Where to start with AI: Lessons from the industry (Panel Discussion)

Trent McLaren: So, we've spent two days talking about all the things: all the different connectors, MCPs, ATO data, XPM data, pricing data, the works. Today, this session, we are talking about where to start with AI, and the lessons from our industry. The rule I gave to everyone, well, I gave it to Jimbo specifically, because I knew this would be… it's already in the chat, was to not say "just ask AI". But I'll throw it out there.

Trent McLaren: You could ask Claude how do I get started with AI, or Copilot, or Gemini, or whatever, and it'll tell you how to get started. What I've asked our group today is to actually focus on practical things: not just "we got started", but here's what we got started on, here's what worked for us, here's what did not, and here's what I'm excited to do next. So, my first question for our panel. Denise, I'm gonna start with you, because you're magically in my top left. Tell us who you are, where you're from, and what was the first thing you got started on with AI?

Denise Twigger: Hi, Trent, thank you so much for having me. For those that don't know me, I'm Denise Twigger. I'm the owner of BAS and Balances Bookkeeping Services, based in the beautiful, sunny Perth Hills. I've been in the industry for well over 30 years, and so AI is not something that I haven't been accustomed to. Obviously, we saw changes come through with bank feeds back when Xero first started in its day. And I guess that was the progression for me to go, holy moly, this software, along with its intelligence, can actually make my life as a bookkeeper a lot more streamlined, which was the big thing. It wasn't as clunky.

Denise Twigger: So I guess, probably the same with a lot of bookkeepers out there, my very first taste of AI was the bank feeds in Xero, where I didn't have to sit there for hours and process a bank statement. Mind you, I use Claude for that now. What I do is I say to Claude, I've got a bank statement, can you please convert this to a QuickBooks Online CSV for me? We've had some drama with open banking, and next minute, I dump the PDF into Claude and it produces a beautiful statement for me. So, even 30 years in the industry, I'm still converting PDFs to CSV to import them into banking, so we can get on with our bookkeeping duties. But obviously we do it a lot smarter and a lot easier today than what I did 20, 30 years ago.

Trent McLaren: Yeah, I mean, wow. Denise, I won't tell you what I was doing 20 or 30 years ago, but it wasn't anything remotely close to converting PDFs into CSVs. But what a journey and a ride to see how much change has happened. That's huge.

Denise Twigger: Mmm, thank you.

Trent McLaren: Yeah, thanks for being here. Hey, Reuben, over to you, you're my next square down. Who you are, where you're from, and the first thing that you started doing with AI?

Reuben Bergola: Hey guys, my name's Reuben, thanks for having me on here. I am the founder of New Wave Group, a multidisciplinary firm on the Gold Coast here. I have about 90-plus staff now. What did I use AI for first? I'm trying to think back, but I'm pretty sure I was just popping in my emails and trying to get it to respond for me in a streamlined manner. All of us are probably bombarded with so many emails, and just drafting, just getting started… in that email section, that's where I got started: popping an email in there and saying, hey, can you just start it at least for me, edit it, send it out. And it just saved me a heap of time. So that's where I started, Trent.

Trent McLaren: Yeah, that's awesome. 90 staff too, man, that's huge! I've got some questions for you after hearing that part. I feel like I missed that in the brief somewhere. Welcome, we're glad you're here. I'm gonna come back to you on all those, particularly about how you're doing this with big teams, because that's very much of interest. Andy, you're my next square down. Welcome, good to have you here. Tell us about yourself, tell us about the firm and where you started with your AI.

Andy Teece: Oh yeah, thanks, Trent. I'm from Atomic Business Advisers, we're in Sydney, and we're sort of heavily involved in AI and have been for some time. I suppose the first actual business use that I was involved in was back in the first quarter of 2023. In Q4 we run planning and strategy sessions. And there's certain template wording for, you know, how a bucket company might be used, what estate planning might look like in terms of you need a will, enduring power of attorney, enduring guardianship, just to give the clients an understanding of what those documents are in nice, succinct wording.

Andy Teece: And I used it for that. It was fantastic. I felt like I created fire. And so, yeah, we've really pushed it from there. As we then prepared those reports for the clients, tidying up the wording, similar to tightening up emails and things like that, we used it on those reports. So that was back in the first half of 2023, when we kicked off.

