Two CPAs Try to Make Sense of AI in Practice

Charlie (00:00): I I really do think that value is gonna have to be redefined, basically, in our profession of, like, what does that mean? Because the client has this idea of value being my tax return, my financials every month, you know, transactions categorized, etcetera. And now in their head, they’re thinking that, oh, it’s taken, you know, the accountant less time to do this because of AI, so therefore, my price should be reduced.

Rebecca (00:29): Right. And I think people are a little afraid of the 7216 thing, not just because they’re worried that the client will be afraid of the security side of it, but they also are worried that they’re gonna start poking holes in the price because of that too. I mean, that I would worry about that. But I think that our clients there was always friction with our clients not wanting to pay the prices that we wanted to charge them because they didn’t value the boring stuff that just has to get done. That wasn’t very glamorous to them of getting the tech.

It’s just compliance. It’s it’s not giving them answers and solving problems for them. It’s just something that has to get done. But now, we actually have this opportunity to be like, hey, we’re actually gonna start providing you value on the things that you are really interested in and now there’s not this friction. Now you’re on the same page of, yes, these are the answers I’ve always been looking for.

Whatever you wanna charge me is great because now we can deliver the value that they want. Whereas before it was, yeah, I’d love to figure out how to do a cash flow analysis for you, but I’m too busy cleaning up your past. It’s taking it’s me way too long. I can’t look at the future with you. There’s only so many hours in the day.

And it and it is stepping out of our comfort zone because now we’re gonna have to learn how to deliver some new things that we’ve never delivered before, but we have AI to help us figure out how to do that and what it all means. We have this brain that we can use to help us fast track our our knowledge and understanding. I mean, I never was comfortable with forward looking financials. I just always felt more comfortable in the past. So I would read things online about how to do cash flow forecasting and it just didn’t click for me.

But if I could have a conversation with Claude to help me figure that out, I could deliver something to my client that is gonna be so valuable and so helpful to them and they’re not gonna question my price. They’re gonna be thrilled that they’re getting more now. And I think our clients of restaurant owner, sure they wanna leverage AI in their business and see if they can save money or improve efficiencies, whatever, but they still have their restaurant to run. So I don’t know that they’re gonna immediately run to say, I’m just gonna do it all myself now. They have a whole life and a whole business that they need to focus on.

So if we can swoop in and say we’re gonna continue to do the deliverables that we’ve always done, but we have some capacity now to go deeper with you and that’s gonna always justify our price and future price increases. Now we finally have an alignment of what they want, we can give them.

Charlie (03:02): Yes. There’s several things there and I think one of them is I feel like the primary complaint with accountants is that we’re reactive. We are looking backwards. And I think now with with AI, we really have the opportunity to switch more into that proactive role and that can provide additional levels of value that we haven’t been able to provide in the past. And then b is that because of our work, I mean, we deliver work to our clients.

Right? And so they see the deliverable. They don’t see the work that we did behind the scenes. Well, now we have AI that can really break down exactly what what we did, what it what we took to kinda get this output and kinda clearly explain that to the client because otherwise, I mean, who would have the time to sit there and say, I did x y and z. I had to clean up all this.

We made all these entries. Did all this work to get to the spot, to get to this deliverable. It just hasn’t existed at the time to really dig deep and explain the the output in that in that manner. So I think there’s an opportunity there and I think that again going back to the proactive nature of it. Say, okay, the deliverable is gonna be the same in a sense but we’re gonna layer x y and z on top of it.

And we’re actually gonna start looking at cash flow projections and we’re gonna start planning out the remainder of the year and what that looks like. And we can get really really granular with that and allow you to make decisions at your business level that otherwise in the past you would never have been able to do unless you would have paid for, you know, more of a CFO type role or controller type role and have someone come in and really be in the day to day

Rebecca (04:47): with it. Yeah. Or don’t give new deliverables, give the same deliverables, but they’re way better. You know, in ALFA, we learned about the ELI5, you know, explain it like I’m five. That could be a funny little paragraph that’s at the bottom of the financials every month that, you know, as an accountant, it’s a little hard to articulate how to word these things in a way that business owner will understand.

We try, but it it can especially with tax too, that’s tricky. But AI is great at that. And, you know, we could even ask our AI, can you not make this look like P&L? That’s not what the language my my business owner clients speak. They’re not impressed by a P&L.

Maybe some big corporations need them for their shareholders to review. But with smaller company, you could revolutionize the entire way you deliver the existing things that you already are delivering without having to add another service cash flow analysis And you’re blowing their minds and they’re just getting more out of the thing that you’ve always given them. One one thing is you could have your AI and look at the conversations that you’ve been having transcripts wise with your clients and ask it, you know, give me some ideas of how I could format this based on what they’re saying their pain points are. What would be meaningful for them to see? Give me some ideas, and it could do that.

I think Brooks Sutherland mentioned this. We can’t think about this as putting our legacy way of doing things into this new technology. This is our chance to do everything differently than we ever have and actually listen to our clients and what their needs are and what’s keeping them up at night. And before it just didn’t feel realistic that we’d be able to deliver those things. But now, anything is possible.

