Rob Collings speaks to Guy Hutchinson, discussing AI initiatives for finance leaders and accounting firms

Speaker A: Welcome to CFO Insights, the leading podcast for finance professionals in disruptive tech, brought to you by the startup CFO community. I'm Guy Hutchinson. I'm the host of the podcast as well as being a tech cfo. In this episode, I'm joined by Rob Collings, chartered accountant, former head of finance and CFO at tech startups, and now founder of Flexel, an AI native accounting firm, uh, built specifically for early stage companies. Rob reveals his insights, uh, on where AI genuinely moves the needle in finance and where it doesn't. We dig into the biggest build versus buy decisions from finance teams and why compliance tasks are proving stubbornly difficult for AI to crack, and how the relationship between startups and their accounting firms needs to evolve in an AI first world. Rob, welcome to the podcast.

Speaker B: Thanks Guy. Very excited to be here.

Speaker A: No, that's all right. Look, you and I've had some great chats over the years. Uh, always really been impressed by how you think about how the finance role evolves and how future finance leaders really need to be thinking about, uh, how CFOs will be out there delivering essentially. Uh, and of course here we are in 2026, and I'd say maybe five years ago, 10 years ago, that discipline of being a finance leader just wasn't changing all that much. Right. It was a relatively, uh, static career in terms of the kind of things you did to be achieving a high standard. Uh, and in 2026 we're four or five months into some ridicul, exciting Claude releases and some fantastic innovations from people like light and Roulet. We're now seeing everybody seemingly talking about AI in finance. Uh, and it's a really exciting time where it absolutely is not a static state of the art and things are changing really fast. And whilst there's a lot of uncertainty for many people, it's also just incredibly exciting. So, really pleased just to have you on to, uh, pick up on that topic and uh, share some wisdom.

Speaker B: Yes, yeah, I think you're definitely right. I think there's a lot of change at the moment. Um, some of it may be a bit smoke and mirrors, but other stuff is actually really having an impact. So looking forward to jumping into some of those things.

Speaker A: Yeah, absolutely. So why don't we spend a couple of minutes just talking about your career background, just just to give a bit of color about, uh, the kind of things that you've done, how you built up your experience and how some of this has informed your thinking about the kind of things that you're working on now.

Speaker B: So I started my career in an accounting firm, probably similar to quite a few people. Um, I did AAT and aca, went through the normal sort of audit training route. And probably like most people, again I thought at some stage I don't know if I want to do audit anymore. And that was the point when I started to really focus on tech and high growth companies. And so they became essentially my whole world for quite a few years. Met lots and lots of founders, really got stuck into what it's like to have a tech company as a client for an accounting firm. And then I reached a bit of another sort of pivotal point where I then decided I wanted to jump into tech startups myself. And that started sort of the next stage of my career, which was head of finance, cfo kind of roles within tech startups themselves, also covering the normal kind of stuff that the finance person, um, works with, the sort of people side, the legal side, the compliance side, all of that sort of thing. And then very recently, um, AI came into the whole world. Lots of things happening there. And I thought this is probably a pretty good time to sort of go back to the world of accounting firms and see if I can start my own firm which is very focused on AI. And that was where Flexel came about.

Speaker A: Yeah, that's really interesting. And uh, what are the key things that you're focused on right now in terms of Flexel?

Speaker B: Yeah, so Flexel is what I would call like an AI native accounting firm. And what that means day to day is that we still have people involved. So our clients still talk to human beings, they still talk to me, still talk to other people in the firm. But we as people are uh, essentially supercharged by all these different AI powered workflows that we're building internally. So we've got a whole range of workflows, some of which are using automation, some are using AI, some we haven't quite automated yet. So they are still manual. But the idea is we can eliminate a lot of that manual processing which accounting firms traditionally do, freeing up people's time to actually focus on helping the clients, um, focusing on figuring out what the best thing is to what the best next steps are for clients, and also improving the quality and the speed and everything else around the work that is coming out of that accounting firm and that work.

Speaker A: Right, Those more manual transactional things, they were never fun anyway. Right. You know, people would, somebody junior might join an accounting firm and be doing AP and bank reconciliations and some work on AR and things like this for years. Right. That could Be like a really substantial part of their career. And yet I think for the majority of people it probably wasn't very challenging or interesting work. Uh, it just needed doing.

