based on what people are buying, store arrangement, that type of thing. Super different to construction, but what then kind of obviously led me over to probably the lens that CFS as a group has now, from its technology basis, was this thinking around data, and this thinking around technology, and how we can try and improve that in the decision making that happens within our business, but also happens on how people are making decisions about how to use our products on buildings.

Charlie Cowan: Yeah, so I think this is good grounding as we go into what we're going to cover, that you had some technical sort of awareness, having worked in the data science space, but by no means a developer or a coder, but just that sort of awareness of, and the belief that.

Henry Chart: Yeah, and the reason I actually ended up getting out of that and going to the family business is that I knew firsthand I was never going to make it as a great data scientist. I was not the best at doing that. But what I thought I was quite good at was trying to understand the kind of the business problem. And yes, I could try and hack some things around in the background to create the solution, but I was never the person who was going to go through all this statistical analysis and come up with this model. But I had an appreciation for how those things worked. And I think that has certainly carried into what we're going to talk about probably a bit today.

Charlie Cowan: Yeah. Well, let's go one level deeper. I know that some of those that'll be watching, this will be from a construction background, but many might not be. And it's useful, because we can talk about what the business, the domain problem is, that we then go on to solve with Clarity. I feel that I've been on this journey with you for a couple of years, and I feel that I could nearly become a structural engineer, knowing my way around a casting channel. But maybe you could just describe what it is that we've got on the screen and how it helps a building stay up.

Henry Chart: Yeah, so this is an example of one of our products and I would say 99.9% of people probably have no idea what any of our products are because ultimately at the end of the day they're hidden within the building and no one ever sees them. But what this product is, is a fixing product. So as I mentioned at the start, we work on major residential and commercial buildings. So if you look around you, if you're in a big city like London, the way those buildings are built is often the frame goes up, whether that's concrete or steel. This product gets used when it's a concrete frame, and then the facade gets bolted on afterwards. And the most common way to do that is through one of these products, which gets cast into a slab, and it enables you to connect whatever the facade is, whether it's bricks, glass, and it enables you to bolt that back onto the slab.

Charlie Cowan: Okay, amazing. And then I guess the challenge, or the design process here, is determining the width of the steel, the number of these anchors that tie back into the concrete, and then the sizes of the brackets, and how many brackets are spaced across a meter, that basically determine, will this building stay up? And there are some fairly complex European regulations that govern these calculations. So maybe you can just explain a little bit about what happens in a pre-Clarity world, where there's spreadsheets, and I'm almost imagining sort of slide rules and calculators to try and work out the right dimensions.

Henry Chart: Yeah, I mean, if we take a step back and you just think about a building in general, so if you think about a high-rise or medium-sized building, 10, 15 storeys that is built, I think the first thing to say is that it is very complex, and there is a lot of parties involved. When each and every junction of the building has many, many different interfaces coming into it, and whether that's different people working on different parts who are responsible for different parts of the building. For example, an architect might be responsible for the general coordination of what the facade is and what the building looks like, but then you have a structural engineer who's getting involved to try and understand what does that work or not. And then you've got all of the different manufacturers that then come in because they will have products that fit within certain areas.

Henry Chart: And so then you have to overlay all of that information. And then on top of that, you then have to understand, well, what is the interface between this manufacturer's product and this manufacturer's product? And how does that work? And so the complexity grows and grows and grows and grows. And that's why it's hard to get right. And that's why the construction industry has lagged in many ways, because it is a very complex process and also the parties within it don't tend to be tech first, and so that interchange of information tends to be quite manual, like you mentioned. So, a manufacturer might have static PDFs, or they might have a standalone piece of software, but it has no integration with any other pieces of software and certainly no integration with other manufacturers' products.

Henry Chart: So then you bring all that together and actually it is genuinely quite hard to get it right. And what then tends to happen is creating a simple solution becomes very, very hard.

Charlie Cowan: Mmhm.

Henry Chart: It is quite hard, it's quite easy to overcomplicate things and that tends to be what happens, and unfortunately these kind of problems get repeated because the information isn't easily available to everyone.

Charlie Cowan: Yeah, yeah. And what we're seeing on this image on the right, this hasn't actually got a casting channel, I think this is, what do you call it, post-fix? Is that just bolted in?

Henry Chart: Look at you.

