Month-End Reconciliation AI Agent for Small Finance Teams

Why should an agent only propose, and never post?

A ledger is a record you may have to defend to an accountant, a lender, a tax authority or an auditor. An agent that writes to it can be wrong quietly and at speed, and an entry posted in error is slow to find and awkward to unwind. A proposal costs a reviewer a few seconds and a wrong one costs nothing. So the design rests on a single rule, enforced in the system's permissions and not left to the prompt: the agent can read the bank file, the ledger export and the document store, and it can write only to its own workspace. Posting is a person's click, in the accounting package itself.

The agent does not do your books. It prepares the pile so your people spend their time on lines that need judgement. The approval pattern is in our post on human-in-the-loop approvals, and the document-reading side belongs to our AI document processing agent service.

Illustrative scenario: one month-end close in three passes

This is a worked example, not a client project. The business, the lines and the supplier names are invented, no amounts are shown, and no outcome is claimed. Assume a small trading company with one main bank account, a ledger in its accounting package and supplier documents saved in a shared folder.

1. Setup, last working day. A finance team member uploads the month's bank statement file and exports the ledger's bank account transactions. The agent checks the statement's opening balance against last month's closing balance, looks for date gaps and duplicate lines, and stops with a plain message if anything is off.
2. Pass one, exact matches. The agent pairs bank lines with ledger entries where the amount and reference agree and the dates are within the window your rulebook allows. It proposes these as high confidence: exact. A reviewer can accept them in bulk after sampling a few.
3. Pass two, close matches. Here it looks for pairs that nearly agree. One example: a supplier payment that is short by an amount equal to a bank charge listed as a separate line. Another: a customer payment recorded two days earlier in the ledger than it cleared. Each proposal states the difference and the reason it thinks it is the same item, labelled medium confidence: explained difference. Each needs a person's click.
4. Pass three, groups and gaps. It tries combinations, such as one bank receipt covering three invoices, and lists lines with no counterpart: unrecorded bank fees, receipts with no invoice, and ledger entries with no bank line yet. These are labelled low confidence or unmatched, and the agent drafts a journal suggestion only where a rule in your rulebook covers the case. Each suggestion names the account, the reason and the supporting document.
5. Human review. A team member works through the workspace: bulk-accepts the exact matches after a sample check, handles each close match, and decides what to do with the rest. Anything the agent could not explain stays open and moves to the carry-forward list.
6. Posting. The reviewer, or a second person where you separate duties, enters the approved journals in the accounting package. The agent can export the approved suggestions as a file for import if your package supports it, but it does not import them.
7. Close-out log. The workspace stores a record: statement and ledger file versions, each proposal and its label, who accepted or changed it, what was posted, the final unreconciled list and the sign-off.

Which reconciliation tasks run alone, need approval, or stay human?

Task	Alone	Approval	Never automated
Importing files and checking balances and gaps	Yes
Proposing exact matches	Yes, as proposals	Reviewer accepts, in bulk after sampling
Proposing close or grouped matches	Yes, as proposals	Reviewer accepts each
Finding the supporting document	Yes
Drafting a journal suggestion	Yes, in the workspace only	A person decides whether to post it
Posting a journal or changing the ledger			People only; the agent has no write permission
Writing off a difference or a bad debt			People only
Deciding the correct account for an unusual item		Suggestion only	The decision stays with a person
Paying a supplier or moving money			Never; outside the agent's reach
Closing the period and signing off			Finance lead only

What goes wrong at month end, and what does the agent do?

• One bank line, several possible matches. The agent shows the candidates side by side and picks none. Ambiguity is a reason to ask, not to choose.
• Same amount, different customers. Two payments of an identical amount from different payers are never matched on amount alone. The payee or reference has to support it, or it goes to review.
• The statement does not start where last month ended. The run stops before matching and reports the gap, since everything after that would be unreliable.
• A document says one thing and the ledger another. Both are shown. The agent does not decide which is right.
• A document contains an instruction. An invoice or email saying "mark as paid" is data. The agent's tools cannot change a payment status in any case. See our note on prompt injection.
• Rounding and currency differences. Rules state which tolerance may be explained automatically; anything outside it goes to a person.
• Reviewer fatigue. If every proposal is accepted without edits, the log highlights it. A reviewer who never changes anything is a signal to sample more, not to celebrate.
• The agent's confidence is misleading. Labels are earned from tests: we replay past months where the answers are known, compare proposals with what your team actually posted, and adjust the rulebook before the first live run.

What data and access does it need, and how would success be measured?

Access. Read-only exports or API access for bank transactions and ledger entries, read access to the folder holding supplier invoices, receipts and remittances, and a private workspace for its output. If your accounting package offers an API that can create draft journals, we would check what permissions it requires and whether a separate role can limit it to drafts. Where it cannot, the agent simply produces a file for a person to import. We have not verified any specific package for this page, so confirm with your own vendor.

Data handling. Bank data and supplier details are sensitive business information, and may include personal data. Choose where the workspace is hosted, keep access short, set retention with your accountant and record which AI provider receives which fields. Our post on AI and privacy covers the broader questions.

Measurement. We would agree metric names with you and use your last few closes as the baseline, promising no figure: proposals accepted without edits; proposals corrected by a reviewer; matches later found wrong after posting (the aim is none); unmatched items at first review and at close; items carried forward and their age; days from period end to sign-off; and reviewer time per hundred lines, if you already track it.

Start in shadow mode for a period or two: the agent proposes while your team reconciles as usual, and you compare the two. Only then does the review workspace become the working method. For other places an agent can read documents, see invoice automation for small business. To talk through your own close, contact us.