Trent McLaren: Awesome! 2023, man, you guys were early on the wagon, that's awesome, very cool. We're gonna come back to that, Andy. Gus, for those that haven't met you and didn't see you on any of the sessions yesterday, do you wanna introduce yourself and what you do?

Gus McLennan: Hi guys, I'm the Managing Director of AI Dojo. We basically operate a secure AI platform to help accountants with their workflows. We run a platform that's hosted here in Australia, gives you access to all the major models, and we've started to build integrations into Xero, the ATO and some of the other key platforms, and build a whole bunch of skills and workflows to help accountants and bookkeepers get their work done. In terms of my first use, though, I'm a little bit like Andy.

Gus McLennan: I reckon I sat down in front of ChatGPT, it would have been 3.5, not even Turbo I don't think, back at the start of 2023. At the time we were running a property technology company, and we literally looked at each other and said, what should we even ask? And somebody said, oh, just ask it what the factors are that are used in the valuation process for a commercial office building. And it just started to magically spit out all of this text. And like Andy said, I was absolutely gobsmacked. I thought this was absolute black magic. And to this day, I still every now and then get surprised and delighted, so it continues to appeal.

Trent McLaren: I'm wondering if next year is the year it all starts to wear off, and we're like, well, yeah, of course it can do that, what do you mean? It'd be like when Optus has an outage, and we're like, what do you mean I can't use it today? Jimmy, James, for people that don't know you and didn't meet you yesterday, please introduce yourself and tell us about your first time using AI.

James Rose: Yeah, co-founder of Content Snare. It's a product for client onboarding, or end-of-year queries, and just getting information from clients. Obviously working a lot with connecting it to Claude and other AI tools at the moment, that's a big focus of ours, and opening up its capability there. It's funny. I know I did a whole bunch with AI, but none of it really sticks in my memory, because I'll be honest, I was not a big fan of AI until about March this year. It was always just useless enough.

James Rose: I used it a lot, had a lot of fun with it, did some silly things, helped plan some trips and stuff like this, but it was never really there. And what changed everything for me was Claude Code. Since Claude Code, and now Codex and all these harness-type products, everything changed. And I think one of my first accounting-type use cases, because I'm not an accountant, obviously: my accountant sent me a list of fuel transactions for various cars, because we'd screwed up which cards are with which, and there was no real obvious way to work out which transaction belonged to which car.

James Rose: Claude, through downloading a whole bunch of crap out of Dext, basically exporting everything, downloading it locally, and I think there were some statements or something, worked out what was for which car by looking at the addresses, and which fuel stations were likely to be close to where that car was in use. And the calendar was used as well, to work out who was at what meetings. And I was just like, this is nuts, right? That just saved me so much trawling through stuff to work it out. And I think it even found some transactions in the list where it's like, this is actually a snack. This is not fuel. So, yeah, that was such a boom moment.

Trent McLaren: Yeah, that's awesome. I got a list of please-explain transactions in a CSV the other day. I dumped it into Claude and said, hey, can you check my inbox and find all the receipts for these? And 2 minutes later, bang, here they all are. Can you then draft the email, send it on, push it into Slack? It was the easiest piece ever. To be fair, I still give most of that list to Casey on my team, who's behind the scenes and manages a lot of that. She's also using Claude, searching everyone's inboxes trying to find all the receipts. So I love all that. Good to have you all here.

Trent McLaren: I love seeing everyone's eyes light up going, oh, the first time I used AI, it was… you can see the joy in your faces. It's a fun thing to see. Denise, I love that we actually have the bookkeeping representation here. Someone was smart enough to ask you to be on this session today. Can you give me a good understanding of where you've started using AI, and where you're getting the best benefits with AI within your bookkeeping roles and function? Because I think a lot of bookkeepers are kind of scared about what AI is going to do for this industry. So I'd love to hear how you're feeling about it, and where you're getting your biggest wins from.

Denise Twigger: Yeah, thanks, Trent. Mine are actually coming from social media. None of us have got that Marketing 101 background, and I like being in control of what I'm doing with my socials, so I've ended up playing with ChatGPT. My business is part of about 120 groups, and I said to ChatGPT, can you tell me the rules and regulations on all of these groups that I want to start posting locally to? I'm about education, so if I pick up a few clients, great, but I want to educate the world of small business about how safe it is for bookkeepers. Yes, I might use AI in business, but I don't use it where there's any harm involved, and to me, that's a really important structure: I don't wash client data through it.

