Ten at night on the last Friday of the quarter. The lights in the accounting room were still on, the coffee had gone cold, and the screen showed a spreadsheet called expense_reconciliation_final_v7_actual.xlsx.
Sound familiar? Every close period looks the same: invoices arriving late, missing paperwork, one supplier chasing payment, and someone scrolling line by line to work out which piece of paper a number came from. That’s the week we started looking into AI for accounting and tax filing.
That week we found a payment that had gone out twice. Same supplier, same amount, two different invoice numbers — so nothing in the reconciliation sheet flagged it. Nobody was being careless. It’s just that no one can hold three months of numbers in their head.
Machine reads and matches. Human decides. Sounds simple, but almost all of the engineering goes into drawing that line in the right place.
What AI for accounting and tax filing actually does
Ask ten people and you’ll get ten answers — everything from “invoice OCR” to “it files the return for you”. What the work actually looks like for us is narrower:
- Reading and extracting documents. PDF invoices, phone photos, confirmation emails, acceptance records, bank statements. The agent pulls out dates, amounts, tax codes, supplier names, line items — even when every supplier formats things differently.
- Cross-checking. A number never stands alone. The invoice has to match the purchase order, the purchase order needs an acceptance record, and the payment has to show up on the bank statement.
- Flagging the mismatches. Duplicate payments, prices that drifted from the contract, missing documents, wrong period, one supplier with two tax codes. The things human eyes skip at ten at night.
- Gathering figures for the filing. Grouping items under the right headings, with a link back to the source document left intact. This is the part you can try on the live tax-filing agent — it prepares 01/GTGT and 01/NTNN returns for small and medium businesses.
- Preparing draft entries. Each one with its documents and a reason attached, ready for an accountant to approve, change, or discard.
That last point matters most, so let’s be blunt: the agent prepares. A person decides.
In the systems we build, no entry reaches the ledger without a qualified person looking at it. The agent has no authority to post, no authority to submit, and no authority to quietly change a number that has already been approved.
One small detail decides whether any of this is trusted: every output has to be traceable. Click a figure and you land on the source document — which page, which line, which date. If it can’t be traced, an accountant won’t believe it. And they’d be right not to.
One close period, before and after
We’re not going to sell this with pretty numbers, so here’s what changed in the work itself.
Before. Closing started with hunting for documents. If the chief accountant wanted to see one expense, the team lost a morning digging through an inbox. Everything ran on memory and on one colleague’s notebook. On the final days the whole room switched into data-entry mode: heads down, keyboards, and the only question anyone was allowed to ask was “are you done yet?”
After. The reading and matching happen before an accountant sits down. By the time they open the queue, the list has already shrunk: matches stay silent, mismatches surface with a reason and the documents attached. Nobody has to keep the answers in their head anymore, because the answers sit in the system.
What we didn’t expect is that the pain moves rather than disappears. Mornings are no longer eaten by searching for paper, but more time goes into handling exceptions — because now the team sees things that used to slip past. Nobody is a document chaser anymore. They’re the person making the call.
And the call is the valuable part.
What AI should not decide on its own
If we keep only one lesson from all those close periods, it’s this: an agent’s value isn’t measured by how much it dares to do, but by whether it knows when to stop.
- Exceptions and unclear policy. When a transaction’s treatment depends on context that was never written down as a rule, the agent stops and asks. Guessing in accounting is the fastest way to create work for your future self.
- Disputes. A price disagreement with a supplier, a quantity discrepancy, a payment both sides read differently — that’s a human job, with calls, emails, and sometimes goodwill.
- Professional judgement. How a particular transaction should be treated is a question for someone qualified, not for a language model.
- Sign-off and accountability. The filing belongs to the accountant and the legal representative. No model carries that responsibility for them.
And one thing said plainly: you won’t find a single rate, deadline, or filing requirement in this article. That’s deliberate. Tax rules change constantly — by year, by guidance, by industry, and case by case. What was right last year can be wrong this year.
That’s exactly why the agent prepares and doesn’t decide. A qualified professional reviews the work against the rules as they stand today, and takes responsibility for what they sign. If someone tells you their AI handles the filing and needs no human check, ask them one question: who signs?
Start with one small process
With a close period in front of you, the tempting move is to do everything at once. It’s also the most reliable way to fail.
When people talk about automate work processes, they usually picture one big project. Ours is smaller: pick one document type with high volume and a clear, repetitive treatment. Reconciling purchase invoices against bank statements, for instance, or checking acceptance records before an expense is recognised.
Then do three things, in this order:
- Measure the current cost. Per close period, how long it takes, how often it happens, where it goes wrong.
- Run the agent alongside the existing process, not instead of it. Each period, compare the agent’s output with the human output and look at where they diverge. Divergence isn’t necessarily failure — it’s data, telling you where the agent can be trusted and where it can’t yet.
- Only once results have matched for long enough do you drop the manual step, then widen to the next document type.
The middle step is the one people skip. It’s also the one that decides whether this works, because an accounting team only trusts a system after they’ve checked it themselves.
Tell us about your close period
If your team spends most of the close period reading and matching documents, that’s actually good news — the process is already repetitive enough to hand part of it over.
At Mon AI we build agents like this for accounting teams: reading documents, cross-checking, flagging, preparing draft entries — and stopping exactly where a person needs to sign. If you’d like to know where a close period at your company could start, send us a question — even if you’re only curious. You can also try the tax-filing agent yourself and see the 01/GTGT or 01/NTNN .xml it produces, or look at how we work and what it costs first.
Then come back on the last Friday of next quarter. The lights might still be on — but not because someone is hunting for a single invoice.
