Faster month-end close with AI: split rule-based and judgment exceptions
Let AI compute the rule-based exceptions in your close; for judgment exceptions, AI shows evidence and a person decides.
Key points
- List exceptions from last month's close
- Split them in two: can a rule answer it?
- Let AI do the math on fee shortfalls
- Odd payers: AI shows why, you decide
Contents
List the items that stalled last month as exceptions
A&A perspective
If you want AI to speed up your month-end close, start by listing, one by one, the items where last month's close stalled. An exception here means a single item where the books and the bank statement did not match as expected, so someone had to stop and think about it.
Hypothetical example
For example, take a company with four employees that has an outside bookkeeping service handle its bookkeeping and month-end close. When a payment does not match an invoice, the bookkeeper sends a question. The close for that month waits until the owner answers.
From the sources
Anthropic has published a case study on Campfire, a North American provider of accounting software for high-growth tech companies. Campfire uses Anthropic's AI model, Claude, to reduce monthly close times. The case study says traditional accounting software had reached its limits with rule-based automation.
From the sources
One Campfire customer says bank reconciliation, matching bank statements against the books, used to consume 80% of their bandwidth.
A&A perspective
Automate & Augment (A&A), the company writing this article, sees the clue to a faster close in the items where work stopped. Until you can see which items stalled, you cannot decide what to hand to AI either.
A&A perspective
Four columns are enough: date and amount, why it stalled, who resolved it, and how long it took. One month of entries becomes your list of candidates to hand to AI.
Split exceptions in two by whether a rule can answer them
A&A perspective
Sort the exceptions you listed into two groups with a single question: can the answer be written as a rule? Those that can are rule-based exceptions; those that cannot are judgment exceptions.
Hypothetical example
A payment that arrives short by exactly the bank transfer fee is a rule-based exception. You can write the rule: if the difference equals the fixed fee, book it as a fee. A single payment that covers several invoices is another: if it equals the total of those invoices, a rule can decide which invoices it pays.
Hypothetical example
A payment whose payer name differs from the customer's name is a judgment exception. Whether the customer's owner paid personally or another company paid on the customer's behalf cannot be settled without knowing the circumstances. A payment that differs from the invoiced amount is another: a partial payment and a discount are booked differently.
From the sources
In December 2024, Anthropic published an article for developers on how to build systems that use AI. It describes a pattern that first classifies an input by type and then sends it to handling specialized for that type. The pattern is said to work well when there are distinct categories that are better handled separately and the classification can be done accurately.
From the sources
A note at the top of the article says that much of the tooling it describes has changed since it was published.
A&A perspective
A&A thinks the same idea applies to exceptions in the month-end close. If you hand both kinds to AI as a single list of exceptions, AI may give confident answers even on payments that need judgment, and someone may end up rechecking every item. Narrowing the split down to one question is what keeps the classification accurate.
| Aspect | Rule-based exceptions | Judgment exceptions |
|---|---|---|
| How to tell | A rule can give the answer | Cannot be settled without knowing the circumstances |
| Examples | A payment short by the fee; one payment covering several invoices | A different payer name; an amount that differs from the invoice |
| AI's role | Computes | Lays out the evidence |
| Person's role | Checks the answer | Reads the evidence and decides |
Let AI compute the rule-based exceptions
A&A perspective
Have AI compute the rule-based exceptions, and move people into the role of checking the answers. Because the rule is written down, a person can tell quickly whether AI's answer is right.
From the sources
The Campfire case study lists automated bank reconciliation, which parses bank statements and identifies discrepancies, among its features. It says customers cut the time spent on bank reconciliation by 90% and shortened their close by three days per month on average.
From the sources
Anthropic's article recommends finding the simplest solution possible and only increasing complexity when needed.
A&A perspective
A&A thinks that some rule-based exceptions may already be handled by the rule features in your accounting software. AI computes the ones whose rule you can write out in words but cannot set up in the software's settings. In that case, write the rule out in plain words and give it to AI together with the bank statement and the list of invoices.
A&A perspective
Campfire's numbers come from the company, and they were achieved in the accounting teams of high-growth tech companies. A&A does not recommend taking them as they are as the target for a small company's close.
A&A perspective
Change how much a person checks in two stages. The checker is whoever handles the close: an in-house staff member or the bookkeeping service. In the first month, a person reviews every answer from AI. Once errors stop appearing, move to spot checks, such as picking one in ten to review. If an error appears, go back and fix how the rule is written.
A&A perspective
Among the exceptions that used to stall on questions from the bookkeeper, the rule-based ones are candidates to move to AI. If you share the rules with the bookkeeping service, they can check the answers against the same rules on their side.
Have AI show evidence on judgment exceptions, and let a person decide
A&A perspective
For judgment exceptions, have AI show which records it looked at and how it read them, and let a person who knows the circumstances decide. AI does not settle the answer; its job is to gather the candidates and the evidence.
From the sources
Separately from automated reconciliation, Campfire built a chat interface that lets users ask about their financial data in plain language. The case study says it explains its logic and reasoning, shows its work, and provides source data. It calls providing source data a critical requirement for accounting professionals.
From the sources
Anthropic's article also describes AI that goes back to a person for more information or judgment along the way, and pauses for human feedback at checkpoints.
Hypothetical example
Say a payment comes in under an individual's name. Have AI give you three things: the likely customer, why it reads the payment that way, and who to confirm with. The reasons might include an invoice with a matching amount, or past payments under the same name. The owner reads the evidence, decides if it holds up, and asks the customer if it does not.
A&A perspective
Among judgment exceptions, those that need a company-level decision are settled by the owner with the tax accountant. How to treat a receivable that is unlikely to be collected is one such decision. Here too, AI's role stops at pulling together the related records and past correspondence.
A&A perspective
Tidying up your customer records moves some judgment exceptions over to the rule-based side. If the same individual's name stalls the close every month, linking that name to the customer in your records turns it into something a rule can handle. Revisit the split after each month's close.
Split exceptions first, then hand them to AI
A&A perspective
Before using AI to speed up the month-end close, list the items where last month's close stalled as exceptions.
A&A perspective
Split the exceptions in two by whether a rule can give the answer.
A&A perspective
Have AI compute the rule-based exceptions. In the first month a person checks every answer, and once errors stop appearing, switches to spot checks.
A&A perspective
For judgment exceptions, have AI show its evidence and let a person who knows the circumstances decide. Those that need a company-level decision are settled by the owner with the tax accountant.
Let AI compute the rule-based exceptions in your close, and for judgment exceptions have AI show its evidence so a person decides. