Use AI to prepare the month-end exception meeting

A worked close process that groups problems and preserves the evidence an accountant needs to resolve them.

Month-end often turns into a meeting about rows that nobody has had time to explain. A missing receipt, a bank difference and an invoice dispute sit together under the same red label. AI can help organise the conversation, but only after the team agrees what each exception means.

A real accounting-agent rollout was announced by Intuit in 2025. The Rel-Cat paper studies relationships in transaction classification. BIS notes the importance of governing financial AI and understanding data limitations. None establishes an independently verified close-time saving for the example below. [Intuit](https://investors.intuit.com/news-events/press-releases/detail/1258/intuit-introduces-ground-breaking-virtual-team-of-ai-agents-to-fuel-growth-for-businesses), [Rel-Cat](https://arxiv.org/abs/2506.09234), [BIS](https://www.bis.org/publications/fsi-summary-financial-stability-implications-artificial-intelligence-executive-summary).

Imagine a fictional wholesaler, Orchard Parts, preparing its April close. Its exception register contains a supplier credit not yet matched, two employee expenses without receipts, an unreconciled bank fee and a customer invoice under dispute. The finance lead wants a twenty-minute meeting that allocates the right next action. Asking an AI to resolve the close would be too broad. Asking it to draft an agenda from a reviewed register is a bounded, useful test.

Create one record per exception. Each has an owner, period, amount, type, latest evidence and next action. Use separate fields for what is known and what someone believes. For example, a purchaser believes the supplier issued a credit; the known fact is that a credit document has not arrived. That distinction prevents the model from converting a hopeful conversation into a ledger event.

The assistant proposes three agenda groups: evidence to obtain, records to match and decisions for the accountant. It puts the two missing receipts in the first group and the supplier credit in the second only if a valid document is attached. Otherwise the credit remains an evidence request. The disputed customer invoice goes to the accountant with the contract and correspondence references, rather than to an automatic write-off recommendation.

A worked agenda item might read: supplier credit, owner purchasing, request the document by Thursday, then confirm the matching bill. Its purpose is clear even if the reader never sees the AI. The owner can correct the action before the meeting. At the meeting, the finance lead records the decision in the register. The assistant may later turn approved decisions into a summary; it does not invent agreement from an absence of comments.

The hardest exceptions frequently span periods. Orchard receives an April invoice in May for goods delivered in March. An AI might select the invoice date because it is prominent. The close process should ask an accountant to determine the correct period under the applicable policy. Include this case in the test set. Include a credit in a different currency, a payment covering several bills and a duplicate receipt photograph as well. Those tests reveal whether grouping helps or conceals the actual problem.

Protect the evidence chain. A generated summary links back to the original exception identifier. If someone changes the amount or owner, the system records who changed it and when. The draft carries its source version so it cannot circulate as a fresh agenda after the register has moved on. Where the AI tool cannot maintain those links, use a plain document with identifiers and a human comparison before distribution.

The potential gain is less meeting time spent discovering basic context. Orchard measures items reaching a clear decision, items reopened because context was wrong and the age of outstanding exceptions. Counting closed items alone encourages premature closure. A reviewer also samples cases the assistant called routine, since mistakes are not confined to the rows it marks uncertain.

There is a cost in maintaining the register, training staff to distinguish assumptions and checking proposed actions. A provider's feature name does not eliminate that work. Intuit's cited release described product-dependent availability in a US rollout; current plans and other countries need direct verification. This approach can also be piloted with an approved language tool and a sanitised export without giving it ledger-writing access.

The meeting succeeds when every attendee leaves knowing the evidence to find, the decision to make and the person responsible. AI earns a place by improving that preparation. The close itself remains an accounting process with accountable people and reconstructable records, rather than a persuasive paragraph that makes the red rows disappear.