'AI native' is a lie: here's how I'm actually transforming my accounting practice with GenAI. (live demo)

Trent McLaren: Hello and welcome to our afternoon session here on the summit of AI. It's been a busy, busy day already. This session is awesome. I'm super excited for you to see everything Dave is about to go through, as well as his views on gen AI, but importantly how he has then built and is using gen AI in his own firm. It is gobsmacking. We talked a lot yesterday about the level and scale of where people are in their journey, and when I look at everything Dave is doing, he's very much, for me, in the top one percent, top two percent of what accountants can do with AI. So I'm super excited for you to dig in and get into this session today.

Trent McLaren: Without further ado, I'm going to hand over to Dave Sellick, who is the founder of Sidgrove. Dave is based in the UK. He's part of the UK's XPAC group, the Xero Partner Council. This is a session that I pre-recorded with Dave last night, so this is fresh as, and I'm telling you, it's awesome. I recommend you check out everything he has built within his own operating system. But Dave, without further ado, mate, I'm going to hand over to you and let you take it away from here. Thank you so much.

Dave Sellick: Nice one, Trent. Well, thank you for having me again. Appreciate it. Oh, by the way, before I go on, we're recording this a day out from the event, so I'm hoping to be in the messages. Feel free as we go through this to put in any messages or questions, and I will do my best to answer those live. But I'm currently at Xero XPAC in the UK today. That's one of the things I'm involved with in the UK, the Xero council. Outside of that, for those that don't know me: hi, I'm Dave. I run a firm in the UK called Sidgrove. I guess uniquely, because I'm 13 years post-qualified, and normally people grow firms and employ people, but I run that firm on my own. So I have a slightly different model. I would say my firm is more like a portfolio head of finance, although I do cover the compliance side as well.

Dave Sellick: And in order to do what I do by myself, I am leveraging a huge amount of technology, in short, and to be vague. We'll obviously talk a little bit about that today again. Outside of my firm, where I'm working hands-on every day across around 50 clients, where I'm very much in the weeds, in the absolute transactional stuff, and obsessed with how we improve those bits, I speak a little bit in the industry, and I work with some firms and other accountants to help them leverage the best technology in the world as effectively as they can in their firms, to improve what they do and essentially be more effective in the way they work as accountants. And that's why I'm here today. I'm not an AI expert or influencer. I'm not an influencer, full stop. I don't have any agenda associated with any software company. I'm here to just speak as an accountant, to share my truth again, unfiltered, and hopefully I can help in doing that.

Dave Sellick: Basically, we can have a little bit of a conversation around AI again and some of the ways I'm using it. My truth from the front line. So today I'm going to set the scene initially around this concept of AI native. So let's get into it. We'll talk a little bit about how I'm using AI in my firm, but first I'm going to talk about AI native. Everyone at the moment, it seems, is talking about this term. If you've been interested and curious around AI, you'll have heard people talking about AI native firms. The big four are all looking to AI native businesses, the startup incubators and the investors only want to invest in AI native startups, and everything apparently is better if it's AI native. Or at least that's kind of what the narrative tells us. So you're kind of falling behind if you're not AI native. But what does that actually mean?

Dave Sellick: It's a good point, really. What does it mean? From my perspective, I would say it's a little bit vague, and I honestly don't really know what it means. And if I'm honest, and potentially this is a controversial thing to say, I don't think a lot of the people who are calling their projects or their firm or what they do AI native necessarily understand what that means either. Or rather, it can mean different things to different people. So it's very vague, much like this AI overview is suggesting: it means being native to AI in an AI native way. Okay, what does that mean? Not a lot, really.

Dave Sellick: But before I go on, I mentioned a little bit about my background. Before we start talking more about AI and what I do, why might you want to listen to my views on all of this, my apparent scepticism towards AI, but also the way that I use AI? Well, as I mentioned, I'm an obsessive early adopter of technology. I have a desk that looks like a NASA spaceship. I am a creative, and I do a crazy amount of experimentation with technology. In the AI space in the UK, I'm probably known for being one of the biggest sceptics of AI, but also probably one of the biggest users. I'm using multiple different Claude accounts. I'm running agents 24/7. I'm burning myself out trying to find out how best to use this technology. And so the reality is, if I'm not doing something or I'm being sceptical about something, then you probably don't need to use that thing either. But obviously that's for you to decide.

