January is when the document chase starts. Firms send the same list again and again, then call, then send it once more. Replies land in a shared inbox and a portal, and some of them get misfiled. By March, review is backed up. Time often gets logged and never billed because nobody can reconstruct what happened on a file. That is the cost firms feel: hours, delays and write-offs.
When a firm owner searches for accounting GPT, the hope is simple. Maybe a chat assistant can take some of this off the desk. It can take some of it. The useful question is which part, and what has to stay with a person.
Collections tell the same story from the other end. When the work is scattered across inboxes and folders, the invoice goes out late, and the client questions it because the value is hard to see. Nobody wants to argue about a bill that was built from a trail of half-finished notes. Every firm would rather do the work and bill it cleanly.
What it does well
A custom GPT is a chat assistant trained or prompted for accounting work. In practice that means language work. It can explain a rule in plain English when a client asks why something changed. It can draft a first-pass client email, which saves the blank-page stare. It can summarize a long notice into the few points that matter, so a reviewer knows where to look. It can turn rough notes from a meeting into a memo with headings and a clear next step.
It is also good at the awkward drafts we put off. The collection reminder. The note that says we still need a few documents before we can start. The explanation of a deadline that moved. None of these are hard, but each one costs attention, and attention is what we run out of in busy season. A first draft that is close is still useful. For most people, editing a close draft is faster than starting from nothing.
Those are real savings. They are also bounded. The assistant is good at the shape of a document. It is weaker on the facts of your specific file, because it cannot see them.
Where the work lives
The work does not live in the chat window. It lives in the practice management system, the shared inbox, the client portal, and last year's file. The assistant sees none of that unless someone copies it in. So it does not know which documents are still missing, who owns the follow-up, whether the deadline moved, or what was decided on the phone last week. Ask it to draft the chase email and you still have to tell it who is missing what. The paste becomes the job.
A client emails a question about last year's return. To answer well, we need last year's return, the workpapers, the email thread, and the latest notice. A chat assistant gets whatever we paste, in whatever order we paste it, and it has no way to ask the portal for the missing piece. The person becomes the connector again.
The gap shows up in review. A summary can read well and still miss the client's facts. A draft email can be polite and still promise a date the firm cannot meet. Someone has to check it against the file, and that check is the work we were trying to protect. If the assistant cannot see the file, the reviewer is doing the same reconciliation twice.
There is a second problem, and it is bigger than convenience. Pasting client data into a consumer chat tool raises confidentiality questions. In the US, IRC section 7216 requires taxpayer consent before return information is used or disclosed for purposes other than preparing the return. That is not a rule about tools. It is a rule about what leaves the firm and why. A firm that treats a chat box as a safe place to drop a client's return information is taking a risk it may not be able to explain later.
So we end up with a capable assistant that is blind to the file, and a rule that says be careful what you show it. Blind to the file and limited by the consent rule, its value has a ceiling. The assistant can help us think. It cannot carry the work.
The approval gate
The next step is an agent that works inside the firm's systems, with permissions, rather than a tab where we paste text. That changes the question. It is no longer "what can the model do?" It is "where does a person stand in the path?" The design answer we keep coming back to is an approval gate.
An approval gate is plain. The agent prepares the work: a draft email, a summary, a completed checklist, a proposed note for the file. The draft goes into a queue instead of going out. A person reviews it and can approve, edit, or skip it. The system logs who released it and when. Nothing reaches a client until a named person says yes.
In pseudo-code, the shape is small:
draft = agent.prepare(client, task)
queue.add(draft, needs_approval=True)
decision = reviewer.review(draft) # approve, edit or skip
log.record(reviewer, draft, decision)
The gate does two things at once. It keeps a human accountable for anything a client sees. It also gives the firm a record, which matters when a client asks who said what, or when a partner wants to know why a file sat for a week. Approve, edit, and skip are enough states for most work. The log is the part firms underestimate. Without it, nobody can answer the question "who released this?"
Someone has to own the queue. Otherwise it becomes the next backlog, and the agent has simply moved the delay. That owner is usually a manager doing triage, not a partner rewriting drafts.
We would put the gate at the client boundary first, not deep inside the preparation. Let the agent gather, draft, and summarize internally. Stop it at the send button. That is where the risk is highest and where review is already a habit. Later, if the firm trusts a narrow task, the gate can move or loosen for that task alone. Start closed. Open it one task at a time, with a reason.
Wider adoption will not come from a better chat window. It will come from agents that can see the practice management system, the inbox, and the portal, with permissions narrow enough that a partner can see exactly what the agent touches. In our view, the wiring and the gate matter more than the model.
Disclosure: this post is from Precision AI OS. Precision AI OS builds an Agent OS for CPA and professional service firms where agents prepare the work and anything that goes to a client waits for a person's approval by default. That default is the part we would copy first.
https://getprecision.ai/?utm_source=devto&utm_medium=social&utm_campaign=accounting-gpt
Top comments (0)