Charlie Cowan: Look at me, look at me, I'm in the wrong game. But the cast-in channel would be in the back here. But you can see just the one specific bracket. There's the size of this bolt. That would be the cast-in channel. There's calculating the width between the slab and the masonry. And then you're calculating all of these. And this is just like one bracket on one floor. And around a high-rise building of, I don't know, 30, 40 stories, this is just multiplied by, I'm guessing like thousands, tens of thousands of calculations.

Henry Chart: Yeah, and obviously this is just one product, right? What's not shown in this section is fire barriers, which this has to interface with, which obviously, rightly so, has grown in importance since Grenfell, cavity trades, wall ties, there's a ton of product detail that isn't being overlaid on the top of here as well, and all of that has to interface.

Charlie Cowan: Yep. I think we will take a look at the platform in a little bit, but there's so many similarities between structural engineers and coding engineers. In that there's a lot of this sort of manual work that these engineers are doing, these calculations, which is probably not why they got into being a structural engineer. They want to spend more time on the taste, the judgment, the creative side of things. And, yeah, just very conscious that there's a lot of manual, spreadsheet, repetitive work going around a building that probably isn't the most motivating work for people to do.

Henry Chart: Well, I'm done. You know, they tend to be risk-averse engineers or architects because ultimately, if a solution has worked in the past, they are inclined to use that solution again, because especially since Grenfell, the liability and where things sit has changed, and so people can't be held personally liable, and the risk of getting it wrong is so high that you're not willing to accept or experiment with better ways of doing things, which obviously creates a problem with trying to improve efficiencies. And what tends to happen, and it's probably true in other industries too, is that what gets done and designed is what the person next to you knows and has done in the past. And so you just learn from kind of speaking to someone, but not necessarily innovating in better ways to do it.

Charlie Cowan: Yeah, yeah. So let's go back, I don't know, a year or so ago. What was the moment when you started thinking there's got to be a better way for this, and what were you playing around with? Were you in ChatGPT? What was the genesis of thinking about this?

Henry Chart: So the idea was created a few years ago. And I think we basically started developing this through the traditional route, and this was before the release of all these AI tools, and particularly their focus on coding. And so initially, we started trying to create a piece of software through a development house. So we started that. That was great. And then I started talking to a friend and he actually mentioned at the time a company called Replit. And he was saying, oh, you can build applications just by talking to this thing. I was thinking, can you? And in the back of my head, thinking, oh shit, because, you know, we're spending tens of tens of thousands of pounds on developing software, and what's this app that can build things without really just talking to it, initially thinking, well, what does that mean for Clarity in the future, if you can just do that? And so anyway, started playing around with that, and that's kind of where this whole thing has grown from, yeah.

Charlie Cowan: Yeah, yeah, it's just that toe in the water and then you think, and then you're hooked. So Replit was your first tool, and I know when we first started talking about this, we were using Claude Code to start. I mean, the key challenge at the start, I mentioned earlier that there are a lot of these European regulations that define each of these products and some of the thresholds, and the calculations that they need to pass in order to be verified and true. And actually just deciphering these complex PDF documents and trying to understand what the actual calculations are was…

Henry Chart: Yep.

Charlie Cowan: …like the biggest first step, before you try and solve something, like, how do you calculate it? And, so maybe there's a little bit of that process of trying to understand what the problem is before you start building a solution.

Henry Chart: Yeah, and I think that's probably true of anything that you're trying to solve, really, and that's something that we kind of, unfortunately, found out the hard way initially. One of the temptations with using these tools is to jump in and build straight away, especially because you can get almost instant gratification on building something that looks quite nice without any awareness of whether it actually works or not. And so, working backwards from that, we developed a process in partnership with you of what we call the Single Source of Truth, and that is a document that we create, which actually lives inside the code base. Not only to help us build the software itself, but also to help feed interactions from users in the future, because they can interact with the platform itself by chatting to it, and one of the reference documents they can use is this single source of truth. And so what that document actually is, is basically explaining to anyone, if they hadn't ever learned about masonry support, casting channels, anything, about the rules around how this product works. And that's something that's been quite useful for the business in general, because what you tend to find is actually there is so much knowledge that is stuck in people's heads.

Charlie Cowan: Mmm.

Henry Chart: And you don't actually appreciate that until you go through and you say, well, what about what happens in this situation? And they're like, oh, yes, well, I forgot to tell you about that. And, fundamentally, whether you're building software or doing something else afterwards, developing that process has been very good for us, because it forces you to think about all of the edge cases.