Dave Sellick: So, back to AI native: what do I actually think it means? Because we're in a period where this form of AI, gen AI, which we'll come on to in a second, is being hyped, my understanding of AI native is essentially taking generative AI and putting it at the core of how, in the accounting context, you run your firm. That usually means using these crazy-looking agents you can see on the visual there, working at the core of your practice, and autonomously, so without any involvement from a human, basically completing tasks from end to end. This is the core of a lot of the hype around what AI native means: the ability to take humans out of task execution by using these agents to do it autonomously. And in firms, in practice, AI native therefore means gen AI agents sitting at the core of the firm doing task completion. Okay.

Dave Sellick: So that's what I understand AI native to be, and what I think most people are referring to in this hyped period of generative AI. What is generative AI? To understand whether this is a good thing to do, we need to understand whether gen AI is good for accounting. Cool. So what is generative AI? Well, we'll hear most people talking about generative AI as simply AI. And I often make the point of speaking about generative AI separately from AI.

Dave Sellick: The reason I do that is because AI has been around for a long, long time. If you've been using cloud technology or any progressive accounting tools, it's most likely that over the last 10, 15 years you've already been using AI day in, day out, through things like optical character recognition tools, things like Dext, and machine learning in the accounting ledgers that we use, Xero, QuickBooks, et cetera. These are all using machine learning in some of their suggestion tools, and have been for a long time.

Dave Sellick: But the real hype of AI has come to the fore through this new form of AI called generative AI. It is a form of AI that is extremely marketable. It is also incredible technology, and it has this incredible marketing duo of instilling fear, because we all worry: what if we don't use it? What if we fall behind? It's going to take our jobs, it's going to take over. That taps into the fear response, the amygdala. And then there's also this element of hype, which activates the dopamine response. Both of these things, fear and hype, drive us to take action, to do something. In this instance, they want us to buy AI, they want us to use AI more. And so it's incredibly marketable, and it has that sense of awe, that wow factor. It's incredible on the playbooks. And so it has been heavily hyped, and it's been a sensation.

Dave Sellick: So this is what we're talking about: generative AI, this new form of AI, which scientifically is defined as large language models, or transformer-based large language models. And these are effectively, if I simplify it massively, very, very advanced autocomplete technologies. They're based on pattern recognition. They're what we refer to as autoregressive: they look at everything that you've given them context-wise and that's come before, and they're effectively using advanced mathematics, which we'll come on to in a second, to predict what comes next.

Dave Sellick: And this is a really key point, you'll hear this a lot when people are describing generative AI: they are probabilistic technology by design, which means they're optimised to be plausible, not correct. So they are probably correct, not correct. And they are very good at hiding when they're not correct, because they are basically, as I say here, optimised to look correct, not to be correct.

Dave Sellick: The way they work, like this diagram is showing, is they take the context of what you've provided them. That could be a bunch of different text and files, it could be a video, it could be an image, and what you prompt it. They will then split this up into things called tokens, which you would have heard of, which, if you take text for example, is around three quarters of a word. They'll then convert those into vectors, so it gets converted into numbers, and then through the model, this transformer technology, and other bits that sit within the model, weightings, a body of language, et cetera, it will predict the next most plausible token.

Dave Sellick: In text, that is the next three quarters of a word. And it will then predict the next one by looking at all the context that came before and what it's just written or predicted, and keep doing that and doing that until it gets to the end and says, ta-da. It is a maths machine. It has no actual sentient understanding of what you're asking it, other than doing a very, very advanced mathematical calculation.

Dave Sellick: Okay, so that's generative AI, that's large language models. So what are these agents? What is agentic AI? How are these running autonomously? Well, they're running autonomously through something called agentic AI, or agentic generative AI, which it most often is at the moment. This involves software, which is not AI, basically looping back and forth to the large language model and asking it, what next? What next? It comes back with the output, and it then asks what next, until it gets to effectively completing a task.