Speaker B: Yeah, definitely. I remember as well when I was, ah, cfo, I was sat there one day and I needed to approve a VAT return. And what I got sent was essentially, um, the face of the VAT return and then pages and pages and pages of transactions. And I was like, this is such a silly way of reviewing things, um, that you're basically just picking transactions at random to see if they're correct or not. And that was actually part of the reason, one of the many reasons why I went into Flexel and started that, because that whole process can be automated so much better with AI, um, and eliminate that kind of, uh, eliminate that sort of mental strain of having to review and approve a VAT return, which is just not fun, not fun at all.

Speaker A: Yeah, I do agree with that. It's definitely not something that CFOs put at the top of their list to value add. So we've picked the topic of AI and finance because one, I know that you're really interested in it. You're obviously solving a subset of that type of challenge in Flexor, and you've already alluded to the fact that there's a mixture of things going on there. Some of it is automation and you might need to be sort of telling it the rules to go through some steps so that person doesn't have to. And some of it's AI where there's some figuring out. And those, those models can be very effective at that, but they're certainly not 100%. Uh, and that's one of the things that, for finance people that are very process driven, um, can sometimes feel a little bit of a challenge as to the degree to which you want to rely on something that's not 100% of the humans. Not 100%. So is it really all that different? But we did say that, um, it would be interesting to start off this conversation because obviously the talk is so big, to start off the conversation with a bit of a sense as to what you think a finance leader should be doing right now in terms of AI deployment and maybe some things that they shouldn't be doing right, because it's not a complete license just to play and see what the tech can do. Um, have you thought about those things?

Speaker B: Yeah, good question. Um, I think I would actually start on the things that I would say we shouldn't do with AI, um, because I think it's helpful to run through those quickly. Um, the first one that I always say to people is don't use AI to replace your writing. Because whenever I get an email or a message or something that has so clearly been generated by ChatGPT, it just makes me not want to read it at all. And I think if as an accounting firm you tried to sell that as a product to someone else, it's just not going to go down very well at all. So definitely don't use it to sort of replace your writing. Um, and also don't rely on it for a lot of important things. So sometimes I might ask a tax question, but the reason I'm asking it that tax question is so it can go and do some background research, it can go and find the sources, it can give me more information which I as a person then go and take and look into in more detail to figure out the answer. Not necessarily to sort of just rely on the first answer that comes from ChatGPT, because it's definitely not the right thing to do. Um, there's a bunch of other things as well that we might go into later, um, which kind of COVID the covers the don't do side of AI. Um, but when it comes to the things that we should do, I tend uh, to think of them in kind of two different types. One is this concept of a, uh, rules based workflow. And a rules based workflow is anything where it's not random, okay? So if you give it certain data, the output that you want should be the same every time you give it that same data. So if you drop the same data into it five times, you should get exactly the same output five times. And that's a rules based workflow. And then the other workflow that uh, I think about is this kind of the, essentially the opposite of that. So like a principles based workflow or something. And that is the one where there is a bit more randomness to it and if you drop something into it, it doesn't necessarily matter if it comes out different each time because the output that you want is a different kind of output that you would get from a rules based work. So uh, an example of a rules based workflow, which I think is quite a good one to talk about, is when you are trying to move data from one place to another. And let's say that your client or your company has a sales system and there's lots of sales data in there and you need to get that uh, data into Xero. If you export that as a CSV and you want to convert it into a file that you can import into Xero. You want that data to be, you want that output to be exactly the same every time you drop that file into the top of the workflow. So if you were to put that data into ChatGPT or Claude or something, it might be different each time. So that's not a great way of doing it. So with a rules based workflow, you probably shouldn't just be chucking data into AI and getting it to spit out something at the back end of it. And then an example of a principles based workflow might be your finance inbox. So you have a lot of emails that come into this finance inbox and you need to decide what to do with them and you need to action them or figure out an answer and then reply to them. Now that's a perfect use case for AI because you can have AI sort of monitor those emails that are coming in, go and do some background research, draft up an email and then if you want it to send it straight out, it could, or it can just pause at that point. A human could then go read the draft email and then they send it manually. But the idea is that it's kind of a very big problem that it's solving and it's speeding it up so much faster than if a human was to do it.

Speaker A: Yeah, those are great examples. And going back to the rules based, uh, piece, how do you envisage sort of teaching, um, the application how to apply those rules? Like are you imagining a world where you are giving it the rules, or are you able to give it access to the historic entries and allowing it to learn the rules or propose what the rules would be for approval. Like, how does it get to learn what the rules are?

