Here's a pattern that repeats in small accounting firms. Someone sees a demo, the firm buys a tool, and a month later staff are checking every output twice because nobody agreed on what the tool was supposed to do, what it was allowed to touch, or when it should stop. The hours that were supposed to come back don't. The tool gets blamed, and the next one gets bought the same way.
If you're building or buying AI agents for accounting work, the fix is boring: write a one-page job spec first. It forces the decisions that matter, and it gives you something to test any product against, including one you build yourself.
What goes in the spec
A spec answers seven questions.
What starts the job? A new client signed, a document arrived, a date on the calendar, a status change in the practice management system. One trigger per job makes it easier to trace where a job went wrong.
What does it read, and where does that live? Name the systems and folders. Limit it to what the job needs. For US tax work this matters beyond tidiness: Internal Revenue Code section 7216 restricts preparers from using or disclosing return information without the client's consent, so the read scope is partly a compliance decision.
What does it produce? A draft email, a checklist, a status update, a note in the client file. Outputs should be things a person can review quickly.
Where does it stop? The point where the work is ready for a person, or where the agent hits something it can't resolve. The stop needs to say what's missing, not just fail quietly.
Who approves? A named role, not "someone." If the approver is on vacation, who covers?
What must it never do? Send to a client without approval, sign, file, pay, delete. Write these down even if they seem obvious, because they're the first things to check in any tool's permission settings.
What gets logged? Every read, every draft, every approval, with timestamps. When a client asks why they got a reminder for a form they already sent, you want the answer to be one search away.
An example
Here's a spec for the job most firms should try first, chasing missing documents during tax season:
job: missing-documents-follow-up
trigger: weekly, Monday 7am, for clients in "waiting on client" status
reads:
- request_list (practice management, per client)
- received_documents (portal folder, per client)
- prior_year_return_summary (read-only)
produces:
- draft_email: lists only items still missing, plain language
- status_note: what arrived since last week
stops_when:
- draft is ready for review
- a received document can't be matched to a requested item
approver: client's assigned preparer
never:
- send without approval
- mark a request complete without a matched document
- delete or rename client files
log: every read, draft, edit and approval, kept with the client record
Nothing in there is clever. That's the point. Each line is a decision someone in the firm had to make, and each one is something you can check a vendor against: can your product do this, and can it be prevented from doing that?
Good first jobs
The best starting jobs have a clear input, a clear draft as output, and someone who already reviews that step today. That last part matters, because it means you're not adding a new review burden, just changing what the reviewer looks at.
Missing-document follow-ups are the obvious one, as above. Close behind is a checklist comparing this year's received documents against last year's return, so the preparer sees at a glance that the rental property schedule has no supporting statements yet. Engagement letter drafts work well when they're built from the firm's own template and the services already agreed, with a partner approving before anything is sent. For bookkeeping clients, a month-end status summary listing missing reconciliations and unusual variances saves the manager from assembling it by hand. And when an IRS notice comes in, matching it to the right client, tax year and return, and laying out what it says, puts the file in front of the right person with the facts laid out. Deciding what to do about the notice stays with them.
What these have in common is that the agent prepares and a person decides. None of them involve the agent committing the client to anything.
How to measure it
Pick one job and time it for a week before anything changes, per client if you can. Then run the agent on that job for a month with every output reviewed.
Track two things. Hours spent on the job, before and after. And the rewrite rate: how often the reviewer had to substantially change a draft instead of approving it with light edits. The second number tells you whether the spec is right. A high rewrite rate usually means the agent is reading the wrong inputs or the output format doesn't match how the firm actually writes to clients, both of which are spec problems, not model problems.
If both numbers improve over the month, write the spec for the second job. If they don't, you've found out cheaply, and you have a document that tells you exactly what to change.
Precision AI OS builds an Agent OS for professional service firms, where agents prepare the work and anything that goes to a client waits for a person's approval by default.
Top comments (0)