I run a tax practice alone, and I have two children under six. Most of my day belongs to them, so the hours I have for client work are few and scattered. For years the limit on the practice was not demand. It was that I could not promise a new client I would have the bandwidth to service them properly, so I said no.
Agentic tools changed that, but not in the way the pitch decks suggest, and not everywhere. Here is the honest accounting.
What actually runs
The list is shorter than people expect. Routines watch for tax regulation changes, including a few niche areas I do not work in often enough to track by habit, and surface anything that moves. On the writing side, agents do every step up to the words: they read what I have posted before and how it landed, find conversations worth joining, and research the angles I could take.
Beyond that, most of my tax work still happens the old way. I have tested the full workflow end to end on spoofed data, and the plan is to run it for real starting with tax year 2026, on an account whose terms meet what this profession expects around client data. That is a date on a calendar, not an accomplishment.
The thing I tried that came back unusable
When Opus 4.6 arrived it was the first model where offloading real work looked possible. I had spent years trying to solve the same problem with remote staff and with local staff, and nothing had worked to my satisfaction. So I handed over the largest thing I had: prepare the workpapers that support a return, and make them good enough to defend it years later under examination.
It came back unusable. None of the worksheets tied. The failure was mine more than the model’s, because I had asked for a whole role instead of a task. What worked was smaller and checkable: pull the historical basis information I need to compute a gain, look up an ambiguous position I am weighing and give me the authority behind it, help me think through the internal controls in my own practice and where the weak points are. Every one of those has an output I can verify in minutes.
What I do not hand over
Client data does not go to agents. That is not a preference, it is the floor this profession sets for itself, and it is why my engagement letters and consent language read the way they do. What I built instead is an anonymizer. It walks a document and replaces every value that could identify anyone, keeping a map of what it changed. The anonymized copy goes through my normal process. A separate step puts the real values back afterward from that map, so the analysis lands with the right context attached. The agent sees the structure of a situation. It does not see the person.
Second, nothing writes in my voice for the public. Agents research what I post and never draft it. I have been accused of using Claude Code too much in what I publish, and beyond that I am proud of my own writing and would rather show it.
Third, the tax numbers get verified by me, personally, and against returns run independently in software I trust and re-verify every year. A wrong tax position files cleanly. It throws no error and fails no test, and nobody finds it until an examination years later, which makes it exactly the kind of error I cannot let a model make quietly.
What review actually means
Nothing reaches a client that I have not read line by line and reworked to match how I talk and what I know about that family. The test is simple: could I say this sentence out loud to this client without qualifying it? If not, it gets rewritten, because eventually I will be saying it out loud.
That standard is also the argument for narrow tasks over broad ones. A workpaper set that does not tie takes longer to audit than to redo. A basis lookup either matches the source document or it does not.
What it bought
Not speed. Capacity. Work moves while I am at the park with my kids, and the agents come find me when they hit something they cannot resolve. Projects move at the pace I want them to move, and I can take on people I would have turned away last year. It has not made me a better preparer. It has made me an available one, and in a solo practice that is the constraint that was actually binding.
If you are trying this in your own practice
Do not offload too much at once, which is the mistake I made in the most expensive way available. Map out what you actually do, in pieces small enough to check, and hand over one at a time. Validate every output, every time, until a specific step has earned the right to be trusted. Write the plan down while you are thinking clearly and keep it where the agent can read it too.
The payoff is not that the machine does your job. It is that the administrative and data entry work stops eating the hours, and what is left is the strategic work, which is the part of owning a firm that was worth doing in the first place.