Dave Sellick: So effectively what you're doing is looping over a large language model. And as I mentioned, if that large language model is probabilistic, so probably correct, not always correct, and you're continuously looping over it, you can imagine what happens when it starts to go in the wrong direction. It amplifies it. But essentially, this is what agentic AI is. Cool. And it uses a lot of tokens, as you can see. It's very expensive. We're not talking about that today, but it is very expensive technology as well. That's a whole other issue.

Dave Sellick: Okay, and like a chocolate teapot will always melt when you try and use it to make tea or coffee, generative AI will always, always, always make mistakes on end-to-end autonomous work, by design, eventually. It might not the first time, the second time, the third time, but it is a maths machine that is using probabilities, and through literally the maths, the fact of the maths is that eventually it will give you an answer that is plausible but not correct.

Dave Sellick: So anyone that says it's just going to get better to the point it will never make mistakes is telling you that they think you can create a chocolate teapot that will never melt. So why, based on everything I've just said, would we want this to be native to the way we do accounting? Why? We don't. As far as I'm concerned, we absolutely do not want that native. We don't want a probabilistic technology that makes mistakes by design to be embedded into the core of how we do accounting natively. It's mental.

Dave Sellick: So my view is that it is a lie that that is the best way to do accounting, that that is the future. And actually, to give credit to those that say they are AI native, or the software that's AI native, the firms that are AI native, I don't think they actually mean what I've just explained. I often think people are using this term to just mean, we use AI a lot. We use AI as the first thing we'll turn to to solve problems, et cetera. They use Claude a lot. And I actually think if you look at a lot of the stuff that's thrown around in our profession and generally, and obviously this is my opinion, if you replace the phrase AI native with the word hype, you'll probably get something closer to the truth of what's going on.

Dave Sellick: And as I mentioned or alluded to, because it makes mistakes by design, my view is that if you're running any of these programs that drive autonomous large language models, or even just do repeated work with large language models through the chatbot over and over again without finding another solution, it will make mistakes eventually. And my view is that firms or accountants will realise this is what happens, and eventually they will turn it off. Which is why I believe eventually we'll realise this is toxic and it will need to be turned off, and it will need to be ripped out, much like asbestos needed to be ripped out because we found that to be toxic too. That doesn't mean it was ineffective for the thing that it was doing, to some degree, which agentic AI is. My view is there is just a better approach that is not toxic. That's all we're going to talk about.

Dave Sellick: So, as I mentioned, it's a false economy. People are talking about using things like generative AI agents, autonomous AI, to save time: you can save time by running AI so you can spend more time doing advisory work. Great. Well, that's like telling someone, in my opinion, to buy ready meals that are nutritionally deficient so they can spend more time in the gym, and that will get them into a better state of health. It's a false economy. In the same way, if you run gen AI agents that make mistakes by design to save time so you can spend more time doing advisory, it's a false economy, because you're doing it potentially based on information that is materially incorrect, and eventually will be materially incorrect. So it's a shortcut with shortcomings.

Dave Sellick: So, with all that said, what is my approach, then? If I don't agree with what it seems most of the profession is pushing at the moment, which is LLMs basically doing repeated work, AI native, what's my approach with generative AI? Because it is still great technology, and I'm still a significant user of this technology. Number one: use it as a co-pilot. This is my mantra for gen AI, which is not human in the loop, but human led. I'm not a fan of human in the loop. I think most of the ways that software or people introduce this concept of human in the loop is more like human as an afterthought. To better describe what works more effectively, in my opinion, you should be thinking about putting yourself in the lead: leading the technology, not allowing it to lead you, not just reviewing it, but instructing it.

Dave Sellick: So, as I said here, you use the technology to build the tools, or to do the process that you're guiding it through, and then you're optionally choosing to execute on it. You're not allowing it, where possible, to autonomously execute a mission-critical workflow. Secondly: know enough to be dangerous. It's fine for non-mission-critical stuff, like designing some stuff, writing emails, experimenting with designing some apps and things, to just go and attempt work without having any understanding of that particular discipline.

Dave Sellick: But my view, and what I've seen, is that the more you understand the domain you're using generative AI in, the better the outputs you'll get from it. For me, it feels disproportionate, and that's just a feeling. But I feel it's worth knowing enough to be dangerous within a discipline. I mean, even use the LLM, the AI, to quiz it and do some research into the discipline before you start using it to do things autonomously within that discipline.