Speaker B: Yes, good question. And I think it depends quite a lot on the problem that's being solved. Um, so with the sales invoice example that I mentioned, that's one I would say you essentially want to use AI to uh, essentially vibe code this rules engine, which you probably just need to do once or maybe tweak it in the future, but the majority of it you just need to do once. So you can vibe code this essentially like a mini app which you can then use on a monthly basis and if you do need to change it, you can ask AI to sort of improve the app itself, but you're not relying on AI to run the whole process fresh every month and then have to go back and like figure out what's happened. And it never quite works out So I think that's an example of where you kind of, you just give it the rules up front and then it just runs with those rules every time.

Speaker A: And while she was just talking me through that, Rob, uh, I found myself thinking about use cases that I could reflect on for my career where I can imagine somebody, maybe me, maybe somebody in the team vibe, coding something. So, for example, in the old world you might have, uh, I don't know, a SaaS type business, but it's got some complicated kind of, um, you know, it's got some complicated issues around recognizing revenue. And so you've got maybe hundreds of transactions, maybe thousands every month. And you might have built a big spreadsheet 10 years ago, but actually maybe in the last five years you might have licensed something like Recurley or ScaleXP, and you've got a product that, uh, is set up to look for start dates, end dates, uh, and other criteria that would be important in terms of recognizing revenue. Uh, but I guess in the current era you might be attracted by the flexibility of. Well, actually, maybe we could just wipe code that, uh, and not have the challenges that it's a rigid thing I've licensed. It's the same for everybody. But I've got something that's bespoke for exactly how my business does things, how things appear on my invoices and how that will feed into, um, the journal entries or the things that I ultimately then post in my accounting system.

Speaker B: Yeah, I think that's very true. Um, I think I've always said that a lot of accounting Software is only 80% complete because it does the majority of what you need it to do. But it's always that last 20% which is very painful for the accountant because it's, I don't know, it's copying data from one system to another system or it's working out these tiny little edge cases. And a lot of the time Those sort of 20% problems end up taking so much longer than the entire process anyway. Um, and I think that's where AI is really strong because it allows the accountant themselves to close that gap without needing to have software engineers or without needing to put it on a software provider's roadmap and wait 12 months or something for the feature to come in. So I think that's a very, very strong way to use AI to close

Speaker A: that gap and do certain things like this. Do they? Is this part of a wider debate essentially around, say, build versus buy? Because I think quite a lot of CFOs that, uh, are Just naturally very process driven people. They've probably got quite a bit of experience. Change can be intimidating. It's easy to be thinking, well, you know, I've been using a certain ERP for uh, 10, 15 years. I kind of accept the world of change. I'll probably be on a new product in a year or two's time. But then maybe they don't want to be an early adopter. They don't want to be the one who's using a new platform when it isn't completely fully tried and tested for at least many years. Um, and therefore deferring the decision to really commit to AI because in their mind it's a buy thing. They're going to buy one day, but they're going to buy next year instead of this year. Uh, because that uh, fits with the way that they think about the world. Is that mistake because a lot of finance people are having this build versus buy dilemma right now.