Denise Twigger: I'm a bit hesitant to put my own data into it, unless I'm playing with KPIs, or something that generally takes a lot of time, where I can just dump it into ChatGPT and it will wash it through. I do pay for ChatGPT, and I do pay for Claude, so it gets to know me as a business owner. It knows my calls to action, it knows I live in the Perth Hills, it knows my direct areas. It tells me the groups that I can and can't post on, so I don't get a bad admin email saying you've been blocked from advertising on our social media posts. As a BAS agent, I don't have time to go through and read 120 admins' rules and regulations, so I used it heavily for that. I've also got a daughter who's coeliac, and we're planning a big trip next year for her 21st.

Denise Twigger: I said to ChatGPT, I've got a coeliac daughter, we're looking at going to Europe, she wants to turn 21 in Italy, I know it's really good for coeliacs, hit me up with your best restaurants. And it went through all these different travel places and put together a menu for me, where I can say, right, we can go there, that's coeliac-friendly. So I use it personally a lot. I went on a cruise recently where I listened to 33 different bands, and I put it through ChatGPT, making sure I didn't miss anybody, and it diarised the 7 days that I was on the boat. It didn't allow for rough seas and bad weather. And I've also used it for a packing list, because I'm notorious for leaving something behind. So I use it as my "I'm gonna forget if I don't write it down" kind of thing, but I use it very heavily in my social media campaigns and my posting with Google My Business.

Trent McLaren: We always have an issue where someone's luggage is way over the weight allowance, and I actually asked it, can you give me a rough idea of how much stuff I could put in this bag to get to my weight allowance? I had to get it to find out what the weight allowance was in the first place, tell me, and then plan that. That's cool. Thanks, Denise, that's a great way to get started.

Trent McLaren: If you're playing along in the chat, and you're a very active chat this afternoon, which is always great to see, tell us what your first experience with AI was as well, personal, professional, whatever it is. Reuben, I wanted to come back to you. There's two parts. You're a very big builder, and I know you love building stuff. But I'm also trying to understand how you're managing that with a big team of 90. So my first question: where did you guys get started in making that decision of, we're going to go all in on AI? I assume that was you, and maybe at a team level, but do you want to talk us through the crux of how you're planning for this around a team of your size?

Reuben Bergola: Yep, for sure. I'm very lucky, because I've had the luxury of not being in front of clients. I'm lucky enough to choose what I do, and I chose to go deep into AI very early. It is much harder with a larger team, mostly because of change management. For us to get people to start using tools, even AI, it's a big, big process. So I had to think about how I'd do that in the most efficient manner. Like you said, I'm a big tool builder, and I think there are three levels of AI: basic, intermediate, and advanced. Basic are the guys that are just prompting on Claude and ChatGPT online.

Reuben Bergola: Intermediate are somewhat coding, creating their own tools, and advanced are full-on agents and automation and so forth. So how do we implement that into a bigger team? I lead with constraints, build tools around the constraints that we have, and then I champion certain people to test the tools. We have many teams within our organisation, so if, for example, we come out with a workflow planner tool, we test it in one team. That goes through a month-long process, shadowing what they do, and then we start implementing it, with change management supported over a larger level. So it is a lot harder with a bigger team. If you have a smaller team, I would just be going absolutely crazy and building AI into your team now, before you have to do the change management itself. It's a big project.

Trent McLaren: As in, if you're a smaller team, building the culture so that as you add team members, you're bringing them into that culture? Is that what you mean?

Reuben Bergola: Exactly right. When everyone's been using the same tools for the last 10 years, it's much harder to change 90 people than it is 3, or 4, or 5. So start now. This is the time.

Trent McLaren: Yeah, awesome. I want to come back and talk about some of the advanced stuff you guys are doing next time through. Andy, you mentioned 2023 was when you got started. I assume you've had new team members come in, and it's a bit of a follow-on from what Reuben just said. How are you bringing new team members into the way you're using AI and what's being used across the team?