Charlie Cowan: Mmhm. Well, I know those watching will love to have a look at the tool. Let me, while you sort of prepare for that, I'll just go to the website here, which anyone that's watching on the video afterwards, it's clarity.site, and, in a moment, Henry will talk about how he built the website all with Claude Code as well. But, yeah, this is really taking the very sort of traditional and manual process into the future, with AI chat, with these 3D designs all in the platform, which, compared to what's traditionally been on offer from these manufacturers, this is just completely game-changing. So I don't know if you want to share your screen, Henry, and sort of guide us through a couple of scenarios and thinking about how AI is helping here. I'll just stop mine.

Henry Chart: You got it.

Charlie Cowan: Okay, fantastic.

Henry Chart: You see my screen, yeah.

Charlie Cowan: Yep.

Henry Chart: Yeah, so this is Clarity. We built it as a web app and kind of structured similar to how a lot of the UIs on these platforms are right now, where on the home dashboard, you have basically a chat entry point where you can start designing a product. And so inside here, I can say, great, masonry support line. In the background here, we've built in Anthropic's LLM, and it's pretty easy to integrate into the software. I can't remember what model we're using, but it's certainly not Opus 4.8, the performance is still very good. On top of that, we built certain tools. So this would be one of them, to start trying to ask the user questions based on information that it needs to do it. And so in here, it's saying, what are you fixing to? We spoke about the casting channel earlier. It's asking for a load.

Henry Chart: It's asking for a cavity, which would be defined usually by the architect. Support level and slab thickness. And then what it's able to do is, so what the AI is doing here is actually very good, useful to collect information from the user and then pass it to calculations. And so we spoke about the calculations briefly, but the calculations in the platform are hard-written in code. They're not up for debate. And that has been one of the things that we have to be very conscious of with the users of this platform. Making sure that, if I ask this same question to the Anthropic LLM, is it going to come back with the same answer each time? And it needs to in this situation.

Henry Chart: Whether the text and how it responds to me is slightly different, but that doesn't matter, but based on the inputs that I just gave there, every single time it needs to come back with the same response. And so we've built that, and it does. So inside here, what's happened in the background, we've kind of codified our logic around these products. It's developed, it's pulled and created this 3D model here. That process would probably have taken us hours, if not days. And also, in the background here, we're running, I think, over 45,000 different calculations to work out what's the optimized design. And so that wouldn't be humanly possible to run that many checks. And what that provides us as a business is certainty that we're giving the best design to the end user. Yeah.

Henry Chart: From here, this kind of works similar to if anyone uses Claude artifacts, that type of thing, albeit we're calling these designs. So it has generated a file here, and then you can then save them into a project. So I can save that and I can create a new project. Oh. And then you'll see here that that design file is within here. You can then open up the full design page. As an engineer or someone technical using it, you can still interact with this product by choosing all of these buttons or kind of the more traditional form-filling way that you might go about it. But obviously, that entry point by using the AI, it just kind of speeds up that process to get to this right result here. And obviously, we're looking at one.

Henry Chart: Product here, which is great, but the vision behind Clarity and trying to solve that interface problem and design problem across multiple different parties is integrating other manufacturers' products into here. So we're looking at integrating fire barriers and how that interaction works with this masonry support, and also starting to create and expand standard details, so a standard detail that an architect could use in particular situations that they can quickly load into their model. Hmm.

Charlie Cowan: Can you just talk a little bit about the 3D model? 'Cause it's super cool. I mean, the first time I saw that, that was right up there with the birth of my four children. The first time I saw that, I was like, that's super cool. So what is that and how accurate is that?

Henry Chart: Yeah, so there's a few tricks we're using in the background to be able to do this. This is modeled in Grasshopper. And so the actual geometry of these products, we can actually download and manufacture from these. We're building outside of Clarity, we're building agentic workflows that, for example, can interact with Clarity, but then also interact with other tools. So these products can be manufactured from and then we can actually send these to other parametric models of our products that can then go and flat-pattern these, so that you could then send that straight into the warehouse to be manufactured. And so, obviously, one of the problems around construction is that I think it's 50% of projects that end in dispute because of design-related issues.

Henry Chart: And often what we see is that detailed design and products get left very, very late. It just introduces risk at every stage, and so if we're able to actually get the design right at a much earlier stage, down to a T, this can be manufactured, we can price, and we can put this and deliver it on site within, kind of, 10 days from doing the design in here, it makes a massive difference. Some of the things that we're thinking about, or certainly that I'm thinking about, is, in the future, people may not be, whether they want to meet and design with Clarity.