So we could just turn everything upside down on its head and still deliver the compliance things, still do a proper tax return and all that. But really help the clients make business decisions in a way that makes sense to them.

Charlie (06:44): Yeah. Think this kinda also goes back to you know, being AI native. Right? What does the AI native firm look like in this new age?

Rebecca (06:53): You tell me.

Charlie (06:54): Well, you know, I think it really is kind of that that shifting of the mindset to where we don’t try to fit AI into these legacy software or the way that we’ve always done things. We literally try to reimagine what this process would look like using AI at the forefront. Right? So not just kinda, you know, putting a band aid onto

Rebecca (07:17): example what a band aid would look with our, you know, think of your dad’s CPA firm or what you know, what what would what would a legacy firm that’s trying to just band aid on some some AI, what would that look and why would that not be good?

Charlie (07:31): Yeah. I think one example could be if you’re trying to do the books for a company. Right? And you have all their transactions, when you have their bank statements, you have their credit cards and you’re trying to figure out I have this software that I use to do the books whether it’s QuickBooks, whether it’s a different type of software and maybe they’re not hooked up and this is typical if you just you have bank statements that you have to get into a software. Well, mean you could use AI to parse all the transactions and get it into a format that whatever the software requires to in order to actually import the transactions.

And so that’s one way if you’re fitting AI into this process because you’re figuring out how to get from images and PDFs into maybe a CSV and then importing that into the software. Whereas if you’re thinking out about it from a more like AI native standpoint, you think about, okay, well what would this look like if it were done completely with AI? Right? And so in in the in the former example, again, it’s kind a band aid fix. Like, I made one part of this process easier on me But the other side of it being AI native is like, okay, well, let’s think about it, you know, larger picture of, okay, how does AI do this from ship to stern?

So and then at that point, you’re thinking about the goal. So what’s the goal? The goal is to get, you know, monthly financials and what does that entail? Starting with unadjusted, getting all the activity for the month or the year, making the journal entries, and getting to the adjusted trial balance. Okay?

Then then once you’re there, then you gotta package it up and deliver it in whatever way. Or if you start with the goal of, okay, I wanna send the client, you know, something they’ve never received before. Right? And that looks like a different template or a different presentation of their financials with that executive summary and you throw in charts and graphs and you throw in forward projections. Just a lot more detail, but also again have that, you know, explain like I’m five maybe on the front page of like, hey, here’s how the month went, here’s how the year went, here’s you know, five, you know, priority items that I want you to pay attention to.

Here’s some brief explainers about them. You know, you have this just this great presentation and then you gotta think, okay, this is where I wanna be. How do I go from raw PDFs and statements to that? Okay? And then what that looks like is gonna it’s gonna vary.

I mean, it really is gonna be dependent on the firm and the software that they’re using, the limitations of their software. But I mean, you can use Claude or you can use ChatGPT and you don’t even need, I mean, bookkeeping software.

Rebecca (10:13): I love that you’re saying zoom out and just look at what the goal is. What are we trying to accomplish here? Saying I’m trying to get the bank statements into a format that I can import into QuickBooks, that’s not the goal. That’s the step that you’re currently doing to try, you know, but zoom way out. No.

What does the client want? What does the IRS want? What really needs to happen? And how can AI support this? And sometimes QuickBooks might have a feature that keeps you in there for now, but sometimes not.

Sometimes you don’t need to be in there at all. And I think they’re going back to to the friction thing. I feel there’s always been friction too in, you know, the accounts that I would want on a P&L as their tax preparer and wanting to keep the books as tax ready as possible. And the accounts that they’d like to see on a P&L, you know, to help them really understand certain things about their business. And I remember those things annoying me.

That’s not correct. That’s not cost of goods sold. Why would that go there? I don’t care that you want it there. That’s not where it’s gonna go on the tax return and I don’t wanna have to adjust the books.

You’re a tiny company. But now, hey Claude, keep two sets of books for me. Always give me the client version and the tax version and every month I want both so that I internally can see what I need to prepare the sales tax return and, you know, do estimates each quarter or whatever. And they can see the things that they’re actually trying to know. And a lot of times clients are embarrassed to ask what the P&L means, so they just don’t look at it.

I mean, I think maybe 80 or 90% of my clients never looked at the P&L that I gave them every month, that they paid me to prepare for them every month. But if it actually said something on it that was useful, they would see the value in paying me and they would wanna stay with me and they wouldn’t care if I increase the price each year, whatever.

Charlie (12:04): Yeah. I think this also kinda goes back to how people are using these models today. And I think when we start talking about basically goal oriented action or like it drives the process when you start with the goal. You know, it’s difficult to achieve in maybe for just in the regular chat. Right?

I think it’s a great place to start. If you’re in just the chat interface of either Claude or ChatGPT and you can just basically ask it like, you know, I do multi financials for my client. This is how I’ve typically done it in the past. Here’s kind of high level of my process. You know, I’m starting to get some pushback from clients because with AI coming on the scene, it’s reducing my time and also therefore they’re requiring or asking for me to reduce their fees.

And that’s not the way that I wanna approach it. Right? So I want to provide additional value so that they are happy with the the services that I’m providing, the price that I’m, you know, charging them, help me understand what that could look like.