Dave Sellick: I've done this diagram on the right, where I feel that quality output with generative AI is proportionate to the extent to which you have domain experience. In accounting, that would be someone using generative AI to do accounting who has significant experience, versus a business owner who has none, or even worse, running it autonomously without any human leading it. And then the other part of this is the extent to which you are letting the AI effectively run by itself autonomously. The ideal scenario, for me, is that you lead it, and that you lead it from a place of having some domain experience. The worst case is autonomous AI without any domain experience.

Dave Sellick: And thirdly, the thing I'm going to focus on for the rest of this presentation: build with generative AI. It is an amazing tool to break down learning curves around building solutions, workflows and tools. And specifically in the coding space, it's a really effective tool to build out solutions for accounting. This is the area of generative AI that has been the most transformational for me, particularly in the recent past. So that's what I'm going to talk about, and what I've been doing with it. Really importantly here, my view is to use generative AI to build deterministic tools. It doesn't need to be code. It could be a spreadsheet. It could just be a workflow in Zapier. It could even be a template for an email or something. But build deterministic tools that run without generative AI.

Dave Sellick: I've spoken a bit about the cost of generative AI, the economic unviability of these big AI labs, the real cost of API tokens. For example, in my instance, I worked out that I spent around $50,000 to $60,000 of API tokens equivalent at real cost, or the cost as publicised by Anthropic, last month alone, using the subscriptions that I'm using with Anthropic. So that is not ROI for me. To be honest with you, I'm making use of the subsidised scenario we have at the moment, but it certainly shows that if you're embedding this into the way you work going forward, the cost can really ramp up, as well as the restrictions in terms of the technology I mentioned earlier. The other thing I would mention is that the market is due an adjustment. At the moment there are significant issues around the economic viability of these AI labs, who are losing huge amounts of money.

Dave Sellick: They're also tied into this circular finance loop where, ultimately, they're dependent on data centres and other players also being successful to continue to deliver gen AI at the levels they are currently and the cost they are currently. There are environmental issues to consider. There are social issues to consider. Basically, what I'm trying to say is, ideally try and remove gen AI from the centre of your workflow in terms of what you're reliant on, and build deterministic tools.

Dave Sellick: What are deterministic tools? Because everyone talks about that a lot as well. Essentially, many of the tools we're already using in accounting. If you're using probabilistic technology, you're putting in the same input and you'll be hoping to get the same output. If you're using deterministic technology, then you will get the same output, or at least you'll have visibility of how that technology is going to process what you give it, so you'll know how it's going to process it before you even give it that input.

Dave Sellick: So things like fixed rules, tools like Power Query, spreadsheet formulas, open APIs, code. I build workflows in technology like Stream Deck and Zapier. These are generally speaking deterministic tools, and we use them a lot in accounting, and they're great to use. So these are the kind of things we want to try and build using generative AI. So how am I doing a lot of my building at the moment? Yes, I am building formulas and spreadsheets and automations and Power Query formulas and queries. But a lot of the building I'm doing at the moment is with code. For context, I do have some domain experience. I've been on coding bootcamps. I've been using code every single day for the last 10 years or so, within VBA, M code, Power Query, Apps Script. So I have a level of understanding, but I've taken that to a whole new level with generative AI.

Dave Sellick: So, vibe coding, for those that don't know: effectively speaking in plain English to a chat interface, Claude Code or a chat interface, and asking it to create something for you. And it writes the code that creates that tool. You can use a whole bunch of different methods to help you with that, like taking screenshots, or pasting in websites or tools that inspire you. But essentially, at its empirical level, it's speaking to the chat interface and getting it to do the work through its agents for you. My go-to is Claude Code, but there are a bunch of different options you could look at here, like Codex by OpenAI, Lovable, Replit, Cursor. These are all options for vibe coding.

Dave Sellick: So this is the tool that I spend most of my time building, curating, evolving and using with my clients now, and it sits at the centre of my accounting workflow. I've been building what I refer to as this orchestration layer tool for around the last eight months or so. Since the vibe coding side of things became good enough that I felt it was producing tools that were decent enough to roll out to my clients, from around January, February 2026, I've been exploring this technology non-stop to see what its potential is to transform the way I do accounting. Before I was using this tool, effectively I had some great spreadsheets, Google Sheets. I've done a bunch of different webinars on what I referred to as amplified spreadsheets: connected spreadsheets that pull from open APIs and push data back into platforms like Xero.

