Over the past six months I've been on more than 80 calls with accountants about AI.
The same problems and discussions kept coming up, so I wrote down 30 things I've learned along the way. https://x.com/i/article/2107893806150619136
30 lessons from six months of talking to accountants about AI
Over the past six months I've been on more than 80 calls with accountants about AI, between the weekly AI Lab for Accountants call and one-on-ones with firm owners working through what they're building and doing with AI.
Here are 30 things I've learned along the way:
1. Implementing AI is a lot like onboarding a new employee
There's a lot of hand-holding at first, a lot of time spent getting them set up, a lot of teaching, and a lot of review. Over time you learn its tendencies and start removing yourself from the loop.
It's also nothing like an employee. A new hire remembers (for the most part) what you told them last week. An agent starts every session with only what's been written down for it, and if that's wrong, it acts on it like it's true. Also, it doesn't get tired, and you can run several at once.
You get better results once you understand both sides.
2. Start by writing down what you hate doing
"Where do I start?" is the question I get most. People come in with grandiose visions of replacing things or building things. I tell them to spend one week writing down what they hate doing or what eats their time, and to have their employees do the same.
Use AI for it. Create a project and say, I'm going to spend this week coming in here and telling you the things I do that I hate doing or that take a long time. At the end of the week, pick the smallest thing on the list and start working on it. See it all the way through before you start the next one.
3. Most "model problems" are setup problems
I did several consulting calls this summer where the person told me the AI kept forgetting things, editing the wrong file, and had to be told the same thing every time they open it. Every one of them assumed those were model problems. None of them were. It's easy to confuse the two and I'll discuss it more in number 12.
A lot of people use AI right out of the box, get bad results, and blame the AI. Setup and foundation are arguably the most critical parts of this. You have to get your harness right, which is everything set up around the model, and get the environment right, before the AI can really cruise.
4. The owner can't be the only one using it
A lot of the firm owners I've talked with were the ones experimenting with AI. That part matters. You should know whatever AI is being used in your firm and how it works.
But keep your employees in the loop, and build an environment where people talk about AI, share their learnings, and work on it together.
5. Your tech stack decides what AI can do for you
I talked to one firm whose practice management software had no way for an agent to connect to it, so everything they wanted to do with AI turned into a workaround. Another couldn't run the desktop AI apps at all inside the hosted environment their firm works in. Another was ready to leave their practice management software until it opened up an API, and now they might stay.
A lot of firm owners are limited by the stack they've already chosen. Personally, I wouldn't sign a three-year contract with anybody right now. Ask whether your agents can connect to it through an API or MCP, and ask how you'd get out if you needed to.
6. Go after the software you pay for and barely use
A firm owner told me he uses eight or ten tools pretty heavily, has a top five in his head, and is paying for a few he rarely touches. I'd target those first.
Ask your AI what it would look like if you didn't have to pay for it and just used AI to do that job instead? Could you build a lightweight version of it yourself? Then keep chipping away.
7. Skills need to be shared and kept up to date
A lot of the people I talk to rely heavily on skills they've created. Not many have a way of getting everyone in the firm access to them, or of knowing when to run them and how to update them.
One firm I talked to does two things I liked. Every skill they've tested has a short video showing how it works, and nobody builds a new one until the first one is rolled out.
Review your skills continuously, and especially around each new model release. You can even create a skill that does this for you.
8. Don't stop at the Band-Aid
Very few have actually implemented AI. Most are using it as a Band-Aid that helps a little in a lot of places. That's a fine way to start. You may need Band-Aids on the places that are bleeding just to free up some capacity. Fixing one annoying task at a time, like in number 2, is how you learn.
But don't stop there. Once the bleeding has stopped, work on preventing it altogether so the Band-Aid can come off. That means going much deeper. Stop fitting AI into the way you've always done things, and reimagine the process with AI at the forefront, from stem to stern.
9. Ask the software how to use the software
This is the first software in history where you can ask the software how to use it, why it's doing something a certain way, and how it could be done better. It's just not in our nature to do that.
When I started, an engineer set me up on Claude Code, and every time I turned to ask him how to do something he'd say "just ask Claude, just ask it." That's still the best piece of advice I've gotten.
10. Start with the goal, not the task
Tell it what you're trying to accomplish and work backwards. Have the model break the goal down into the steps, workflows, and tasks it takes to get there.
Don't accept that at face value either. Question its methodology. Then have it create a project map for the goal so you can see it, and tell it to track its progress. Zoom out and then zoom in.