Rebecca (13:05): And you don’t have to type that. Right? You could just press a button and talk it.

Charlie (13:09): Yeah. I mean, ChatGPT has a great great voice mode. I mean, you could have this conversation when you’re on a walk, you know, on the weekend or whenever and whenever you have free time. It’s not something you have to sit there and labor over, you know, during the day. This could be any time because again, you’re thinking kinda high level.

Right? And at this point, you’re not building anything. Right? You’re not building a new work paper, you’re not building a new template. You’re just trying to get these thoughts out into a manner to where you can give it to the AI of again your your overarching goal or this issue that you’re running into.

And then after you get to a solid point to where you feel like, okay, you know, me and AI are kinda aligned with what we’re doing here. You know, what I do a lot of times is I’ll ask it to create a handoff package. Right? I’ll just say, hey, can you Based on this conversation that we’ve had and the conclusions that we’ve come to, can you create a hand off package so that I can give it to a coding agent so that we can start actually building this out and seeing what this would actually look like. And it’ll do that.

Right? It’ll create, you know, likely just a markdown file.

Rebecca (14:16): So give me an example of, you know, let’s say I had 12 bank statements and I’m trying to do, you know, put together a year’s financials for a client and I know what my chart of accounts should look like. I have prior year financials and I’m talking to chat about this bigger picture idea of normally I would, you know, get a QuickBooks ledger account or something to do this or you know, whatever. But I kinda wanna just see if if you can do this with me. And if you can, then maybe we could do this on a monthly basis for the client moving forward. Why would I need to then hand that off to Code?

What’s better about Claude Code or or Codex and ChatGPT? What’s that gonna do for

Charlie (14:56): I think the the primary difference is that these coding agents just have, you know, better tool calling abilities. And basically, when I mean tool calling, like, they’ll they’ll call Python. You know, Python just a coding language to where they’ll use Python to use Excel or use a different application or they’ll call the web search tool. So they’ll actually go and use, you know Google or another web search feature and actually go and look up something. And you know in the chat function you can do these things.

You can still it still has tool calling capabilities but it’s not as great as it is in these coding agents. These coding agents have a larger library of tools that they can call, they can use and they’re a bit more efficient and faster at using them. So if I was gonna do something like this, I would not do it in the chat feature. Right? I would migrate over to Cowork or Claude Code and I would approach it very similar to how I would in a chat because in the chat feature you can create a project, you can give an instructions, you can give a document and you can kind of work through it in there.

Well, I would just shift my thinking a little bit and go to Claude Code and I would create a folder on my desktop or on my computer. And wherever you know folder directory that that you want to work in or you can start on your desktop and then you can have Claude move it or you can have the AI just okay. Built this here in this wrong directory. I have a client directory where I want to put it and this is how I wanna break it down. And even if you don’t fully know like where to have it set up, you can just ask it.

Right? I would just say, hey, again, here’s my goal. I have twelve months of bank statements, I have twelve months of credit cards, I have a prior year trial balance. I have the chart of account. And what I wanna do is use you being the AI to help me prepare this year’s financials.

Again, based on kind of these reference documents that we have of the prior year trial balance. So Claude can have the beginning balances. He can add that and I would kind of walk it through like, you know, in the past, this is how my deliverables looked. It’s really just been, you know, the balance sheet, the P&L, you know, maybe a general ledger or whatever it may be. But I wanna think differently about it this year and I really want to deliver something that is above and beyond what the client’s used to and even what I’m used to and what does that look like.

And just have this conversation with it and ask it to basically set you up. So get you set up on your on your local computer. It’ll set up the file and folder structure, it’ll place the documents where they should go and then you can really just start working in it in the same type of manner that you do in chat. The difference being is that Claude now has a better tool calling capabilities that we talked about and it’s, you know, more efficient and faster in Claude Code. These coding agents are just, you know, more powerful than these LLMs that are in the in the chat

Rebecca (17:56): and Okay. Wait. Coding agent, you keep saying that. So I open up a chat in Claude Code. So it still kinda looks like chat, but I’m in the code part now.

Charlie (18:07): Mhmm.

Rebecca (18:08): Am I talking to an agent? Is that an agent or is what the conversation I have, it’s gonna go find agents to help it do what I ask?

Charlie (18:17): Yeah. So in Claude Code, you can choose the model that you’re using. Right? Something you can do in chat. You can pick these models.

And the difference is again is that you know, a coding agent which is what you’re using in Claude Code versus a large language model which is what you’re typically using in the chat. Now again, these large language models that are in the chat are getting more and more capabilities to where they’re starting to get I mean, they’re not on the same tier as a coding agent. But they’re again, they can do kind of, you know, basic things. They can look up, search the web. They can create a Google sheet or a Google document, they can create a PDF, they can create some of these things as well but when you go to a coding agent, I mean, you unlock a lot more capabilities.

I mean, you can have MCPs, you can have all kind of tools and integrations with with various different softwares and they’re just more efficient. They’re better, they’re faster and I would think about it like this. If you want to do a project, right, and a financial statement could be a project, a tax return could be a project, Then I would use Claude Code or Codex. If you’re just trying to have, you know, kinda like a a conversation or maybe something that is like not applicable to any one thing, you don’t wanna save it. The way I think about it is if I’m gonna have a throwaway conversation which means I’m not really gonna do anything with it, then I go to chat.