Speaker B: Yeah, I think it's a very common thing, particularly in finance because if something goes wrong it is quite painful to fix it. Imagine if you tried to bytecode your own version of Xero and then six months down the line you realized it had all gone wrong and you needed to go back to zero. Suddenly six months worth of transactions that you're going to have to painfully reconcile, um, because you made a mistake at this point. So I think it's a very, very reasonable approach to take. But on the flip side, a lot of this tech is already saving people so much time. And we're at the stage now where you can kind of figure out all these things that work, figure out what doesn't work, and then you can really get ahead of other people by just being up to date with all of this kind of stuff. And I've been thinking about this kind of buy versus build decision quite a lot because at uh, flexible we have a lot of workflows that we're trying to automate. And some of these things we ideally, in a perfect world we want to automate everything and just build it all ourselves. But that's not practical. So what we tend to think about is like what workflows should we just buy a product that exists today versus which workflows can we vibe code something ourselves or automate ourselves? And I think a big part of that comes down to a few very specific factors. The first one is how complicated is the workflow that you are trying to solve. And a good example of that is payroll. Like there's so much stuff that could go wrong with Payroll and so many different edge cases that if you went and vibe coded your own payroll software, that would be pretty impressive but also pretty difficult. And when the tax rates changes and various other things happen, your app is going to struggle probably quite a lot. So that's the first one. Like how complicated is it? If it's very complicated, probably buy a piece of software for that. The other one is like what are the edge cases that could crop up? And I think a good example of this is your uh, sort of general ledger. Because so many different transactions that could come into a general ledger. So many different things which you need to be able to handle in a certain way. And if you vibe code your own tool for that, that's again going to cause a lot of problems because two months down the line an edge case comes up and you need to fix that. And then more edge cases just continuously pop up and you end up spending so much time trying to fix all these problems that for 20 to 30, 40 pounds a month you could have them all solved just by buying a product. And then the third category that I would say is what is the impact and importance of uh, the solution or the problem that you're trying to find a solution for? Because if you are building uh, like a dashboard or something which the board are going to be looking at on a regular basis and they're going to maybe make decisions from that dashboard, that's probably not one that you want to fibcode something for because if that's wrong, that could have quite a material impact on the company. So probably would buy a tool in for that which has a bit more robustness to it. So then on the flip side, I guess the types of things that you would want to vibe code are things that probably don't fall into any of those categories. So if it's a tool just for your own use and it's fairly small workflow, that's probably a perfect example of something you could uh, you could vibe code similarly if it's just an internal tool for the finance people and ah, actually the output doesn't matter that much because you just want it to be directionally correct or within a good margin, that's probably like a good thing that you could vibe code similarly. Um, if it's something that you just want a very quick solution to, you're not too fast if it goes wrong because it's such a small thing that you can fix it easily yourself. That's another good example of something that you can five code. So when I think about that Kind of buy versus build discussion. They're the sorts of things that kind of cross my mind.

Speaker A: I uh, completely see that. I think there's lots of wisdom there. As you were describing those things, I was thinking about some of the compliance changes. Right. So if you think about a typical startup or scale up business that you might have in the startup CFO community, you know, it might be a business out of the UK or from the Netherlands or somewhere growing 10, 20, 30, 40 million in terms of revenue, multi market. And so you've got compliance in the uk. So we've got like HMRC with the MTD making tax digital program where essentially HMRC are uh, setting themselves up to look over the fence into your accounts in the future. Uh, in the US you've got this very complex setup with sales taxes where it's different state by state and not just the rates but the rules are different state by state. Things like Avalara help you to support that. And then in parts of Europe they've got a, ah, phase rollout of E invoicing, uh, again similar to HMRC's MTD program. Um, another thing where the tax authorities are having a little bit more of uh, a view on um, what's taking place in different businesses and how those transactions are building. It's definitely a world where you don't want to be, you know, kind of messing up those things or even just having to jump through hoops to meet compliance requirements. You in the end want to be on a platform where somebody's done the hard work behind all that compliance, it's been approved, it fits those use cases and you're reducing the chances that you accidentally trip over any rules that you shouldn't be tripping over.

Speaker B: Yes, exactly. Yeah. Can you imagine trying to vibe code an app which handled tax in multiple different countries? That would be impressive, but uh, very difficult.

Speaker A: Yeah, yeah. I think a lot of CFOs would think that was borderline insane, but you're quite right. If somebody got there and maybe somebody will, uh, it would certainly be very impressive. Moving the conversation on a bit at this moment where finance execs are sort of maybe finding it hard to juggle the day to day with looking after their own learning and development. Right. How would you help a finance person to think about structuring their week? So there's always a little bit of learning and development time where they can try some AI and that doesn't need to be like full whack vibe coding something. It might just be sort of finding a use case which is real which is tangible, deploying some Claude to it, um, just, just, just establishing like a, ah, routine by which they can learn.

Speaker B: Yeah, good question. I think for me particularly in the early days of all these uh, AI products, having chat GPT on my desktop was such a game changer because I could just press a keyboard shortcut and it would open up and then I could ask it the question straight away. So I think that is something that I'd tell everyone to do. Just have it very easily accessible and just ask it random questions throughout the day. If it's anything like maybe you've just done a process or you've um, maybe you've just done a process and you want some ideas on how to improve it, you could ask it that or you could uh, I don't know, maybe you've had a discussion internally and you want to generate a few ideas, ask it that and then just keep pushing on different things and go down different avenues. Because I think the more you use it, the more you get used to it just throughout your day to day job, the more you start to discover new things that you didn't think you could previously do. And the other thing that I would say is speak to other finance people who uh, are maybe a bit further, further ahead in the AI journey and just get ideas from them. Like even when I speak to people who I thought weren't doing a lot of AI stuff, maybe they're a bit pessimistic about AI, I come away from those conversations and I've learned things that they think can't be done with AI and then it generates ideas for me for to be like, oh actually we could do that with AI because it could work like this. Um, so I think having those conversations often sparks quite a lot of ideas for me personally, which I obviously then go away and play around with and see where I get to and just keep that persistence to make it work.

