Contents5
The first serious test of AI agents on real tax files, published in late October 2026, is clear: the best of them pass about one assignment in ten. We analysed it in what the first test on real client files shows. So the useful question for a firm is not “can the agent do the file?”, but “which part of the file can it do, and where does the accountant take over?”.
In short
The agent does the volume work: documents, filing, reconciliations, first draft.
The accountant keeps three moves: validate the facts, question the client, sign.
The agent always returns the list of what it could not verify, and that list is read first.
The file split into two columns
What the agent prepares
Receiving, naming and filing the client’s documents
Reconciling statements, the ledger and prior returns
Spotting missing documents and unsupported amounts
First draft of the workpaper and of the letter to the client
What the accountant decides
The facts the file rests on: family status, residence, date of an event
The questions to the client about what they did not mention
Edge cases and tax elections
Sign-off and sending of the deliverable
This split is not caution for its own sake. It follows exactly the three errors the test found: the agent misses the deadline the client never mentioned, reasons on an assumed marital status, and treats an expense as supported when its receipt is missing.
How a file flows
- The client drops off their documents.
Through a link, a portal or a messaging app, as with the bots that record receipts or our bookkeeping agent on Telegram, which only records an entry above a confidence threshold.
- The agent files and reconciles.
Each document goes to its category, each amount is checked against the statement and the prior year.
- The agent returns a checkpoint.
Not a finished file, but a draft with the list of what it could not verify, the assumptions it had to make and the missing documents.
- The accountant decides.
They read the list of doubts first, call the client if needed, then correct the draft.
- The deliverable goes out, signed by a person.
Never sent to the client automatically.
Key takeaway
The list of doubts is worth more than the draft
An agent that says “I could not find the receipt for this reimbursement” saves more time than one that produces a file that only looks clean. You set it up to flag, not to fill the gaps.
The checkpoint that brings doubts back up
A list of doubts does not write itself. Part of it fits into simple rules, written once: no receipt, no deduction; a missed closing date, an alert. The rest requires reading a situation: “is the family status this advice rests on proven by a document?”, “could an upcoming deadline change the chosen option?”.
For these fuzzy questions, a decision model such as Jev answers yes or no with a probability, in under a second, without writing anything. The firm sets a threshold per type of file, and anything below it goes to the accountant with the precise question. We cover these models in our article on AI decision models; their reliability is measured on the firm’s own files before anyone relies on them.
The time saved, and where it goes
In a firm, data entry, filing and chasing take up a large share of billed hours without being advice. Our view: that is where the agent belongs, and the time freed goes back into what AI does not do, the client relationship and advice.
The model is chosen task by task: a small model hosted by the firm is enough to file documents, a more powerful one is kept for drafting a memo. That keeps costs low, and it is the principle of our AI at the right cost offer. To connect the agent to the firm’s accounting software and CRM, see our CRM offer.
Frequently asked questions
Who is responsible if the agent gets it wrong?
The firm, as for a junior’s work. That is why the deliverable is always validated and sent by an accountant, and why the agent must say what it did not verify.
Do you need one agent per type of file?
No, but one set of instructions per type: an individual return, a company filing and an estate file do not call for the same checks. You write those instructions once, with the accountants.
Does client data leave the firm?
Not if the model runs at the firm or in a closed environment. That is often the condition for a firm to accept AI on tax files.
Jim · founder of Brain Plus AI
Also in business ai
Understand12 min
AI agents in tax work: the full DAYJOB: Tax file
Fifty real tax files, 1,789 documents, ten of the most advanced AI agents: the best passes 10.8% of assignments. What the test really shows, case by case, and how an accounting firm can use it now without handing its judgment to a machine.
Understand7 min
Accounting firms: one bot for receipts, one for client files
One bot records the receipts clients send in, another answers accountants' questions about a client file: two different uses, and both can run on the firm's own servers without losing quality.
Choose and price9 min
Private AI in business: the four levels of control
A "private" AI is not a product, it is a degree of control over your data. We distinguish four levels, from an API used as is to a model installed on your premises, and a simple way to choose task by task.