Now for a lot of people, if we go back to what we were just talking about starting in chat and then creating this handover, is perfectly fine as well because if people are more familiar with chat, then I don’t wanna say just do everything in Claude Code. Right? Because that’s probably not ideal. For me, I start in Claude Code. But I’ve been using it for several several months now.

Rebecca (20:03): Which kind of reminds me of like, why do we have to keep telling people just ask Claude or just ask ChatGPT? And I don’t mean that in a mean way. I just when there’s so many questions that we get asked that it’s so obvious, it clicks instantly to me that, oh, that’s a great question to ask. Why aren’t people all thinking that way? I know at the beginning, I didn’t think that way.

So how did that how did my brain make that switch?

Charlie (20:33): Yeah. I think it’s just counterintuitive. Right? This is a new technology, a new software. And prior to these large language models come on the scene, every other software, whenever you got a new software, you had to go somewhere else to learn how to use that software.

Right? So you would go to YouTube, watch a YouTube video on how to use it. You would go to Google. You know, you maybe would even read the instruction manual of like, how do I use this software? Well, now this is like the first software in history where we can just ask the software, how do I use the software?

And so I think that’s just kind of, again, it’s counterintuitive. Not many people have really grasped that. And I think for me, when I started, I had an engineer set me up on Claude Code and I remember literally sitting on the couch next to him and I was using it and I was like, so, like, how do I he and he was ask Claude. Just ask it. Just ask.

So I mean, again, it’s just not in our nature to say, well, I’ll just ask this thing that I’m using. How do I use it? But honestly, that is the best piece of advice that I would give any and I still give it to everyone that I talk to. I say, with all due respect, and I mean this, you know, in not a bad way, just ask it. You know?

I mean, I’m happy to help but I I try to instill in people to just ask the models. Ask whatever they’re in.

Rebecca (21:57): I remember we had a question in TCR recently that was, how should I use AI to help me with marketing my firm? And I just thought, well, why would you ask us? Ask the thing that you want to do that thing. They know better than we. Ask AI how you should use AI, you know.

And I know what they what they meant, how are other people doing it and you know, it was still a great conversation. But, yeah, that’s an example of just ask Claude, you know.

Charlie (22:25): Yeah. I think a great thing that people could do is that whenever they’re working in chat, I would ask whatever if it’s Claude or if it’s ChatGPT, say, I’ve been working in chat a lot, you know, and maybe they haven’t navigated over to to Claude Code or or Codex. I would just ask, okay, so, you know, why should I go to Claude Code or why should I be doing, what should I be doing in Claude Code that maybe I’m doing in here and what would the differences be? Explain it to me like I’m five. Right?

Just really break it down and maybe even walk me through what that would look like and how I should go about that and even create this handoff document, that I could give to Code to basically take whatever they’re working on currently in the chat and just start working on it in Claude Code. And then say, hey, you know, get me set up. Let’s set up a CLAUDE.md based on whatever we’re working on. Let’s transfer over any context that this chat has, any of the instructions and it’ll build the entire document for you. And you could just hand it right over to Claude Code and just continue.

And even once you get to that point and you’re using, you know, these coding agents and then I think the next level is using multiple different of these coding agents and comparing outputs and kind of pitting them against one another because I found that way you get very good output when they start reviewing each other’s work and you can kind of come to this consensus of like, you know, this model has kind of this approach, this model has this approach, Well, we kinda came to this conclusion and it’s kinda the best of both worlds.

Rebecca (24:05): Yeah. Yeah. And so to do that, let’s say I’m in Claude, would I just open another kinda conversation in Claude Code and just say, look at this or would I go to a whole different model or a whole different company or

Charlie (24:22): Yeah. I would use a completely different company, a completely different model from a different company. So like, you know, OpenAI and Anthropic. Right? They they all have their different model types.

I mean, you could use models within the same family, you know, there’s Opus 5, there’s Fable 5.1, there’s Sonnet. I don’t know if it’s entirely beneficial to run the same, basically task through the same family of models. I mean, you could try. I know certain of these models kinda are better. Maybe one’s better at writing, maybe one’s better at knowledge work.

But what I do personally is I work in ChatGPT and I work in Claude. And I’m right now using the best models of both, Fable 5.1 and Astra over at OpenAI. So I have, you know, them both up on my screen and I’ll run a task. Usually kind of the same task and I’ll just, right now, just copy and paste the output or be like, know, this is what another agent said, what’s your opinion on it? And then I’ll get that response and I’ll give it back to the other agent and say, hey, this is what another agent said about your work.

How do you feel about it, basically? And then just kinda go back and forth. And usually, I let whichever model or company kinda has the edge, I’ll let that be my driver. Like, right now, it’s it’s GPT Astra. It’s it’s working, in my opinion, you know, very, well.