Henry Chart: Obviously, at the moment, we have a web app that you can interact with here, which is great, and I have to try and tell myself every day not to care too much about the UI, although I find it very hard not to, but I think in the future, we're thinking about, can people talk to Clarity from inside Revit or Tekla, and just get the answer back to what they need directly from the tool that they're working within.

Charlie Cowan: Mmm. Yeah, and thinking about that, just like the interaction that you have built within the platform about chat versus filling in fields, as you were talking about there. Because the engineers have probably got numbers, and rather than having a long conversation, they just want to get the numbers in. What have you seen from the users about how they're liking to interact with the tool?

Henry Chart: Yeah, so for us, onboarding is really critical. And I know that's not a brand new thing with software, right? Like onboarding, you have to get it right to be able to get people to almost hit that inflection point. I think the nature of our user base at the moment is engineers, and engineers, I think, like quite traditional ways of working. And I think there is a trust element they need to get over within using the actual chat functionality. So obviously in here, you can create a design and fill in the form manually, but you then have to work to find the best result as opposed to getting AI to work for you. And so there's a couple of things which we're thinking about to actually try and fix that.

Henry Chart: The first thing is, on the home page, building pre-built form-filled kind of templates, so that the user can still fill in a form like they might be used to, where they feel like they've got control over the inputs, but actually, when they fill those in, they're sending them to the AI, and the AI can work out what the best result is, as opposed to the user kind of going back and forth and saying, oh, well, that didn't quite work, let me try something else and different. An example of that might be if I find, in here, this is just a channel design. But in here, at the moment, this channel here, you can see, it's because it's red, it's heavily overutilized. And so if I'm doing it the traditional way, which is, oh, okay, well, here's what I filled in, why doesn't that work?

Henry Chart: Okay, well, should I try reducing this load down to 20, not 35, rerun that, that may work, but okay, well, it still doesn't work. So that's a way that traditionally might have been done, but actually, if you interact with this through the AI, you can give it the challenge and say, well, this doesn't work, go and find me what I could change to make it work. And here's the parameters you can play with and here's what you can't. And it can go away and start doing that work for you. And so that's obviously through good onboarding, that I think people are starting to lean into.

Charlie Cowan: Yes.

Henry Chart: Oh, here we go. Okay, fine, I get it.

Charlie Cowan: And, yeah, I mean, I think about how humans would have been doing it. It's a bit like the kids' game of pin the tail on the donkey. You're just sort of trying to get something that works. But it might not be the most efficient one. And so one of the key optimizations of this is that the manufacturer, or the construction company, wants to have the most cost-effective solution. So you need something that works, but not 1% more than that of efficiency, otherwise you're paying for more steel than is required, and so you might have a human that comes up with a solution that works, but it's over-optimized, and therefore you're just spending money on steel, too many brackets.

Henry Chart: Yeah, yeah, exactly. And you often don't have the time to check all these different situations as much as, construction, everything is needed yesterday. And so people are always under pressure, so you don't have time to fully optimize things or run different scenarios. But obviously, using the LLM inside the platform, it enables you to run those very quickly. And also stress test different scenarios. Something that's quite common in construction is things are out of tolerance. It's all good and well designing something like this in a piece of software. This represents a slab on a 30-story building, and there are tolerances involved. And so those tolerances can impact whether the product works or not, but you're able to kind of stress test those ideas probably early on and get the results quite quickly as opposed to doing that manually.

Charlie Cowan: Mmm. Cool. One final question on this before we talk a bit more about your process, and maybe if you go back to that chat that you kicked off, is this balance between when to use the AI, and when to use raw calculations, and I think in your chat, it was saying, you know, calculations have been passed, and so, you were explaining there that the AI goes off to go and come up with multiple variations, but having then worked out which ones are the most likely ones, it's hard deterministic calculations that are in the application that verify whether that design passes.

Henry Chart: Yeah, that's correct. So I think, probably this is best explained by going, if I go over to that project and open that design file, there we go. So, inside here, if I go into the results, and I download this PDF, in the background here, the calculations that say whether this product works or not.

Charlie Cowan: Mmm.

Henry Chart: Are hard-coded. They are not up for debate. So you'll see in here, I'll just show you an example of them. As an example, we won't go into what this is, but here's a calculation. This calculation is fixed. What the AI is doing is it's taking input data and feeding it into these calculations and saying, does this work or not? And so in this example, obviously, we're looking for something that's under 100% utilization. So in this case, that's okay. And so what it enables it to do is it's very good at extracting information from a user and passing it to these calculations, which are hard-coded, and trying to say, yes, no, does that work? And it can run 100 different scenarios all at once.