Andy Teece: Yeah, so probably 18 months ago, I was very conscious that we need that consistency, and I'm conscious of the change management as well. So we started what we call AI in Action sessions: 15 to 30 minute sessions, twice a month, just to run through some of the key things that were cropping up. We shifted from everyone doing their own thing, because I'm a standardised type of person. I'm also conscious that as accountants we can often reply to an email and just start typing. I like my emails to kick off with "I hope you had a nice weekend".

Andy Teece: So I set up some skills or instructions that forced that, and I pushed those out across the firm, just to make sure there was some standardisation, so that people don't have to think about the pleasantries at the intro and the outro. They get embedded in the process. We have our method, the Atomic way, as I refer to it, as no doubt all firms do. So we've embedded those key things into our skills and instructions. In more recent months, rather than ask everyone to upload new skills into their own versions, we've got a business plan now, and looking back on it, I probably would have done that 6 months earlier, because we're able to roll it out.

Andy Teece: I'm a big believer in projects. The skills and instructions and your prompts are super important as well, so we've had a lot of training in prompts: being very clear, being very direct. Your ChatGPT, your Claude, whichever large language model is your favourite poison, they're very verbose. So my favourite prompt is "reduce by 25% without losing meaning". Sometimes it's "reduce by 50% without losing meaning". And it's incredible, when it condenses it, how the meaning is retained. The effect is there, but it has been able to strip it out, which is sometimes really hard to do as an individual.

Andy Teece: I prefer point form rather than long paragraphs, because I'm mindful that a bunch of clients will be reading on their phone or an iPad. So I embed that point-form instruction into our instructions or skills, and force it across the firm.

Trent McLaren: Andy, you come across like someone that's done a lot of reps with AI, and I like that about your answers. You all do, but that was an awesome example, mate. I loved hearing the team values, the culture, the Atomic way, how we're gonna do this, the prompt structure, the business plan, the shared skills. When we come back around, I want to ask where you guys are focused right now, and what you're trying to expand on within your workflows. Gus, I know you speak to a lot of people that are not doing anything with this, or are dabbling in it. What's your view of the low-hanging fruit for accountants to get some good, quick wins with AI?

Gus McLennan: Well, I was gonna show a graphic, if everyone would be interested, to basically show how the platform is being used. I'm hoping some of this will act as a bit of a prompt for someone. Can everyone see that?

Trent McLaren: Yep.

Gus McLennan: So, what we did is we took a thousand chats randomly from our platform. There was a reason we did it: it was all around chat intents, and being able to get our auto router to route the chat to the right model to respond to you. But some interesting insights came out of this classification. Maybe this won't be a surprise for some people, but the number one thing is still "help me write and make my emails better", funnily enough, or write an email, or draft a document, maybe social media as well.

Gus McLennan: After that, you start to get into compliance and workpapers. A lot of that would have been the skills we've built into the platform, like BAS prep, that our users are using. Tax advisory, document creation, and then it kind of tails off after that, right down to a handful of stuff around financial modelling. There's still a pretty broad spectrum of things that you can do with AI. But for me, the most surprising thing is that a lot of people are still using it really just to help with some of those very basic tasks, like preparing and sending an email.

Trent McLaren: Yeah, okay, awesome. Thanks for sharing that, Gus. Jimmy, you guys have been helping chase and get information for quite a long time, and there's a lot of cool automation you do in Content Snare, but is there any low-hanging fruit for ways to take that further with AI, in the way you're seeing people use AI with Content Snare?

James Rose: I mean, the big thing is just connecting it to your assistant now. It's just incredible. I know you don't want me to say "just ask Claude", but it kind of is. Just use Claude.

Trent McLaren: I'll say it once, but just say more often.

James Rose: Can I say "just use Claude"? Because there was this moment in time where all the products were putting their own AI within the product, like a sidebar that's an agent or something. And look, I don't want to insult everyone, but I always thought it was pointless, in a way. Not always. There was a window where it seemed to make sense, but it quickly disappeared, and now I believe you may as well just connect the tools to your agent, as long as you've got a good one, and let it do all the bits and pieces. Because once it's connected to lots of tools, the capability opens up as well. What's insane about this is it used to be a problem to have separate tools for everything. People want it all in one, even though it doesn't do as much, doesn't do all the things as well.