And so I’ll kinda have it do something, take the output, I’ll give it a Fable 5.1, have it review, and I’ll come back to Astra and say, you know, here’s a review of the work you did. Don’t take it at face value. But is there anything, you know, that you should take into account for this review? You know, look at it with a with a clear head. And, you know, that’s worked a lot of times because a lot of times they’ll say, yeah, I actually agree with this and this but I’d push back on these points because of x y and z.

And then I’m kinda in the middle of like, okay, yeah, that makes sense or I don’t really agree with that either and let’s kinda go this route with it. And then I’ll just say, okay, well, let’s continue under, you know, whatever agreement or assumptions we decide to pursue.

Rebecca (26:26): Your judgment still matters. You’re Yeah. Kinda the boss and their employees sharing each of their opinions and then you have the final say on based on what you’ve both presented to me, here’s how I think we should proceed. And is this good to use for finding errors to check work?

Charlie (26:45): Yeah. Very very much so. I mean, that’s primarily how I use it. I mean, depending on the work you’re doing, you know, I’m building several applications. Right?

So I’ll have one of these models basically create a PR or make some code change and then I’ll just tell the other model go in and review that PR or go in and review this recent commit or these recent changes and basically give me your opinion on it. Know, what is there anything that looks off or should have been taken into account or just really give me any feedback on on this change and it’ll come back with its findings basically. And then I’ll hand it back to the original model and say, had another agent review your work and this is what it says. What do you have to say about this? And again, they’ll agree sometimes, it’ll push back on other points.

A lot of times it’s just that these models work differently. Like, right now, in my opinion, I think Fable 5.1 is slightly better at coding than than Astra is. But Astra is very good at, you know, creating it’s very good at visual design and it’s very good. I mean, he’s still very good at coding and it’s in my opinion a little bit better at writing than Fable is. I think right now knowledge work and financial like financial statement type accounting type work, think Fable is a little bit better just from my experience of what I’ve seen.

So I like seeing each other’s point of view whenever I’m doing the depending on type of work that I’m doing.

Rebecca (28:12): Whenever I hear people say that, oh, this one is better at that. I’m always like, wow, they’re so smart. How do they know that? But part of our job if we’re gonna start using AI in our firms is to do a lot of testing. Right?

Almost the way you experiment with different employees and you know, give them a task and see how they do and if they do great, maybe you keep giving them more work. So it’s not all just I use AI and everything just gets done with some things maybe, but it’s also we have a lot of human responsibility, especially with high stakes stuff, taxes and financials, where we have to be constantly testing them against each other, having them review each other, and then keeping up with all the news. Just we always keep up with all the IRS updates and changes. Now it’s all these models and things. So it’s not such a cakewalk, you know, like our clients might think.

It’s a lot of responsibility and a lot of upkeep. Right?

Charlie (29:08): Yeah. I think the the best way to think about these these new software is like it’s a new employee. Right? So like when you hire a new employee, there’s a lot of hand holding at first. There’s a lot of teaching.

There’s a lot of trial and error. There’s a lot of review. Right? Over time, you start giving them more responsibility. You start maybe not reviewing as much as you used to.

You start to learn the employees tendencies. Oh, they typically miss this. They typically don’t do this. Whatever it may be, it’s the exact same. And so when you’re working with AI and you’re kinda just getting started, you should be in the loop a lot in the beginning.

I mean, you should be constantly kinda reviewing and the key there is like learning. Learning how these agents work, what they do, what the tendencies are, how to use it well. And then as your knowledge of it and your experience with it starts to increase, you start removing yourself more and more from the loop but never completely. Like I disagree when when people you know want agents and they have to do everything. I just in this line of work that we do, don’t know how that’s possible.

I think they’ll get to a point where they’re very very good and we only have to come in the loop, you know, very seldomly but I don’t see a I don’t wanna say I don’t see a world but I think it’s gonna be I think we’re gonna be in the loop for a long time. I think it’s super important to be in the loop because again, yeah, it’s our name on the return. It’s our name on the header of the the financial statement package, you know, so it’s on us and we can’t just go back to the client and say, well, my agent made a mistake and I didn’t catch it because, you know, that’s that’s not gonna fly. And we have obligations and responsibilities that, you know, requires to really be in the loop. But I think the key is again learning through trial and error and through use of these models of these agents so that you can start removing yourself more and more and so you can hand off and have more responsibility to these agents.

And then really just come in when you need to and when it’s required and before anything goes out the door. And I think that’s that’s the way people should be thinking about it. I mean, that’s how I think about it and how, you know, I approach these models. But, you know, for people to expect them to work just right outside the box, I think that’s also a mistake. It’s just you have to get them set up right.

You have to learn how to use them. You have to figure out which model is best, what are the tendencies, and then really start to kinda refine your own process and your own way of working with these models. Because out the box, you’re gonna get bad output. And the more time you put in, you know, the better the output will be. And I think that’s what people should keep in mind when they’re using them.

Rebecca (31:45): So if I was brand new and I was just creating my ChatGPT account or my Claude account and I, you know, purchased the teams level, you know, because I own an accounting firm and I know that that’s supposed to be better for security. Or I already have it, but I as I’m listening to you talk, oh, I don’t know that I’m doing any of that stuff. What’s a good reset or a good just starting point if you’re brand new of what conversation should I be having with it to kind of reset and start doing my setup the right way. It’s not too late. I could still set things up differently.