Dave Sellick: And they worked great, but ultimately they still weren't particularly scalable tools. They were disjointed. When I made an update in one client, it wouldn't update all of them. So I basically went about taking all of my spreadsheets and turning them into a product, one app basically. And that's essentially what Sidgrove Intelligence is. It's a productisation of essentially all the work I've been doing over the last 10 to 15 years as a head of finance, and then a practice, working with multiple different companies at the same time.

Dave Sellick: And as you can see, it sits at the centre of my workflow. It brings in data from a bunch of places and pushes data largely into Xero, but also into Slack, which is my communications tool with my clients. Importantly, it's augmented by generative AI. It's not AI native. Wherever possible, I'm exhausting building deterministic tools as much as I possibly can, and then I'm using generative AI to fill in for review, exceptions, anything else basically. But essentially, it's using generative AI to build tools that run without it at their core, wherever possible.

Dave Sellick: Cool. So I'm going to do a little bit of a demo, just because I like to show the tools that I'm actually talking about. I think talk is cheap to a certain degree. There's so much talk about AI at the moment that I just want to do a quick whistle-stop tour of the app, so you get a feel for some of the things that are actually built with it, what it looks like, the functionality. So I'm going to come out of the presentation, and hopefully you should see there now the overview for Sidgrove Intelligence.

Dave Sellick: There is a login process initially. I won't log out, just in case it turns off my demo mode, to be honest. The login is through either a Google or Microsoft login, so I'm delegating the authorisation to log in to Microsoft and Google. Then there is a practice view, which is what you're looking at at the moment. My clients will have their own view for their specific company, or multiple companies if they work across multiple companies.

Dave Sellick: So I'm just going to take you through some of the functionality. There's no particular structure to this; I just want to give you a bit of a view of what's in this app and what I've built, so you can see some of the potential. I'm not suggesting you copy it or do the same thing. This is just to show you something real. This is in action at the moment across my clients, like right now. I am in a migration period in terms of transitioning to full use of it, but many of the features in here are being rolled out and used across my client base. The overview page I'm mainly using for navigation purposes at the moment. It does have some insights around cash balances, transactions going through the bank, certain metrics around their staff levels, deadlines and things like this. I'm trying to optimise this a little bit. I can click through to these and then go to those areas of the app.

Dave Sellick: At the top here, I've got various different notifications about things I might want to resolve, like overdue payrolls in this instance. There'll also be deadlines and things that will come up here if they're overdue or coming up. I'll come on to these bits at the top right, but these are essentially some features in relation to my AI usage in the app. So what I'm going to do is click into Circle, which is my demo company. This is completely made-up data. By the way, these links here will also take me straight into the different platforms I have for that specific client. So this is a nice way to navigate quickly to the Xero for that particular client, or the payroll for that particular client, et cetera. That's why I've made this my homepage, essentially. When I log in, it's the first thing I see, because I can navigate from here to anywhere.

Dave Sellick: Cool, so let's go into one of these companies. The first thing you'll see is this overview page, which shows a bit of information around the client's cash, a view on recent financials, deadlines, cash flow. Again, this is something I'm trying to work on and improve. It's a constant work in progress. I've also got this aspect here where I can change what's on here, so I'm building out these different widgets, so that for a particular client I might want to put different things on here. So that's the client overview. The first thing I've got here is cash flow. This is a tool that is showing live cash flow, the cash flow as it is right now, by supplier, which refreshes from Xero every single morning. It's giving my clients a view of their cash flow at supplier level. This is grouped at the moment by account code.

Dave Sellick: I could show that flat, and then I can sort this however I want to see the data, so I can go and have a look at what's driving cash in and cash out. I can also model, on a basic level, burn and a projection of my cash flow at the top here. So it's a nice tool. My clients really love using this tool, particularly the smaller companies that may not be so obsessed with looking at P&Ls and balance sheets.