James Rose: It's sometimes just easier to have it in one place. Now you kind of do: it's just in Claude, or your Codex, or your agent, and it connects to all the things. So the way we're seeing people get crazy results out of Content Snare now is literally just connecting it to Content Snare, maybe to SharePoint, and being able to go, get our generic checklist that we would normally send to every client, but customise it for this person based on the documents we got from them last year. Have a look at what bank statements are there. Instead of just asking for "send us your bank statements", split that up into the 10 questions for the 10 different accounts we know they have. And it could just build that, right? That is the big unlock. And that's why we probably won't build an agent inside Content Snare, because it's not necessary anymore.

Trent McLaren: Yeah, I always push back on people in two ways. I would always say that your top 20% of people that like Clauding will Claude, but you'll actually find 80% of people are gonna…

James Rose: Oh, no doubt.

Trent McLaren: …like, move, so I think you kind of…

James Rose: Most people won't, but I'm just hoping they hurry up and come along for the ride.

Trent McLaren: Yeah, I don't think they will. I think you need both. I think you'll have some that sit in both.

James Rose: Because it's already at the point where it's two clicks to connect any of these tools to an agent, right? If they're set up properly: add it, OAuth, done, you're connected. And if someone's not doing that, come on. Come on!

Trent McLaren: I'll tell you a story about someone I've been pushing on this for, like, weeks, and they're like…

James Rose: Yeah, we can share stories, I've got plenty of them.

Trent McLaren: Denise, you talked about the marketing, the social side. I'm keen to understand how you feel about building more confidence with AI, because what you represent in terms of feeling nervous is echoed in all the chats I've seen in the last couple of days. People are worried about what am I doing with it, and what am I going to put into it? Do you ever feel like that's going to disappear? Are you hopeful that you can use it for all these other things as well?

Denise Twigger: Oh, totally, Trent. And I guess the guys have already mentioned it: it's a case of getting in there and dabbling with it. Like I said, I use it heavily with socials. It's knowing where you use Claude and where you use ChatGPT, because they're the two main players that I play with. I use Claude to create checklists and templates to make sure that I'm not missing anything. Obviously, we're heavily embedded with quality management systems, so it's taking the burden of administration that so many BAS agents get drowned in, and pushing that into another platform such as Claude, where I park it.

Denise Twigger: Because I pay for it, and I've made sure of my sensitivity settings, it doesn't go out to the world at large. It's helping me put together a checklist, so if we onboard a new client, I know that everything is in place, and it basically checks me as well. I know that a lot of bookkeepers are frightened about it. I know that we like to make sure that we've got control of everything, and you still do.

Denise Twigger: But you could start off with something that you inherently trust, whether it's taking a photograph of a receipt when you've been out for dinner and you need to split it between 8 different people, and make sure everybody pays their fair share. It's breaking down basic bookkeeping to make our lives a lot easier, where we're not wasting our time on the mundane stuff. Like I mentioned before, a PDF straight into Claude: turn this into a CSV for me. It does it instantly. I know that there's other applications out there that do it, but I've always got Claude or ChatGPT on, so why not just chuck it through that?

Denise Twigger: It's open, it's live. I see a lot of bookkeepers are frightened about it, but start off slowly. Start off with what's in your fridge. Take a photograph of what's in your fridge, and tell ChatGPT to make a menu for you for dinner. Start somewhere, but start.

Trent McLaren: I had you in my head the other day, because we were chatting about it, and I was like, I need to actually eat something really healthy. Stuff it, I'll do it. I've been eating minced beef bowls with veggies for the last couple of days, trying to find healthy things.

Denise Twigger: You're welcome. I'll send you a note.

Trent McLaren: So it wasn't Claude, it was you. Reuben, you mentioned before the idea of more advanced builds. We've talked about some of the beginner stages, but I'd love to hear about some of the more advanced stuff you've been building, and where that's really taking it. You did mention it might blow a few minds, so just to preface for everyone else, this is a view of what could be possible.

Reuben Bergola: Yeah, sure, happy to. Can I share my screen? Is that alright?

Trent McLaren: Yeah, yeah. Please do.