Charlie Cowan: Yeah, I think this is a really good point for not just anyone in construction. This could be finance, it could be legal, it could be any other domain, of how do you combine the AI with some deterministic calculations, because often people will say, how do you trust it? How do you trust it? Well, you can trust what's coming out of Clarity because it's deterministic calculations that are verifying a design.

Henry Chart: Yeah, that's great. I mean, that's really our pitch, to be honest, in the future as well, which is why you would use something like this over just using a generic LLM to try and figure out whether these products work, is, we guarantee that the result that comes out of it is correct, where you can't necessarily do that if you're just talking to Claude, the desktop app.

Charlie Cowan: Yes. So a couple of questions on your process and maybe the website, before we talk about internal agents. Yeah. So you come to this as a sort of non-technical CEO, gone down the rabbit hole of Replit, Lovable, and then you loaded up Claude Code for the first time.

Henry Chart: Yes.

Charlie Cowan: So talk to me about your evenings and weekends and your ability to iterate on the design and the functionality as a non-technical

Henry Chart: Yep.

Charlie Cowan: leader.

Henry Chart: Evenings and weekends are fun and stressful at the same time, I think. Very early on, and I'm sure you probably had similar feedback, which is, you would try something, and then Claude would produce a kind of a front end on something called localhost, which obviously we weren't really aware of what the hell that was before we started using it. And it has gone all completely wrong, and you're having a major panic attack that you've broken the whole thing. I think that is largely gone as these models have got so good. And so I think some of the things you're doing back in the day in terms of creating, making sure that you're keeping a track of the task that the product is doing, all these things have been built into the actual software or the agent core code that is growing out now. But in terms of the way that I work, we have the code stored on GitHub, like everyone else. I will then pull that code, and I will use Claude Code like I have down here. I actually, oh, this is a design, here's the code. We just.

Charlie Cowan: It's in your web browser, I don't know if you…

Henry Chart: Okay, let me just change to a different screen. Desktop. How's that.

Charlie Cowan: Perfect. Oh, yeah. There she is. Hello, old friend.

Henry Chart: Okay. So GitHub, fine. That's not new to anyone who's probably writing or using any sort of code. I pull that and now work on it. Now, up until very recently, I had been using Ghostty. Ghostty was, it's just a different terminal, and I was kicking off Claude Code sessions in here. As of, I would say, 2 weeks ago, I think I've largely switched to using the code inside the Cowork app. And the reason for that is I think they've just built more things that make it more user-friendly for a non-developer, personally, for me, using. So things like here, where it's doing reviews versus kind of the main GitHub account, and it can show you the process. And also some of the things that Claude Code are bringing out in terms of things like goal and other ways to interact with it. I find that inside the Cowork app they actually enable you to kind of upskill the way that you use Claude Code without you necessarily being a domain expert. I don't know if you find the same thing. Yeah.

Charlie Cowan: Made it, made it very easy. Ahhh. Yeah, and I know we've talked before about voice as well. So what's your experience of literally talking to Claude?

Henry Chart: In my experience talking to Claude, it's my wife telling me, "Who the hell are you talking to all the time?" But I use, what's the plug?

Charlie Cowan: Right, yeah, yeah.

Henry Chart: Yeah, yeah, yeah. So Wispr Flow is obviously on my laptop, as you can see here. And you basically just end up talking to it, and then once you set off your, the code tasks, you then have to figure out the balance of do you go and start another piece of work whilst it's working, or how do you try and jump between different sessions and keep the context in your head of what's going on. But, I guess, to summarize, GitHub is where the code is stored. I use Claude Code to actually play around and change things. We have a developer, full-time developer, that actually works on the platform, and we have a part-time product manager. And so, I think my role, I absolutely love playing around with these things, but I don't have the full time to play around with these things all the time.

Henry Chart: But what is very useful for me is I can go to client meetings, I can understand certain things, and I can actually make those changes myself. Figure them out, I could even do them live within a meeting themselves and actually push that change and say, is that what we're thinking about? And so, I think it is important to say as well that CFS Group, we are probably 13, 14 million turnover, but we're not a big organization in terms of headcount. I think where that means that we are is I've been quite hands-on. I think if you're in a super large organization, I think it is important to understand the art of the possible. But I could see how these tasks would probably be delegated to someone who would run that division.