Dave Sellick: It's an evolution of looking at the bank statement, without me doing, importantly, any accounting work, because this is essentially just replicating the bank account or bank accounts. Forecasting is a work in progress: short-term P&L and cash flow. This isn't three-way; it's just a quick tool to flex for clients that want to do a bit of forecasting, and then fixing budgets on the P&L side. There's also a short-term cash flow tool, feeding in from the different ledgers and data that I can pull through from Xero, and the bank statement.

Dave Sellick: The next one is the management pack. I was previously doing this via Excel, and I still am, but at the moment I'm trying to replicate it in something I can provide in an app interface, so that there's more functionality, where they can drill into items. They'd actually see the attachments here, if it was pulling through attachments; I have none in this demo company. Effectively what I'm trying to do is replicate what I was creating in Excel in a productised version, because as much as I love spreadsheets, this will be created instantly, out of the box, once the Xero data is complete.

Dave Sellick: And again, you've got that additional functionality, like being able to drill down into the items. What I've also done is bring some gen AI in here. These things that are being flagged are being driven, I think, by a combination of rules and also generative AI that is being run over the financials and identifying any exceptions. Again, a work in progress, but that's the management accounts side of things. I do also have an area where I can upload my own packs in here.

Dave Sellick: Cool. Live reporting is my version of Xero's reports. This is effectively reconstructing what you would typically see within Xero and putting it into an interface where we can see everything. We can see the attachments. We can see all of the transactional breakdown. We can then go in, if we want to recode something, and we can come in here. I don't know what's going on there, but it typically will open the find and recode with the contact, and it will then search automatically. So I'm trying to work this incredible workflow into the tool, within the app, rather than having to go into Xero itself. And I share this with my clients too. Bookkeeping is the way I get information from clients. This pulls through the outstanding items from Xero, and that will then get sent to them. So I'll just show you how that looks.

Dave Sellick: So once I've refreshed this, I'll send it through to them. I might make some comments in there, and then they'll get a notification here that just says it's ready for review. I've got a whole bunch of those across the app, various different notifications that get sent through to Slack that then point them towards the app. Year end: I've got working papers in here. I also have the actual year end uploaded in here. That again would be shared with the client through Slack, just for a provisional sign-off, and I would then look to do a final sign-off through DocuSign. Cool. VAT, or GST in Australia. This is not very well served by Xero, so what I do is drop in the VAT audit report, and it creates something that looks like this, which is much easier to navigate. I also run some checks in here, some deterministic, rules-based ones, and then also some generative AI here to pick up on anything else I might want to look at.

Dave Sellick: Payroll is pulling through from my payroll software, Paycircle, and is identifying any variances month on month. I do actually have some generative AI built in here that would explain any variances if there are any. And then from here, I can post journals across all my clients by going one by one, hitting post. And I can also post the spend money transactions here as well. So a really nice workflow around posting payroll as well. I've got some gen AI-driven checks in here.

Dave Sellick: I've got a whole bunch of different month-end schedules that are closed systems, so they're never out of balance on the balance sheet. Things like prepayments and accruals, where we're actually using gen AI to look at the invoices and extract the dates from them. That's all accessed through the live reporting, or if we go back to month-end schedules here, and go to the live P&L through here, I can grab an item, not that one, because that's a manual journal.

Dave Sellick: But if I grab any of these invoices, I should then be able to just set a prepayment or an accrual from here. It will add it to the schedule. If it has an attachment, it will try and look at what the dates are on the attachment. Obviously I can change this, or I can just hit tick. So a really nice workflow around speeding up prepayments and accruals and similar schedules, in the context of it being a closed system. It does checks versus the ledger. Similar for payroll reallocations, direct cost reallocation, stock postings, corporation tax recharges, doing the various adjustments on corporation tax, R&D accruals: all the month-end schedules that I previously had in spreadsheets, I am doing within this app, in a closed system, in a very efficient way.