Reuben Bergola: Okay, cool. So, while I'm sharing my screen: my biggest thing is that when you do take Claude, or any AI, to the next level, it has to be done correctly. Data's a big thing, so foundationally, before you do any of this, make sure that your data is washed in a way where it's safe, and it's in a safe environment. But why I wanted to show this today is that our biggest constraint, and it always has been, is capacity and efficiency in our business. Capacity through not being able to find the right team members when we need them. People get sick, and then write-offs, and all those sorts of things happen. That's the life of the accounting firm.

Reuben Bergola: I've never been more excited to be able to bring some of that back. For me, I'm not about building every single tool. You'll see this: do not do what I've done, because this is my ADD brain just going nuts. However, you can build tools around the data and the main software that you already have. I'm not going to switch off Karbon. There's no point. The change management of Karbon is going to be mental. Same as Ignition. It would take me months and months to try and recreate my own platforms. I'm not going to recreate Xero, there's no point in doing that, but I'm going to create the glue. The glue and efficiency is what I wanted to do.

Reuben Bergola: This is what I've built over the last 6 months. A lot of things are in beta, but a lot of things are already in play. My thought process is: what are the things that we need the most within our firm? Attack those first. So we started off with simple things like creating documents that connect to Xero and Karbon, so everything flows into one. Instead of taking 15 to 20 minutes to do an annual summary, it takes us 5 minutes. Simple things like adding in all the documents from the prior year and then creating structure charts: it's a two-minute process for us. We extract it all, it puts all the information in, builds the chart for you, all the tables are done for you.

Reuben Bergola: And then as you keep going, this is where the addictive part comes in. You'll probably find yourself coding more than you should, which is what I've done. But you will start creating things and saving unbelievable amounts of time. It's all the boring stuff within your business that you can automate and systemise. I was wanting to show you, but I've got clients in there: we have got bookkeeping to 90% at the moment. Our team are more front-end now, making sure queries are out. They're reviewing the actual work that the agents are doing, but these are running 24-7 on multiple clients. And yes, we had to update our engagements and proposals and all those sorts of things, so make sure you're having a look through the TPB guidance and making sure you're following it.

Trent McLaren: Good point, Reuben. I want to quickly ask on that, because it keeps coming up in the chats, and Tanya just mentioned it, which is why it's top of mind for me. The advice we keep giving everyone, or we've heard from different people on the panel, is that you've got to de-dupe the information, you've got to redact certain things, you've got to do all those secure parts. But then the other part of it is, well, I just want to connect the MCP and let it rip, and I need to be comfortable that it's not training on the model and it's in a secure environment. You guys are a big team, big firm, lots of clients. How do you tackle that? How do you remove the friction of the redaction component?

Reuben Bergola: Yeah, that was probably the biggest part. The longest part was figuring out, with developers and security specialists, how we get that into an environment that takes detail from Karbon and all the other places. We have ours in AWS, in Australia, and it does not take certain information into the system. So that then allowed free rein to go ham on figures like this, automating capital projections and advisory, and everything just flows into one, all backed with AI.

Reuben Bergola: I think the ultimate goal is to have a localised model internally, so it does a lot of the automation internally here. We are basically going through each process in our business and getting it to a point, one by one, where 50% of the job is done. And then we're focusing our time and attention on how we create better advisors, reviewers, and people at that front end. It's getting to a point where, yes, the automation is real, it's here. And you can do this sort of stuff just with the normal browser as well. This is us just taking it a little bit further, but anyone can do this.

Trent McLaren: Yep. That is the fun part, I think. You finally have that good idea, and there's always a good time for it. Was there anything else on this you wanted to quickly show?

Reuben Bergola: I'm just trying to think what would be of value. For example, here, we've created our own workpapers system as well. This is not for everyone, but if you really get into Claude, and you are a smaller business, and you want to get the IP that you have in your minds… all of our IP as accountants, where does it live? Excel, basically. Excel is great, but it also breeds issues and double data entry and all that sort of thing. You can build that in here and connect it to Xero, which I've connected all of my clients to here already. And it just creates this seamless symphony, which is amazing. That's why I'm so hyped up about every single firm, and the change that's gonna happen in our industry. In a good way, in a great way.