Charlie Cowan: Yeah. And I think for anyone that's watching that is a non-developer, it's important to see what's on Henry's screen is not code. It is you talking to Claude as if Claude is your developer, and the instructions you're providing are, I want to change the way this page looks, I want to move that button to the left, I want the user to be able to do X or Y, and so it's a lot more like you're delegating a task to someone than actually being a developer. Claude will work that out. So there is that onboarding, but once you're there, suddenly it's like you've got superpowers.

Henry Chart: Yeah, exactly. And I think, as you can see here, I haven't actually sent this task off, but this is the kind of common sort of interaction that you might find, is that I've taken screenshots of something from a UI perspective that I quite like. And I would just talk to it and say, I'm thinking about this front-end work, can you explore how you might integrate this? And it will just go away and do it. I think the coding part of it isn't necessarily the challenge anymore. There is a…

Charlie Cowan: There's.

Henry Chart: …hurdle to get over in terms of using these tools. But once you kind of do an initial upskilling on that, you can do it. I think the hurdle actually, in many cases, is on taste and not building stuff that…

Charlie Cowan: Yep.

Henry Chart: …he wants.

Charlie Cowan: With that, and we'll wrap up on the Clarity bit, do you want to share your screen, your web browser, and the Clarity website? Yeah. Because there's some amazing 3D animations, which I know you've brought to life, and, for those of you that are watching that, I'm not trying to build a construction app, so how would this be useful for me? Well, no doubt your company's got a website, and no doubt up until recently, well, maybe you still do, if you're watching, that you've got an agency that you have to go to and they quote you 50K to do a small change on the website. So maybe you can just talk about how you've been able to bring the ideas that you've got in your mind to life with Claude Code.

Henry Chart: Yeah, we were certainly in that camp. This website, I think, two years ago, for something that was more simple than this, we were quoted, I think, 15 to 20 grand to build it, and probably a couple of months. And I built this in maybe a couple of days, and if I don't like something in here, my process would be that the code is stored on GitHub, I can work on the code, I can chat inside the desktop app, and I say, can you change this, or make it look like this? And within five minutes, it's changed it and it's done. This visual here, this is a 3D visual, which looked pretty cool. Again, I probably would have spent a couple of thousand pounds to get someone to try and generate that visual and link it in, all the back and forth.

Henry Chart: That's built with something called remotion.dev. Again, I didn't necessarily know how to use this, but one of the beauties of using the LLMs is you can just point it at it and tell it how you get this to work. And so that's the kind of process that I would follow, which is I would get this and ask Claude, how do I get this to work? And then I just carry on having a chat and you end up getting it to work, but that's what these visuals are built on here, and certainly some of the other ones down here, which are not just great for a website, but obviously, if you're posting on socials or anything like that, it's a great way that you can build this interactivity into the actual site.

Charlie Cowan: Yeah, unreal. Awesome. Well, let's switch gears a little bit. I know that separate to building out the product, you've been diving into building agents internally in the business. So you were talking a little bit about the size of the team there. Where have you got to on building agents and how are they complementing your human employees?

Henry Chart: Yes, so I think similar to you, the strategy around it is we're trying to, in the future, particularly process and admin roles, we're looking at where we can build agents to come and automate a lot of that. So, at the moment, inside the business, I think I might have this link live here. These are the agents that we have live, so we've got three, basically. Ada, which works as a purchasing assistant. Archer, which works as an external SDR role. And Eva, which works as a sales administration, so processing orders, effectively. Now each of these agents can also talk to each other, and so the general setup is that we have these agents which are very focused on a particular task.

Henry Chart: But then they all write to memory in a brain, and so they can all, over time, the intelligence of those agents gets better and better as they start to pick up and learn certain nuances, whether that's a customer responds in a certain way, or a supplier always does this, and that information is stored. We built these using Claude's managed agents, so they're hosted there, and then we have a kind of a wrapper around it using Vercel, which actually triggers certain runs, and when they actually get used. And you can also interact with them, so there's nothing on here, but you can also interact with these guys in Teams. So, for example, this is an example of Archer. So Archer is an external SDR. So I can ask him about certain things and he'll go away and go and find the contacts of the project.

Henry Chart: We've connected him to using Apollo, so he can go and grab contacts. He also is connected into HubSpot, and so actually puts the contact into HubSpot and can enroll them in an email sequence to kick off comms. It's grabbing project information, making the actual sequence hyper-personalized. So yeah, crazy possibilities. I think the only thing I would say on this is that I would definitely make sure that the business case is there, because something that I think is now waking up in terms of the news is that if you don't get the design right, or don't think about it, they can be quite expensive and can rack up a bill quite quickly.