Speaker A: I can see how that process would work. Yeah, yeah, it's really, it's really insightful to hear you talk through that. And then um, the other lens that I thought was quite interesting, right, you started your career in an accounting firm and now you're coming back to that having been a finance leader in some high growth businesses. And that must give you quite a unique lens of like what is it that founders or CFOs want from an accounting firm? Need for accounting firm where accounting firms can be really exceptional, where they let themselves down, is there is a framework for say a founder, uh, who might have to sort of pick an accounting firm just because there's some basic compliance to be done and they need to have some books essentially. Uh, and there might be like a little bit of support on cash flow planning or something. Is there a way that they can think about having a simple framework to identify what, what kind of firm would best suit their needs?

Speaker B: Um, yes, Good question. I think the first thing that I would say is the relationship with your point of contact at ah, the accounting firm is always quite important. So you really want someone who's got experience of startups. So when you talk to them about your VCs or you talk to them about R and D or share options or something, they just get it. But uh, you don't need to explain it to them, they don't need to go away and research it before they can answer the question. They just understand it. It makes it so much easier. So that sort of startup experience, probably someone who, like myself maybe who has worked in startups and in the accounting firm side of things that's always, well, obviously I would think this, but that's always quite useful because you've kind of seen it from both sides of the table. Um, another thing that I would say is an accountant that is easy to speak to, easy to contact. Like with all of our clients we have the availability on slack. So if a founder has a question, they can just slack us like they would with the rest of their team and we can get back to them pretty much straight away. And I think that's very important because you can just bounce ideas around and you can solve problems much faster than if it was an email away, then you have to wait three days for a response, that sort of thing. So ease to work with, ease of communication, that side of things. And the last one, obviously I would say this because of our uh, conversation, but their familiarity with AI and their ability to sort of figure out what the right tools are, uh, figure out what the right processes to implement AI are, uh, and be able to sort of get ahead on that side of things. Because ultimately you don't want to be kept paying an accounting firm for someone to sit there and manually process transactions in Xero. You want to pay an accounting firm because you need the expertise and you need a certain type of skill set. And that skill set isn't just paying someone to essentially do manual labor to put things into a software when you can automate all that kind of stuff and have it done so much quicker.

Speaker A: Yeah, and that's really a combination of those technical skills that you know that they're behaving in a really modern manner and they're going to be efficient and they're going to use like the latest technology to make sure that you get a high quality of service, but you're not necessarily paying more than you need to be paying. But also like this important piece around people skills, relationships and using really contemporary communications channels like being in touch on Slack and having that sense of immediacy like they're part of your extended team. It sounds like it's that combination of things which you think is really key.

Speaker B: Yeah, yes, definitely.

Speaker A: Brilliant. Excellent.

Speaker B: Rob.

Speaker A: Um, look, I wish you the best of luck.

Speaker B: Right.

Speaker A: It sounds like a fantastic thing. Uh, I'm sure be hearing a lot more about Flexel in the months and years to come. Uh, so yeah, let's definitely keep this open because I'd love to get you back on the podcast in a year or so's time to talk a bit more. I think what you're seeing through your work is a really unique lens on this AI challenge. It is a little bit different from what the CFO see, uh, but also super interesting. And uh, yeah, I've really enjoyed our chat. So thank you very much for joining me on the podcast.

Speaker B: Amazing. Likewise. Been very fun. Thank you.

Speaker A: I hope you enjoyed our discussion today. We're really proud of the podcast following we built up as we run our um, podcast in conjunction with the startup CFO community. We're able to access many of the experts who are changing the fans pace of the modern finance function, allowing us to hold these discussions, playing our role in shaping the modern finance leader and informing career journeys. In the age of AI, we're seeing the most dramatic changes in how CFOs apply themselves to supporting high growth businesses. It's a time where peer support will add more value than ever before. And lastly, if you're not in our group already and want to join, just go to startupcfo.com tech and click to apply to be part of our exclusive community offering.