Trent McLaren: It's super fast moving. And coming back to the comments, obviously you've got to make your own judgments and decisions in your firm about what you're comfortable with. You should do your own due diligence. I think I said in an earlier session today, you've got to decide. Andy, it sounds like you guys made your own decisions on this too, right? Your values, the Atomic way, and all those bits and pieces. But everyone here in the chat has to make that determination. What are we going to do? What work have we done to determine that we're comfortable, we're happy, we're safe to do the integration? Have a good IT support team if you have that as well, or bring that in-house, whatever makes sense for your size. That's awesome. Thank you for showing all that, mate.

Reuben Bergola: No worries.

Trent McLaren: I've got a few more questions to round out with Andy, Gus and James. Andy, my question for you is: with all the things you guys are doing, what are you focusing on for next year? What are your opportunities to expand your use of AI at the moment?

Andy Teece: Yeah, and I'm very mindful of what Reuben was going through there. From my point of view, we're pushing, but we're a 60-year-old firm, so I'm very mindful about getting it right. I think the Tax Practitioners Board, CA and CPA are great in their theory, but no one wants to commit. No one wants to put their hand up and say, if you do this, then you're okay. So things like getting your engagement letters right, getting your privacy policy right, getting your internal and external AI policy right, and getting your staff trained are absolutely critical. Sanitising your data to make sure that sensitive information is not going in is critical, and a lot of that comes down to staff training.

Andy Teece: My team's not as big as Reuben's, but it's big enough, and we've put some guardrails in place. This is where those instructions and skills are stopping people from hurting themselves and the firm, and protecting their clients. All that being said, where we're going moving forward: I think the large language models are wonderful as a base, but when you put some further parameters in place and train it, that's where you really get some power. So we're developing some skills and instructions that work with accounting files, again with sanitised data. We've got a preparer skill, which is probably at an intermediate level, and a reviewer skill.

Andy Teece: At the moment, those are heavily guided by an intermediate and a senior or manager, so the human in the loop is there, but we are finding it is so good, and it follows the bouncing ball. One of the things I did the other day: I put some data in, and it told me the GST was out. Normally you then go back through, period by period, and identify the period. Sometimes, if it's below a certain amount, you'll just amend the BAS, or whatever it might be. I said, can you please identify what period the error took place in? And it came back with two dates and two amounts, where it had gone wrong.

Andy Teece: And I was able to go and amend those BASs, reallocate the transactions in Xero, and make sure that we were right. That's what we're really trying to hone further, moving forward. Will the time come that we're comfortable setting that preparer off as an agent on its own, setting the reviewer skill off as an agent on its own, and then having an accountant do a final review? Yes, there is a point where that will be the case. I'm trying to sit right at the front of the curve, but just half a step back, to make sure that we're okay and our clients are protected. But it's extremely powerful.

Andy Teece: The other thing that we're developing: I run my client meetings a certain way, other staff run things a certain way, and no doubt everyone's doing things pretty well, but again, I like the Atomic way, one way. So I'm developing skills to look at the presentations, set those up, and give the accountants a cheat sheet. Because even if the presentations are set up the way I like, if my staff who are presenting don't understand why I've got something in there, or a comment, or whatever, it becomes difficult for them to convincingly present from a position of understanding.

Andy Teece: So that cheat sheet is the extra bit that's going to help them cram before the exam, and step in and provide a really good presentation to the clients that's going to be of great value. Then, with my little tweak at the end, we're providing really quality meeting summaries. The next step is to update the client profile, all integrated. That's where we're going.

Trent McLaren: Yeah, awesome. That's great, Andy. I love that whole well-rounded view. Thank you for sharing all that. I can see Jack's comments on that as well. You can hear the level of work you're putting in, and you can hear how it's working for the rest of the team, so I love all that. We are coming into closing moments, but Gus, James, I want to give you guys one last piece of the conversation before we wrap. Gus, do you have any final thoughts? If there's anyone here that needs further help or wants to explore further, anything you'd recommend they do? And if they want to learn more about what you're building with AI Dojo?

Gus McLennan: Oh, listen, I think there's been some pretty impressive stuff here, and the people that you've assembled are probably at the forefront. I loved what Andy was saying about how they've gone and built out all these skills. From our perspective, we position ourselves as the platform for those accountants that probably don't have either the capacity or maybe the capability to get in there and do this.

