I get asked to talk about AI a lot, and most of what's said about it in my industry is either fear or brochure. I'd rather tell you what actually happened when we put it inside a working tax, accounting and hospitality operation — including the parts I had to undo.
Some background, because it explains the bias. Before I was a tax strategist I spent six years as a Director of Information Technology, and years before that as an assistant controller who also ran IT because the property wasn't big enough for two people. I've been on the systems side and the numbers side. So when AI arrived I didn't experience it as a new thing to be afraid of. I experienced it as tooling, and tooling either survives contact with real work or it doesn't.
We've shipped four platforms. The first customer was always us, with our own money on the line, which is a very effective way to find out whether software is any good.
What held up
Reading unstructured documents. This is the boring win and it's enormous. A client sends a shoebox — statements, notices, closing documents, a photo of a receipt taken at an angle. Getting that into structured data used to be the whole job. It's now mostly solved, and the time saved goes back into thinking about the return rather than typing it.
First-pass review. Not the final review. A first pass that flags the things a human should look at — the number that moved 400%, the classification that doesn't match last year, the missing form. It doesn't decide anything. It shortens the list.
Drafting the explanation. Most of what a good advisor produces is explanation, and explanation is expensive in time. Drafting it and then editing hard is faster than writing it cold, and the client gets more communication rather than less.
Hospitality reporting. A hotel P&L behaves unlike anything else, and the variance commentary that owners actually want was always the thing that slipped when the month got tight. It doesn't slip now.
What broke
Anything where being confidently wrong is expensive. The failure mode isn't that it can't do the work. It's that it produces something plausible, formatted well, and wrong in a way that takes real expertise to spot. That is worse than a blank page, because a blank page doesn't fool anybody.
Citing authority. Early on we let it draft positions with citations. That went badly enough, fast enough, that it is now a hard rule: every citation traces to a written authority a human has opened. No exceptions, no "it's probably right."
Judgement calls with no clean answer. Reasonable compensation. Whether a position is defensible for a specific client with a specific risk tolerance. These aren't knowledge problems, they're judgement problems, and judgement is the thing you're actually paying a professional for.
What I rolled back
We tried putting AI in front of clients for intake — letting it ask the questions a preparer would ask. It worked technically. We killed it anyway.
The reason is that the intake conversation is where you find out what's really going on. A client mentions in passing that they're thinking about selling in a few years, and that one sentence changes everything you'd advise. A model asks the questions on the list. A human hears the thing that wasn't on the list.
The interview isn't data collection with extra steps. It's where the work actually starts.
The boundary I'd draw
The honest line, and I'd defend it on any stage: AI is very good at gathering, structuring, drafting and flagging. It is not good at deciding, judging or being accountable. A professional's value was never data entry. It was judgement and responsibility, and both are still ours.
What changes is the ratio. If the mechanical work shrinks, the same team can do more of the thinking — which is the part clients were underserved on in the first place.
What I still haven't figured out
How much of the review layer this can genuinely carry before a human has to own it again. I don't have a clean answer. I have a conservative line that I'm fairly sure is too conservative, and I move it when something earns the move.
If you want the version of this with the demos and the arguing, that's most of what I do on a stage now. It's the talk I'll go longest on.
I'm Ryan Otto — an MSCTA and fractional CFO in Allen, Texas. I run Accent Financial Services, The Taxsmiths, Dayspring Hospitality Solutions, and the AI studio Dayspring IdeaForge. This is education, not advice about your situation — your entity, your state and your own facts all change the answer. Ask me directly if you want it applied to yours.