A year ago I started running parts of my CA practice on AI in earnest — not experimenting with it, actually depending on it for work that goes to clients. Here is an honest account of what changed, what did not, and what I got wrong.
Where it genuinely changed the work
Reading large volumes of documents. This is the biggest single gain. Ledgers, bank statements, agreements, notices. Work that used to be a day of reading now starts with a structured summary and a list of the things worth looking at. I still read the important documents myself. I read them knowing where to look.
First drafts of everything. Replies to notices, memos, engagement letters, client explanations. A first draft that is eighty percent right in two minutes changes the economics of written work completely. The remaining twenty percent is where the professional judgement is, and that is where my time should have been going all along.
Reconciliation and anomaly spotting. Comparing what is reported in one place with what is reported in another, and flagging differences. It does not decide anything, but it reliably finds the rows a human eye slides past on the two hundredth line.
Explaining things to clients. Turning a technical position into plain language, at the right length, for a specific reader. This used to be a genuine time cost in my week. It is now close to free, and the quality is better because I am not writing the fourth explanation of the same thing on a Friday evening.
Where it did not help at all
Anything requiring a current, authoritative position. Rates, thresholds, due dates, the state of a provision after the last amendment. A model that sounds fluent while being out of date is worse than no answer, because fluency is persuasive. Every figure that reaches a client is verified against the source. Every time.
Judgement calls with consequences. Whether to take a position, how aggressive to be, whether a client’s explanation will survive scrutiny. These require knowing the client, the officer, the history, and what the business can absorb if it goes badly. That is not in the training data.
Responsibility. When something is signed, a person is accountable. That has not moved and it should not. The tool does not share the risk.
What I got wrong
I over-trusted it early. The output is confident and well-formed, and confident well-formed output reads as correct. I caught a statutory reference that was plausible, precisely formatted and simply wrong, which was the moment the verification discipline became non-negotiable rather than a good intention.
I also automated before I had standardised. I built things to speed up processes that were themselves inconsistent, which produced faster inconsistency. The order that works is: make the process repeatable, then automate it. Automating a mess gives you an automated mess, with less visibility into it than before.
And I underestimated how much of the value would come from small things rather than large ones. I expected a dramatic replacement of some big workflow. What actually happened was a hundred small frictions disappearing — the summary, the draft, the reformatting, the first pass over a file — and the cumulative effect being larger than any single automation.
What I will not do
Client data discipline is the line I hold. What goes into a tool, where it is processed and what is retained is a decision made deliberately for each kind of information, not assumed. Professional confidentiality obligations do not soften because the recipient is software.
I also will not let it write in a voice that pretends to be considered advice without a professional having considered it. Everything that leaves my desk has been read by me. That is the whole basis on which a client is paying for a named professional rather than a subscription.
The honest summary after a year
It has made me faster at the parts of the work that were never the point — reading, drafting, formatting, summarising — and it has made no difference at all to the parts that are. I spend more of my week on judgement and client conversations and less on mechanics, which is a better job than the one I had.
The firms that will struggle are not the ones that avoid the technology. They are the ones that adopt it without changing what they charge for. If you are still billing for the mechanics, the mechanics are exactly what is getting cheaper.
Most of my firm work — audit, cross-border structuring, transfer pricing — runs through AeTx, and the same principle applies there: the tooling changed how the work gets done, not who is accountable for it.