Do I just tell it, ask me some interview questions about myself so that you get to know me and what I do so that you can help me get this set up? Because you know, mine agrees with me all the time and someone said, oh, you can tell it to not agree with you all the time or you know, because it’ll it’ll make a mistake and I’ll say, oh, that’s wrong. Oh, yeah. You’re right. It’s well, just stop agreeing with everything I say.

I want your honest opinion.

Charlie (32:44): Yeah. I mean, there’s a couple of things. Right? So in the settings, you use the desktop app or the or the web app, you can change the communication style. You can have it be concise, you can have it be more of a coaching style, you can make some broad adjustments there.

And then even within the web app and the desktop app, you can also give it information about you. But what I did at least when I started was I mean, first I would download Wispr Flow or fluid voice or some kind of dictation app to that Why? You just talk. Well, so you just it’s just much easier to just talk to it. At least I found than type.

And then even these models like ChatGPT and Claude, they’re the voice mode is getting really good too. So you can just have a conversation with it by talking to it. And it’s just faster. I mean we Yeah. Typically talk faster than we type.

And when you’re typing things, a lot of times you’re aware of what you’re typing so you may not type that thing because you just you’re just you’re aware that you’re typing. When we talk, you know, especially when we’re just kinda one on one, we’re just talking. It’s kinda first thing that comes to mind. So in my opinion, it’s like a bit more raw. It’s a bit more real.

And I think that’s super key when you’re working on these models and AI in general is just to be raw and real of like, hey, I just downloaded you. I’m an accountant. I do tax and accounting work. You know, I wanna use you to help me do the work that I do. Right?

I wanna know some of the things that I should be aware of because I don’t know what I don’t know. And also, what I think is super important to tell these models is that I’m not technical. I’m not an engineer. I’m non technical. You know, again, I’m an accountant.

I’m a CPA. So whenever you talk to me, explain things to me as a non technical accountant. And what I did early on was I told it to explain everything to me in accounting and tax analogies.

Rebecca (34:35): Can you give an example of how it did that? Because that that’s so clever. I would have never thought of that.

Charlie (34:40): Yeah. I mean, a good example and the reason I did this was because I wanted to learn how these models work. So I actually wanted to learn like how they work so that I could actually better use it. And I knew that if it explained it to me in something that I’m very familiar with, then I’d have a high likelihood of actually having it stick. But you know, it’ll it’ll explain things of a good one is like the file and folder architecture of like, hey, think of it as a, filing cabinet.

Like, it said that to me before when it talks about MD file or a context MD file or something like that. But it’ll just explain these analogies and a good one is just how to think about these agents in general. Like a new employee that you just hired, you have to give it a these SOPs that you’ve created that you hand to the agent. And so it’s just a great way of kinda making it stick and think that just goes a long way. And so I would tell it, you know, I’m non technical.

I’m an accountant. I’m CPA. Explain everything to me in that terminology. Don’t use technical terms. I don’t love explaining to me like I’m five because I just I just don’t like I don’t like saying that.

I’ll say explain it to me like I’m 15 or 18. I don’t want it dumbed down too much. Right? But I do want it so so I can consume whatever it’s telling me. But that’s where I went the route of like, explain it to me in tax and accounting knowledge because I can understand that really well and that that works very well for me.

So I would I would start there. I’d start in the chat and I would just have a conversation with it. Right?

Rebecca (36:07): I remember you did that with the GitHub in ALFA. None of us really knew what GitHub was and you wrote us up something that explained it all in the context of accounting. Yeah. Do you remember that?

Charlie (36:21): Yeah. Yeah.

Rebecca (36:21): Members being Now I understand. So that’s smart to something that you do know really well, ask it to connect everything to that to help it really sink in and really click.

Charlie (36:33): Yeah. And and by it having that context on you that you are an accountant or that you’re a CPA and you do this type of work, it has that in its knowledge and it has access to that. So like it basically keeps that in mind whenever it’s working with you. So if you ask it a question, like a good one is like I send it X posts or YouTube videos all the time and it just knows to apply it or how it could be applied to the work that I do because it knows the work that I do. Right?

And so

Rebecca (37:00): Give me an example of that. Think of one that you’ve done recently because I’m so curious. I

Charlie (37:06): did one. Let’s see. I did one this morning or last night. I saw a post on X on it was about I can’t remember what it was called, but it was basically a a new operating system or actually a new desktop. Right?

And so this I saw this you know post on X, about this startup releasing this new desktop and you it’s basically meant for like knowledge work. So you can like open up like excel document like in there and you can have you know, it’s really good for agents and you can have, you know, emails. It’s just a much better kind of desktop than kind of what exists today. I just send it a link. I don’t send it anything else.

And it’s like, you know, it will talk about like privacy concerns. It’ll go and look up who this is, who the founder is, you know. If it’s open source, it’ll go and look at the the GitHub repo to check it out. But, you know, it’s it’s typically like for someone in your profession, you know, there’s the privacy is this and, you know, the the way that you could use it could be, you know, having excel and pdfs and kind of one of you while you’re working with an agent and dah dah dah. And then I have it just store whatever I send it in a Notion database.

