Trevor Nash

Take us back. What did your week look like before you brought AI into it?

Most of my week was reading. A client would call with a question about an S-corp reorganization or a multi-state move, and I would open three or four research services, pull the code sections and rulings, and build an answer from scratch. A single thorny question could take half a day before I wrote a word the client would see. Busy season made that worse: the research backlog and the return volume competed for the same hours, and advisory conversations, the part I actually value, got squeezed to the edges. I was accurate, but slow, and I carried a lot of it in my head. When a similar question came up months later, I often rebuilt the analysis rather than trusting my memory of it.

What was the first thing you tried, and how did it go?

I started with the obvious experiment. I asked a general model a real tax question and checked its work. The prose was confident and readable, and the reasoning was often close. The citations were the problem. It would name a code section that did not say what it claimed, or invent a revenue ruling that did not exist. That was an early, useful lesson: I could not treat the tool as an authority.

What changed my mind was grounding it in my own material instead. My memos, the annotated code, the guidance I already trusted. When the model draws from references I control and quotes them back, I can check every line against the source in minutes. The judgment stays mine, but the blank-page part of research mostly disappears.

So concretely, how does it fit your routine now?

Three ways, mostly. For research, I feed it the client facts and my own reference set, and it returns a first-pass memo with the relevant sections pulled and quoted. I read it against the sources, correct it, and add the judgment calls. For planning, I use it to model scenarios: how an owner's salary and distribution mix changes the tax, what a capital-gain event or a move to another state does to the projection. It lays out the trade-offs quickly, so I can walk a client through options in one meeting instead of three. Then there is client communication. I take a finished, correct analysis and ask it to translate the technical version into a plain letter the owner can act on. I edit that for tone and accuracy, but the first draft saves real time.

Where do you keep the tool on a short leash?

Everywhere a number, a citation, or a filing is involved. Under my professional duties, due diligence is mine and cannot be handed to software. So every figure, every code reference, every calculation gets verified against the primary source before it reaches a return or a client, and I sign nothing I have not checked myself. The IRS has been explicit that AI is a starting point, not a finished product, and that a lack of technical competence with these tools is itself a risk. I take that seriously. Confidentiality is the other line. I am careful about what client data goes into any tool and which tools handle it, because I owe those clients the same protection I always have. The model can draft and organize. It does not decide the tax position, and it does not get to be sloppy on my signature.

Was there a moment it clearly earned its place?

A manufacturing client called in the fall, weighing whether to make a large equipment purchase before year-end or wait. The owner wanted to understand the trade-off in real numbers, not in a memo he would file away. I pulled the projection and used the model to run several versions: buy now, buy in January, different financing, the effect on the following year. In the past that is an evening of spreadsheet work. This time I had a clean comparison inside an hour, checked the depreciation treatment myself, and walked him through it that afternoon. He made a confident decision the same day. What struck me was not the speed alone. It was that the client got to have the strategic conversation while the question was still live, instead of after the window had closed. That is the part I had been losing to backlog.

What would you say to a CPA who is skeptical?

I understand the skepticism, and I share the version of it that matters. If a peer is worried about accuracy, they are right to be, and the answer is not to trust the tool more, it is to verify everything and keep the judgment where it belongs. What I would push back on is the idea that the choice is adopt blindly or refuse entirely. Start narrow. Use it on a low-stakes drafting task, ground it in your own references so it stops inventing citations, and check its output the way you would check a new associate's. You will quickly see where it helps and where it cannot be trusted. The competence the rules ask for cuts both ways: you have to understand the tool well enough to know when it is wrong. Once you do, it becomes a genuinely useful first-pass drafter.

How has it changed the work, on balance?

The clearest change is where my hours go. First-pass research that used to take hours now takes minutes, and that time did not vanish, it moved into advisory conversations, the reason clients hire me in the first place. I got through last busy season without adding staff, because the tool absorbed the drafting load that would otherwise have meant another hire or a lot of overtime. For clients, the difference is responsiveness. They can get a considered answer to a real question quickly, often while the decision still matters, and the explanation they receive is written for them rather than for another accountant. None of that changed what I am responsible for. I still own every position and sign every return. The work is just weighted more toward the judgment and the relationship, and less toward the mechanical first draft.