Charlie Cowan: Mmm.

Henry Chart: And for reference, Ada costs us about 300 USD a month. And that's probably true of the others too. So you're probably looking at around close to $1,000 a month for running these three agents. Which is obviously still more efficient than human. But again, cost can spiral.

Charlie Cowan: So that's running on an Anthropic API key, and do you have a sort of a reload? We have ours reloading at $45, because I'd rather have lots of reloads and I can keep an eye on it than get a reload at $19,000.

Henry Chart: I get slightly anxious every time I get an email from Anthropic thinking, oh no, it's happened again. But yeah, we have the exact same thing, basically. And I just hope that I don't get an email from them too frequently. Yeah.

Charlie Cowan: A special folder for that email.

Henry Chart: Yeah, yeah. But, I mean, that's the thing, the reality is, is that we were planning to hire humans to do this. So there is obviously an easy business case around this, whether how it's improving, from Ada in particular, we have kind of real KPIs where we're tracking performance, like our on-time delivery has gone from kind of the 80% to the 90% since we started running Ada. So there's real genuine metrics and improvements that these things are actually delivering on.

Charlie Cowan: Yeah, brilliant. And I think just, to reiterate, Henry's coming to this not as a CTO, not as an engineer, it's just been getting in, dabbling, keeping on top of what's going on with Anthropic, and it's often one of the best ways is just to get in and start to learn, and I think the first one that you built out was the procurement agent, and if I remember rightly, your head of ops was saying, I want to go and hire a procurement agent, and you said, I think I can build one.

Henry Chart: Yeah, that's right. For example, you can see up here on version 91 of Ada. So there is a lot of tuning that has to happen.

Charlie Cowan: And…

Henry Chart: …but this system prompt here, you probably look at this and think, what the hell is that? This takes a lot of time. We've got the skills set up, the servers that it can talk to, and all that type of stuff, and the schedule of the different jobs that are run automatically. So, it does take a bit of time, but once you know the basics, anyone is capable of doing this. The one thing I would say, just as a slight sponsorship of Kowalah, these things are moving so fast.

Charlie Cowan: Then.

Henry Chart: It is also very hard to keep on top of and make sure that the maintenance around the agents and how the tech stack is changing, and obviously Claude or Anthropic have dropped Claude Tag, and okay, well, how does that change what we've built with these specialized agents and all that stuff? So, I think it's very useful to play around with it, and you can build and productionize these things, like I've shown you can. Having said that, to keep and maintain and keep it kind of tip-top, I do think that's obviously where partnering with someone to deliver that is useful.

Charlie Cowan: They need, I mean, I think about ours, we've got about 20 of them running right now, and I think we need to treat them like human employees. They need coaching, they need feedback, they need ongoing training. If you build one and just let it go, then just like any employee, you're gonna get, you know, bored, go down the wrong route, so…

Henry Chart: That's right. And I think the other thing that we're battling with, not battling with, but I think it's something to bear in mind, is when does something become an agent versus when does something live in the code, the Claude Code desktop app? Because obviously using the desktop app is a much more efficient way of getting some work done.

Charlie Cowan: And.

Henry Chart: You have the monthly subscription, you have an amount that you can use, but you're kind of fixed to that. But I find that you can get a lot more usage out of the app without being charged as much.

Charlie Cowan: Yeah, yeah, yeah.

Henry Chart: So that's obviously a balance that we're trying to fit internally.

Charlie Cowan: I'm going to share my screen. There was a really good post, I'm just going to put it into the chat here and I'll put it on the YouTube video as well, from the CTO of Uber who posted today all about how they are driving AI adoption across all functions, not just in engineering. And so he talks through this six-step process, but they're taking an engineer, so that could be either you've got developers in-house or your friendly Kowalah engineer, but they're going and sitting down and spending two or three days just doing a day in the life with those teams. What's your process? What are you doing? How do you do payroll? How are you hiring someone? How are you processing customer support tickets? And then those engineers are then building out whether it's an agent, whether it's a skill, whether it's a workflow. And in his post, he documents the percentage of savings or new revenue that they're getting through this. So I like this agentic pods idea of connecting up the domain expertise with someone that can build.

Henry Chart: Yeah, I mean, that's exactly what we're trying to do, albeit without the scale. I had an example of that literally yesterday where, say, someone who sits in our finance team, one of the tasks I dread is going through the Barclaycard receipts each month.

Charlie Cowan: Mmm.