And so like I if I ever wanna revisit it, it’ll have the original link and it’s kinda like synopsis in there. And so if it’s high applicability potentially, then I just kinda have it stored and then maybe I go back and visit it. But at this point, there’s just hundreds of things in there that is just kinda building up but maybe one day I’ll I’ll start chipping away at them. But again, I think it’s key to give it the context of what’s the goal. Right?

So my goal when I sent you this tweet and I and I told her this, I’m like, I’m gonna send you tweets. I’m gonna send you YouTube videos and I want you to figure out what it is I’m sending you, how it could apply to me based on the work that I do, and any other synopsis do some brief, some quick research on what it is and kinda give me this TLDR of what it is and then store it accordingly.

Rebecca (39:04): That’s a huge time saver.

Charlie (39:06): Yeah. It’s the the big, probably the best way that I use these agents, or not not the best way, but just a very valuable way of just sending it something, explain it to me. Should I spend time figuring out what this is, learning more about it? Or is it just NA for me based on what I do? So it does save a lot of time.

Rebecca (39:27): And the TLDR thing too. I had never thought about that. And now in when I’m asking the question, I always say and give the response in TLDR. And then it’s so much less reading for me, you know.

Charlie (39:42): I tell it that all the time because I’m a visual learner and so working with AI has has kind of been originally before I really kind of finessed how how the output is, you know, it just it’s just walls of text and it’s just like I struggle to read like a huge wall of text, like I need to see it. And so, you know, that turned into me telling it to create HTML files, right, and visuals and give me TLDR and just really break it down. And now, like, don’t even have to tell that to the models anymore or to the AI that I’m working in because it has that context on me. It knows how I work and knows the preferred output that I that I want and just gives it to me in that way which is super helpful. And you know, it’s it’s still sometimes with the walls of text, it doesn’t do perfect every time but I think that’s another thing that people get tripped up on is that they, for some reason, expect these models to be perfect and they’re not.

I mean, they they make mistakes. They do things wrong all the time. I mean, the key is catching it when they do, telling it when it does. Again, going back to like this It’s an employee. Yeah.

This is just an employee. Hey, you messed up here or like, hey, I don’t like that output. This is what I prefer. Can you make it like this next time? And you know, these models are great.

They’ll say that to their to their memory and they’ll update any context files or instruction files and the next one ideally will be better. I think that’s just how people should treat them. You know, they’re not perfect. I’m not perfect. No one’s perfect.

And that’s the same goes for these models.

Rebecca (41:09): Yeah. That makes sense. This has been amazing. I I mean, I have so many more questions and now I’m thinking a lot about our employees. First, we learn this and then we have to tell our team, human team, how to do all this too.

We can talk about this next time, but I just always remember at the big firm I worked at that I would be so scared to ask one of the senior people or managers a question, a tax question because they I knew they were gonna say, did you look it up in the book first? You have to go, you know, and I’m like, this big book doesn’t have the context of exactly the situation though. Like, it’s so inefficient for me to sit here when I know you know the answer. I just and I’ll learn from it, but they were so annoyed to be bothered. And so it was this fear thing.

I was so afraid to ask them a question, and I had to really exhaust my resources before I went to them. And it was part of our review of how often you were asking them questions. It was a bad bad thing to do. So I’m just a man. If I had this, I knew how to ask the questions with the proper guardrail so that receiving answers from credible sources and stuff.

If I knew how to use it properly as a staff accountant, I would fast track my career. I’d be a senior in no time. I’d be a rock star.

Charlie (42:24): Yeah. What’s interesting is a lot of firm owners that I’ve talked to, they’re the primary ones using AI. And I really think that’s kind of backwards. You know, I don’t have any employees. But if I did have employees, I would want my employees to be just killers with AI.

I would want them to be using it heavily. Now you gotta have a conversation, right? What does that look like? How are we using AI? You gotta update your AI policies, your WISP, you know, everything like that but that takes no time.

Like that’s not an excuse. Just sit down, write it out, use AI to help you write it out and get that over with. And then it’s like, okay, you know, give your employees the freedom within reason to learn how to use AI and these models to streamline various aspects of the firm because they’re the ones that are in it day to day. I mean, they know probably, you know, to a better degree the bottlenecks, where the bottlenecks are because they’re living in it every day. Whereas if you’re a firm owner, you know, you’ve removed yourself from from a lot of that rightfully so.

And so your bottlenecks are much different than theirs. But they’re spending, again, probably a lot more time on things that probably have more potential to be automated and more potential to kinda implement AI into the pieces that they’re working in that maybe the firm owners does. Because the firm owners, you know, likely spend their time on really high value judgment heavy work where, you know, you can’t just hand that off to AI.

Rebecca (43:49): Yeah. And I think everyone is a little shy at the beginning and afraid to say what they did with it or even think of what they could do with it. But I’ve seen it, you know, talking to firm owners, where it’s the domino effect. As soon as one employee finds something that saved a ton of time on this monthly task that they always have to do and they get excited and they share it with everyone and then they’re sort of praised for doing that and look at all the time you saved the company and you’re awesome. Then the next person is kinda motivated to, well, I wanna find a way to say it.