11. Stale context is worse than no context
I ran an audit on my own setup and it found a file still telling Claude to save documents to a folder that had been moved and no longer existed. When context is missing, the model will often ask. When it's wrong, the model acts on it like it's true.
So have the AI create the right context files and keep them current. Ask it if there's a CLAUDE.md or AGENTS.md, which are the instruction files these agents read at the start of a session, and if not, should there be? Ask if there are instructions for what it's working on, what they say, and whether they're up to date. Ask if there's a README. Ask if there's a CLIENT.md when you're working on a client, a PROJECT.md when you're working on a project, and so on.
12. Pick one model and learn it
If you're getting started, stick with one, whether it's Claude, OpenAI, Copilot, Grok, whatever. Go deep with one provider before you split your time between several. Save the experimenting for when you're really dialed in.
One of the best things I did was commit to going deep with Claude in May of 2025. I really learned how the different Anthropic models worked. So when I started using several providers earlier this year, I was in a much better position to see how they were different and which provider/model was best for which job.
13. Get out of the chat
If you're still doing everything in the chat window, the next step is to get out of the chat and start working with coding agents via Claude Code/Codex/etc.
In the chat, you're the one carrying everything. You upload the file, explain the situation, copy the answer back out, and do it all again tomorrow. Claude Code/Codex work where your work lives. It opens the folder, reads the files that are already there, makes the changes itself, and can run things to check its own work. It reads your instruction files every time it starts, so you're not re-explaining your firm in every conversation.
That's the difference between getting help with a task and handing the task off. And don't let the name throw you. You don't have to write code to use it.
14. Tell it what finished looks like
"Get this ready for review" means something specific in your firm, and the model has no way of knowing what unless you say so.
Think about handing that assignment to someone new. They'd need to know how you handled things last month, which accounts need attention, what they can resolve themselves, and what needs to come back to you. The agent needs all of that too. Without it, you still get something that looks finished, and you won't find out it wasn't until you're reviewing it.
15. Don't build a rigid skill for work that has exceptions
Skills are great when the steps barely change. Roll the dates forward, reformat the export, same shape in and same shape out. Build those all day.
The problem is the year something is different. A rigid skill follows the steps and might not tell you they stopped fitting. The output looks right because the process ran, and the exception ends up buried in a workbook.
The instinct is to add a rule, then an exception, then an exception to the exception, until the skill has boxed the model out of what it does best, which is reason. A skill doesn't have to be a list of steps. For work with judgment in it, give it the goal, what the output should look like, and what you'd want flagged, and leave the path open.
16. Stay in the loop until the evidence says you can step back
Don't wait until the end to review. By then you have little to no context, and by the time you've figured out what the agent did and why, you'd have been faster doing it yourself the old way. That's the catch-up cost: the time, energy, and effort a human spends figuring out what an agent just did before they can trust any of it.
What's fixed this for me is breaking the work into its natural steps and putting myself in the loop between them. More checkpoints, yes, but less time at each one.
And stay there for a while. You remove yourself from a step once you've seen enough evidence to warrant it, the same way you would with an employee.
17. Have a fresh agent review the work
The agent that built it knows the whole story: what you were building, why certain decisions were made, and what you planned to fix later. That's invaluable when building and a liability when reviewing, because it fills in missing intent and treats unfinished pieces as understood.
A fresh agent has none of that backstory, so it can only evaluate what's actually there. Give it the requirements and the source documents, and leave out the builder's explanations. Build with context, review in a fresh context, then bring the findings back to fix them.
18. Think of these agents as lazy geniuses
They know an absurd amount, and they'll often do the minimum amount of work necessary to produce an answer. If retrieval isn't part of the workflow, you can't assume the model actually checked anything. So be explicit: open the source, read the entire section, and make sure everything ties.
Where that language lives matters. If you only type it into the chat, a new session starts without it. Put it in the files the agent loads at the start of every session, like your project instructions or CLAUDE.md, and ask it what it loaded. A README only helps if something tells the agent to read it.
19. You need a way to undo it
Before you delegate anything, the agent has to know where it is and what it must never touch, and you have to be able to undo what it did. That last one is the one almost nobody has.
A project with no version control is not a project you can delegate. Everything else improves your odds of a good outcome. This is the only one that makes a bad outcome survivable.
20. If your clients are using AI, don't fight it
Help them use it better. Treat it as a collaboration, where both of you are using AI to serve them. If you aim to be AI-native, that can't be one-sided.
For me, this looks like getting a client's harness and environment set up so we can both work with AI together. Figure out what that looks like for your firm. It might even be a new tier of service with an AI collaboration piece built in.
