It is the middle of January. A preparer is on the phone again asking for a missing statement. The return waits. The reviewer is on two other returns. Another client is still waiting on last week's answer. Most of that time never gets logged, so it never gets billed. The season's real cost is hours lost to the chase, returns that slip to extension, write-offs nobody bothers to fight anymore, and clients who leave over the wait.
So when a partner asks for the best AI for accountants, the honest answer is a question back. There is no single answer. Anyone who gives you one without asking about your worst week is selling something.
The worst week looks different in every firm, but the shape is familiar. It's usually one of these:
- Chasing documents clients swear they sent
- Sorting an inbox that mixes client questions with newsletters
- Drafting routine client emails from scratch
- Turning review notes into something a client can act on
- Rebuilding time from memory, then chasing balances nobody wants to call about
Pick the one that costs you the most and judge every option against that job. A tool that is excellent at a job you don't have is a subscription with a nice demo.
Start with the job, not the tool
Most firms have bought software that arrived with a promise and left behind a rollout nobody finished. The pattern is familiar. The demo is clean. The sample data is tidy. Then real work starts and someone has to export a file, rename it, upload it somewhere else, and paste the result back. The tool works. The workflow does not. Staff quietly return to the old way because the old way, for all its faults, is one step shorter. The old way is not good. It is just familiar, and familiar wins when the alternative adds steps in the middle of busy season.
So the first question is not what the tool can do. It is where it sits. Does it work inside the practice management system, the email you already use, and the ledger you already keep? Or does it ask your team to copy and paste between windows? Copy and paste is where adoption goes to die. It's also where small mistakes creep in, every time a result gets retyped or pasted somewhere new.
The second question is about data. Where does client information go, who can see it, and does it train anyone's model? You want a plain answer, in writing, before anything is switched on. If the answer is vague, treat that as a no.
Tax work adds a constraint that is easy to forget in a busy week. Under IRC section 7216, preparers are restricted from using or disclosing tax return information without taxpayer consent, and the rules set out specific consent requirements. This isn't the place for a full walkthrough. The point is simpler. Before you put return information into an outside tool, know what the rules require and make sure consent is handled. "It was just a draft" is not a defense anyone wants to test.
Five questions that separate a real option from a demo
Here are five questions to ask every vendor, the same way every time.
Does it work inside the systems you already use, or does it create a new place to copy and paste?
Where does client data go, how long is it kept, and does it train anyone's model?
Who approves before anything reaches a client?
Can you see a record of what it did, in plain language, after the fact?
Can you measure it on one job for thirty days and see the hours?
Write the answers down. A verbal yes on a sales call has a way of becoming a maybe by the time someone else asks. That last question matters more than the rest, though. Pick a single job, the document chase or the inbox or the review notes, and run the tool on it for a month. Count the hours before and after. Count the returns that slipped and the emails that went out late. Do not count the demo. Do not count the enthusiasm of the person who liked the sales call. Count the work. If the hours do not move, the tool is not solving your worst week, whatever else it does well.
The approval question deserves its own attention. The bigger risk usually isn't a draft that's obviously wrong. It's one that sounds plausible and goes out. A client email that sounds right but misstates a deadline is worse than a slow email, because it creates work to undo. Anything that touches a client should have a person in front of it by default, not as a setting someone can forget to turn on.
Red flags, and the rule that outranks the rest
A demo built only on sample data tells you almost nothing. Sample data behaves. Yours does not. Ask to see the tool work on a messy version of the job you actually have, or ask for a trial where you supply the mess.
A vendor who cannot answer the data retention question, or who answers it differently the second time you ask, is telling you something. So is pricing that changes per seat with no cap. A per seat price sounds small until the season adds people and the bill does not stop where you expected. You want to know the ceiling before you sign, not after.
The rule that outranks the rest is simple. No client-facing output goes out without a person approving it. Not because the tool is bad. Because the relationship is yours, and the client will call you, not the tool, when something is wrong.
Then measure, on one job, for thirty days. If the hours fall and the work still gets reviewed, you have found your answer. If the hours only move because review got skipped, you have found a problem.
Precision AI OS builds an Agent OS for professional service firms. In it, agents prepare the work, and anything that goes to a client waits for a person's approval by default.
Top comments (1)
Hi Precision AI OS. The export, rename, upload, paste-back cycle you describe is what actually kills a rollout. I've watched review notes sit in a PDF until someone retypes them into a client email, and that's the step everyone quietly skips. If it'd help, send a sample review note and I'll show what a client-ready version could look like.