Dave Sellick: So they are posting journals into Xero, they are pulling data down from Xero. It's this two-way sync via the API. Not MCP, via the API. MCP would be getting gen AI to go and do it for me, effectively putting a roulette wheel on top of an API. Wherever possible in this app, I'm using the API to connect the data, which means that it will work the same way each time. Deadlines, pulling through from Companies House in the UK. Company information, pulling through again from there. To-do lists with my clients, notes, key links, a bespoke tool here. And then I have group functionality. For groups that I manage, I also have this intercompany functionality, where it's bringing through a matrix and doing side by side across different files. It's a really nice tool, with similarities, I guess, to tools in the market like Mayday.

Dave Sellick: Wow, okay, so I'm talking pretty quickly. There are so many things to go through here. I'm just going to go back to the practice view and very quickly show you some of the tools I've got here. Again, a personal tax tool I've built, which links through to the companies where people are directors. Something called Setlist, which is basically my to-do manager, which again ties into task management within the app. Payroll journals, which I mentioned. Forms, which are currently being used just for new starter forms, or at least I'm building that at the moment, predominantly because I was using a form software that I wasn't getting huge value from, other than for payroll new starters. I also like the fact that it can link in with payroll within the app. Designer: if anyone's seen my LinkedIn posts, this is what I'm using now to create those visuals quicker. It's just something I experimented with.

Dave Sellick: Send a report: this is the function I use to upload documents that go through to my clients and then get pushed through into Slack. Notes: again, it's a notes technology I can use with my clients, within the app. It's a bit like Notion, to a certain degree. Deadlines that I can see across the entire firm. Automations: this is where I'm keeping track of all my automations, my use of AI across the app. I even have a market review, where I'm looking at all the different AI models and whether I'm using them as effectively as I can across the app. Notifications, where I manage notifications into Slack, my communications tool, and also into email and WhatsApp where I'm using those, which I use sparingly. Proposals, something I'm experimenting with at the moment: the ability to create proposals in here for sign-off. A chart of accounts amendment tool; that's typically very hard in Xero, so I've created a tool for that.

Dave Sellick: A screenshot-to-spreadsheet tool: I take a screenshot, paste it in here, extract it, and it will use gen AI to pull the data out into a table that I can then copy into a spreadsheet. This is a tool that I use across a bunch of niche workflows in my practice. And finally, something called a spend importer, where I drag in or copy in items from Amazon order reports or Uber, and it turns these into spend money transactions, which I can then apply coding to, even project coding. It will also group together transactions that are paid on the same date. So it really optimises the way I approach my bookkeeping around things like Amazon and Uber. Perhaps that's a video for another time. Loads of stuff I've been through there.

Dave Sellick: Let me just take you into Slack. I haven't been able to go through all the different notifications, but you can see some examples. For example, on the cash flow side, I send a weekly notification showing all the cash flow movements. I have a flow around bookkeeping, as I showed you, the management pack, VAT, and a bunch of different notifications that get pushed into Slack to cover those off. I also have automated notifications that come in when things have been filed, which are quite helpful as well. Cool. Okay. So that was kind of long. I just want to do one final thing, actually. We'll come on to this in the presentation, but if any of that stuff looks helpful, in terms of wanting to understand how you'd even start on any of those features, I have built something called Stems. I've taken every single module in what I've built, and this is continually updating, because I'm continually building and improving.

Dave Sellick: I am essentially going to be making some of this, if not all of it, public, so that you can come into any of these modules, download the code, and copy a prompt that, when you put it into your Claude Code, Lovable, Codex, et cetera, effectively rebuilds what I've created, and then you can curate it into your own environment from there. That's just one way I want to help people get a bit more of an understanding of how to use this technology.

Dave Sellick: And I know I'm tight for time, so I'm just going to finish off by talking a little bit about how I created this technology, some of the structure around how I vibe code. Claude Code, as I mentioned, is my go-to. GitHub is where I then send the code. It keeps my code and the version control; it's almost a bit like a Google Drive for storing code. That will run some tests as well whenever I push code. That then interacts with something called Vercel, which effectively hosts the code on a site. And then Supabase is the database I use behind it.

Dave Sellick: All of these technologies are really well suited for vibe coding, and this is the stack that I use. So, some of my key workflows in vibe coding. I don't use many skills or plugins, particularly. These are the two that I probably use the most, or that are most central to what I do. The first one is called shadcn, which effectively creates a UI/UX design system for your app to use, so that your design stays relatively consistent across the app. Impeccable.style is one I recommend to avoid your app looking like AI slop. It's literally designed for that. It looks to optimise against the default slop look and just improves the design of the app, basically.