Gus McLennan: So we've got our own skills centre: you can go in there, have a look at a whole bunch of pre-built skills, download them, make your own updates to them and start to improve your workflows. I think that's where I'd start. I saw some of the stuff Reuben had built, and that looks super impressive. It looks quite similar to some stuff we're working on ourselves, which hopefully means directionally we're in the right place as well.

Gus McLennan: But I think there's that evolution from building out some skills for highly repetitive, linear workflows, and then probably moving to the next level, into real custom AI workflows that are more software-driven, for really high-value processes like doing a company tax return, the year-end process, something like that.

Trent McLaren: Yeah, awesome. Thanks, Gus, appreciate it. James, any final comments or thoughts from you, mate? Any best practices or best places this group can start to get a leg up on AI, or even in your problem space around chasing clients, gathering docs, and so forth?

James Rose: I'm not allowed to say it, Trent. That's the one rule.

Trent McLaren: You can do.

James Rose: Yeah, look, it is just start connecting things. No, I didn't say "just ask Claude". It's connecting things, you know? I was just saying offline before this chat how I had not connected it to our Slack until last week, because I didn't feel comfortable with it. But now I'm starting to see that it is accurate enough for me to start doing some drafting or whatever, and already I'm like, oh my god, this is saving me so much time, right? And then I connected it to my signing tool just today, so I was like, can you just take what we were talking about here, grab that document out of Slack, stick it in there and prepare it for signing with all the boxes?

James Rose: And this stuff is only possible once you've got it connected to things. And this is irrespective of the chat that's been going on here, that I've been trying to snatch responses out of, around data security and stuff. I expect, and I've seen some apps already, that there'll be apps that sit in the middle of your AI and all the apps, so that Claude hits that app, and then the app goes and gets the data it needs from all the other apps and presents that back to Claude. There's no reason that app can't also be responsible for the redaction. And I do wonder if we'll see that kind of thing. Maybe there's a founder of one of those listening to this who's now gonna go and build those features.

Trent McLaren: I think we're more likely gonna see a localised, like Australian sovereign AI. I think that's maybe more likely, and might be an easier step for some people, rather than using US-based models and things.

James Rose: Yeah, but people are still real scared of putting stuff into AI, because I think people just do not trust that the AI companies do what they say they're gonna do, and not train on your data, right? I don't know if you saw that thing with the mathematical proof that was recently released, where the real human researchers were closing in on it, they'd been working on it for 10 years, and then just before they released it, OpenAI's like, hey, we solved it!

James Rose: And everyone's like, hmm… So those kinds of things don't bode well. I don't think people trust that they're doing what they say they're gonna do. Some people use Azure-based environments, Microsoft, with enterprise agreements. If there's anyone to trust, it's probably that. They would get dragged over the friggin' coals. I know they'll all be in hot water, but I just think the amount of risk on them, if it was found out they weren't honouring those agreements… I don't know. It is the Wild West, though.

Trent McLaren: It's an interesting space. Look, none of us are here to decide what's gonna be right or wrong, or how it's going to go, but it is fast moving. There's plenty of really good questions, and unfortunately, we don't have the right or wrong answer for you all. We can only share different points of view and perspectives from people at firms of all sizes. So jump on that React button for me. Audience, please thank our panel, you've all been amazing. I appreciate all the conversations, the diverse discussions, and the chat that's gone sideways at least 5 times in different ways, which is always a fun part of it. Thank you so much for coming along.

Trent McLaren: I hope you've been able to get some good discussions out of that. And keep having the conversations. This is not the end of this chat. Let's all keep figuring out how we help each other through those points. Yeah, I appreciate that, Tanya. Like I said, I think it's up to you. You're gonna have to make that choice on what you're comfortable with. I process refunds through Stripe, through Claude, if that helps you understand the scale of where I'm at, alright? I'm giving Claude everything. Make refunds, do upgrades, run my life, and come back and tell me when you get stuck. That's where I'm at. So I'm on the most opposite end of the scale to most people, is what I would say.

Andy Teece: Or just usher it through, then?

Trent McLaren: That's what I mean. Like, Trent, what do you think I should do? And I'm like, what do you think we should do? You're running the show here, just go. Anyway, we're gonna wrap here. Thank you all. Bye-bye.