So you could make it fun, make it contest, let me, you know, make it, you know, let’s all see what we can do. I I know that there’s also balancing, you know, the the anxiety that your team might have about, does this mean I’m not gonna be, you know, needed anymore? I know of some of firms where the employees are resisting using it because they’re afraid that it will kinda uncover, you know, the reality that AI could just replace them. So I don’t really know, you know, you can tell your employees as much as you want. I promise, you know, this isn’t to replace you but I don’t know how you make them more comfortable with that.

Charlie (45:01): Well, I think I mean, in my opinion, it’s a bit overblown because it’s extremely difficult to steer these models. Right? If you are responsible for setting up a model to achieve a goal or to produce an output and you’re worried about, okay, now it’s running nice and it’s kinda producing this output. I mean, you still have to maintain it and so is the firm owner gonna let this employee go and they’re gonna switch to maintaining, figuring out how or what this model does to produce this output. You know, I don’t think that’s the way that that we should be looking at it.

Know, again, I think we should be given we should be empowering our employees to say, look, optimize and automate this part of your workflow and then let’s go tackle the next. Or once you get that dialed, now you can spend more time of these other opportunities that we want to pursue in the firm that we haven’t had the time to because we spent all our time doing this. And now we took this from, you know, a four hour job down to an hour or even less. Now we have all this time freed up to where we wanted to kinda offer this new this service line. Well, let’s start again talking with AI and figuring out what that would look like because now we have the time.

And so I think that’s the better route that we should be taking as firm owners is empowering the employees, you know. Let them know the reality like these models are very good. But I think the firm owner should be honest and say, I’m not planning to to replace you. Or if they are planning to say, you know, just be honest. Let them know that.

Yeah. It’s it’s their business ultimately. But I think the key is honesty, getting on the same page, and then working towards the same goal.

Rebecca (46:42): And I think each employee basically just got a huge promotion. Now you have this second brain that you can use to be at a more senior level than you could have been before. You can you could take you know, one of your bookkeeping clients and and tell your, you know, AI all of the problems that that that all the inefficiencies in the in the operations of that business. And AI could come up with some amazing tools and ideas and service add ons that you could then present to the client and charge them more money to give them solve more problems within their business. Now you’re doing business consulting using tools and implementing those.

Now you’re doing a whole different thing and staff accountant could probably do that, you know, because now they have this amazing tool to kind of fast track their career.

Charlie (47:33): Yeah. If I was an employee, then I would want to become the AI guy or the AI girl. Right? I would want to whenever there’s a problem, I would want to be the one that figures out the solution by using AI. Cause if you become that person, then every time there’s a problem, you know, management or the firm owner is gonna say, well, you know, where’s Charlie at?

Like, I bet he can figure this out or whatever it may be and you know, problems are just gonna continue to exist. I think new problems are gonna come now because of AI. Now we’re just gonna have different problems that we’re having to deal with. I mean, management. What does it mean to hand off much more responsibility to an agent, to AI?

Where do we now put ourselves into the loop or where do we remove ourselves from the loop and then what happens in both of those scenarios? Like, if we’re if we’re removing ourselves from a loop or from a process, then what does that mean? What does that look like? What happens downstream from that? If we put ourselves into a new loop that we’re not used to being in, what does that look like?

What does that mean? Yeah. And so there’s just gonna be new problems for us to solve, new value for us to add that I think if you can be the one that can be this problem solver or this value adder and use AI to leverage that, then I think that’s a huge huge desirable trait that if I was hiring someone, I would want an AI native person for sure, a 100.

Rebecca (49:03): Which was kind of saying that about a business hiring an accountant. You would want them to be able to add all this value and solve all these problems. Same thing with an employee. There’s kind of a a similarity there.

Charlie (49:15): Yep. I think this is only gonna get more and more important too as AI adoption continues to increase on both sides. Right? Our clients are using AI more and more. We are using AI more and more.

That’s gonna present new opportunities and new problems. So I think that’s where, you know, we really need to start thinking deeply about what does that look like and what does that mean for us and how can we position ourselves today to be ready for this, I mean, pretty unclear future. But I think that, you know, the one thing that is clear is that things are changing. And so we need to, start preparing for it today so that whatever that change looks like, at least we’re in a better position to kinda handle it and even adapt our firm and our works so that we can kinda fit this new age. And it’s gonna look differently for everybody.

You know, it really is depending on if you’re a solo practitioner, if you got a team of 20 or if you got team of a 100, you know, that’s gonna look vastly different for everyone which I think is another opportunity and a great chance to see kinda how others are doing it, approaches that people have taken that maybe didn’t work, why it didn’t work, or if something did work, why did it work, and just kinda learning from kinda what’s going on. I think that’s what I like the most about what we’re doing at AI Labs is, you know, we’re just kinda sharing. Right?

Rebecca (50:40): We gotta talk about this. We can’t all be in a silo on this. We gotta help each other. Our profession needs collaboration on this. We gotta figure this out together.

Charlie (50:51): Yep. Yeah. Alright. We gotta wrap this up.

Rebecca (50:53): Yeah. I mean, we could go on and on, but this is this is great. We’ll we’ll do this again soon.