Dave Sellick: I do use files, or text files, to give context to the AI agents whenever they're coding. I have three key tools: one is called Bible, the other one is a roadmap, and the last one is called handover. The Bible essentially sets all the default styling that I want across the app. Any ideas I have about what I want the agents to be thinking when they're coding, I will ask it to put into the Bible. That could be things like, I might say: every time you approach the design of a page, I want you to ask the question, does it make you feel joy? Does it make you feel more human? Is this as inspiring as it can be? Maybe rate that out of 10, and if it's not a nine, then start again. These kinds of things, as well as some more basic things around how I want it to approach errors that are being made when it's coding. Anything and everything that I want the agents to consider when they're coding each time goes into the Bible.

Dave Sellick: The roadmap is exactly that. Anything I'm looking to create, it will add to a roadmap, and it'll pick it up autonomously when it's got nothing to do. And the handover is what I give it at the end of the day, or what gets populated at the end of the day, before it starts a new session. Then I just say, let's go, and it basically picks it up. I have a product called Soundcheck that I use to gamify improvements in the app. So quickly, if I just go into that, that is here. Nothing is open at the moment, but essentially it gives me different options for things it's going to suggest coding, and then I pick the one I want, basically. It'll rate it out of 10, and it gamifies the process. Cool. And as I mentioned, I have a start-up routine where I just say, let's go, and it just cracks on and does various things.

Dave Sellick: Vibe coding is a great tool to create internal apps and prototypes. I use it for things like these slides, simple websites, brand assets. But just be aware; I always want to make it clear that there are risks to this. If you're creating an app with client data and you don't have domain experience in coding, and you haven't had an engineer look at it, which I have in my instance, just be wary of security. Be very wary. Make sure you're working with someone who knows what they're doing. It's fine if you're not putting client data in it, but be a bit more risk averse. Then, be wary of rabbit holes. It's very easy to build things that perhaps you can buy off the shelf. Generally speaking, I'd say if you're using a product that's well priced and that you like, you don't necessarily need to go and build it.

Dave Sellick: And the other thing I would say is, don't build 72 different apps. My advice would generally be to build one app and then put 72 features in it. I hear a bunch of accountants talking, and they may even be talking about it at this conference, about the thousands of different apps they've built. For me, that's just thousands, or 72, different ways in which the data can be insecure. So, key takeaways. AI native is a lie. It's built on a technology that is probabilistic and makes mistakes by design. This is the last thing we want at the centre of our mission-critical work in accounting. If you're using generative AI, then default to using it as a co-pilot, never the pilot. Do not let it lead you; you need to lead it. Use it to enhance. Use it to build. And ignore the hype. Go and use the technology. Explore for yourself. Be wary of those that are selling it, and the agendas that they have.

Dave Sellick: And if everything I've just said, and all the fast talking over the walkthrough, which I kind of had to do because there was so much to get through, has gone over your head, or you're still feeling disillusioned, my advice would be: start in chat and stay in chat. This is still an unbelievable way to leverage generative AI. Those that are using Claude chat or ChatGPT in a really effective way can get most of the gains that most people who are using AI can get. I think if you start moving beyond those without a clear idea of how to add value with autonomous agents and vibe coding, then I believe it can do more damage than good.

Dave Sellick: And you're probably miles ahead of most if you're just using the chat interface effectively. So yeah, that's all I'd say. I've mentioned Sidgrove Stems; keep an eye on that on my LinkedIn. And on that note, thank you. If you want any more information on what I'm building, or what I've spoken about, or the Stems solution that I'll be sharing to help people get a fast track on vibe coding, then just keep an eye on my LinkedIn. That's all, really. Thank you, guys. Thanks, Trent.

Trent McLaren: Thank you, Dave. That was awesome, mate. I love everything that you're doing, and I'm so supportive of it. And I've said this to you before: keep going, keep building. We're excited to see what's happening and where you go with everything you've just showcased. Other than that, thank you for joining this session. I hope you've enjoyed it. If you like Dave, you can follow him on LinkedIn. Thank you so much. Bye bye.