Henry Chart: And then she has to reconcile all of those through all of these PDFs. And, so I just said to her, write that process down and we'll build a skill, which will automate that whole thing.

Charlie Cowan: Yes. I do find this walking around clients' offices. You're just looking at stuff. Skill, skill, skill.

Henry Chart: Okay.

Charlie Cowan: I was in a client's office the other day, and there was three people who just had all their receipts out on a table, obviously from, like, client events or something, and they're just sat there taking pictures. I'm like, okay, A, this is just a pointless process, but the opportunity cost of the amount of these people just taking photos of receipts is like, it's ridiculous.

Henry Chart: Yeah, well, we haven't got time to talk about it, but you make an interesting point, which is obviously trying to think about how much of communication or things that can be actually trying to get them in channels so that AI can interact with them.

Charlie Cowan: Just, yeah.

Henry Chart: Yeah. Okay.

Charlie Cowan: Yeah, pick it up automatically. Yeah. Well, we're coming up towards the end. What I'd love to just wrap up thinking about, and I'm going to skip AI in the news because this is way more interesting and useful, is just thinking about someone that might be watching this that is like, well, I didn't work in data science, I haven't got maybe that belief that I could go on the same journey. What would you share to someone that's in that situation?

Henry Chart: I think there's a few things that you could do. I think having a plug, again, to Kowalah, having a conversation with someone like yourself would be hugely valuable, because if someone can teach you the basics, as long as you can speak English, anyone can master these things. Because as soon as you do not know anything, you just have to ask the AI, help me understand what's going on here. And so I think.

Henry Chart: My push would be just get started and start playing and just moving outside of your comfort zone of how to use certain things, and if you don't know, then you've got the world's best teacher in the app to tell you what you should be doing, and so I would just encourage anyone to get into it and play around with these things, because until you do, you really can't work out the art of the possible, and that's what I found anyway, is, things are changing so fast, and how can I decide what we should be doing, or how this technology is going to impact us if I don't understand it myself?

Henry Chart: And so my encouragement would be to get stuck in, but also, I am conscious that people sometimes don't have the time, or maybe don't have the curiosity to do that, and therefore, if that is the case, I would encourage them to have a conversation with someone like you, and then you can help guide them through that process.

Charlie Cowan: Yeah, yeah. I saw this post from a guy called Ethan Mollick that I follow, and for those of you that are on Claude, you might have heard of this new model, Fable 5, that came out, went away again, came back again, and his post was that I think founders, that leaders of companies, are lowballing the challenges or the tasks that they're giving to Fable. And that actually, you should think of what is the craziest, most outlandish outcome that I could possibly want for my company and throw that at Fable and go, right, how do we do this? So, that might be, we are a traditional manufacturing company, I want to reinvent this business as an AI-enabled technology business. Help me figure out how to do that, and it's these…

Henry Chart: Yep.

Charlie Cowan: …really weighty problems. I mean, I was doing this with Fable. I flew to the US on Sunday to go and do some work with a client, and I was chatting to Fable the whole way over. I mean, I went deep, it was bad, but I was like, imagine all scenarios are on the table for us to just dramatically scale what Kowalah's doing, and the conversation I was having, it was just unreal, what it was able to do.

Henry Chart: Doing yourself out of the job in the future.

Charlie Cowan: Yeah, yeah, yeah. Oh, well, brilliant. Henry, I knew it'd be a great conversation. This has been absolutely fantastic. Thank you so much for sharing your story. Really excited to see Clarity grow. I've shared clarity.site in the chat. I'll add it on the YouTube channel as well. And I know Henry and his team are very able to answer any questions anyone has on that. I'm going to skip through AI in the news quickly, because we've only got a couple of minutes left. And just want to talk about what we're going to discuss this time next week, which is really going back to that start of the journey. So a lot of the organizations we start working with, there is some sort of empowered executive, maybe it's the CEO, a CIO, maybe it's a COO, someone has made the decision to go Claude, to go Anthropic.

Charlie Cowan: And then the question comes, right, how do we do this in a structured way? The things that we build are the most impactful to our business. We've moved from the years of experimentation and giving people tools and just seeing what happens. Now companies need ROI, especially as these token costs continue to expand. So we're going to walk you through our discovery process that we run in the Kowalah platform. We're going to run you through how we do all of the AI use case scoring, and then how you can start building the things that are most impactful. So we will see you same time next week, 3pm UK, 10am Eastern Time. And, once again, thanks, Henry.

Henry Chart: Thanks, Charlie.