21. Ask your clients what they wish they could see
I took a new client to lunch and just asked questions. What do you wish you could see? What's important to you? How do things work in your business? He told me he'd love to see his cash position across all of his businesses in one place. I said, cool, I'll just build it for you.
I built it over the weekend and emailed him Monday. He loved it, and asked if I could build a cash projection for the rest of the year. I told him I'd have to charge more for this, and he didn't blink.
You don't have to roll something like this out to everyone. Try it with one client who's willing, and ask for raw feedback.
22. If you build your own software, you own it afterward
I built my own client portal and quit paying for the old one. Hosting, a secure database, and backups came with that decision.
Start with the goal in mind. If you plan on hosting it, or giving your employees or clients access, the agent needs to know that when it starts building, because it's very hard to retrofit into an app you've already built. Tell it up front that security is paramount, and so are data storage, maintenance, and monitoring. And start small, which goes back to number 2.
23. Think about the future version of your firm
A lot of people are focused on the here and now with AI. I think we should be working out what we want our practice to look like a year from now and three years from now, and working backwards from there.
A firm that runs heavily on AI and agents won't happen overnight, and you'll likely have to rip out and rethink a lot of existing processes. Get aligned with your agent on where you want to go, build that roadmap, and start working towards it.
24. Put your 1040 clients in buckets
If I had a 1040 shop with a thousand returns, I'd be putting them in buckets. Who's going to stay with me for the next five years no matter how good AI gets? Who could go either way? And who could I see leaving, the ones I talk to once a year with a pretty straightforward situation?
Then I'd quantify it. If most of the first bucket stays, half of the middle one stays, and most of the last one leaves, what does that do to revenue, and how do I make up the difference?
I don't think any of us should be losing sleep over this. I do think we should be planning for it. These models can already do taxes and accounting, whether we want to believe it or not, any they're getting better and better every week.
25. The ones furthest along didn't wait for permission
The ones that are furthest along with AI are those that didn't wait. Nobody told them what to do or how to do it. They picked something, got going, and figured it out along the way.
This goes back to something I believe pretty strongly: In the age of agents, you need agency. Go do it for yourself, and don't let anybody tell you it can't be done.
26. Micromanaging is a stage, and you're supposed to leave it
A CPA about a year into his firm told me AI speeds him up some, but it mostly feels like he's micromanaging Claude instead of delegating to it. I think that describes almost everyone early on.
Getting from there to delegating takes structure. Write instructions that match how you think, get one process right from start to finish until it's repeatable, then do the next. You earn delegation one workflow at a time.
27. Spend more time thinking with these models
These models are good. They have access to a lot of information, and they're great at reasoning. You should spend more time thinking with them and treating them as a thought partner. Ask why it's doing something a certain way and how it could be done better, and go back and forth.
On a lot of my calls I have the person share their screen, and then I talk to their AI myself. One firm owner I've been working with told me that what he learned in our first session, just from watching how I prompt it and how I talk to it, was priceless. That alone has been a big unlock for people.
28. Ask it how to make this 10x better
One of my favorite things to ask is how do we make this 10x better, whatever it is that I'm working on. It works because the model already has the context: what the thing is for, how it's set up, and where you feel it's falling short. It's pretty good at finding the gap between where it is and where you want it to be, and coming up with ways to close it.
I think about my own work the same way. The first version of anything I build for a client is V0 and the worst it will ever be.
29. Rethink what's possible
Forget how things have always been done. AI gives us a huge opportunity to change processes and workflows we've run the same way for years, and to add our own flair to them while still getting to the same outcome.
I started out working for my dad. He had done things the same way for forty years, and what I heard the most was "same as prior year." I hated that, because there's no creativity in it, and this year is never the same as prior year anyway.
Now you have not only the freedom but the capability to make your firm into whatever you want it to be. So get creative with it. Pick one thing you do the way you were taught to do it, and put your own flair on it.
30. You're not behind
Most of the people I've met with feel behind. For a sense of scale, only about 2% of U.S. households were paying for an AI subscription as of early 2026. If you're paying for one of these tools and using it in your work, you're in a smaller group than you think.
It's not too late to start and you don't have to do it alone.
Where does this leave us?
None of this is finished. I'm still figuring a lot of it out myself, and I'll be wrong about some of it. But I don't think anyone is going to hand us the answer, so I'd rather work it out in the open with other accountants.
That's what we do every week in AI Lab for Accountants. If you want to work through this with us, come join: https://ailabforaccountants.com.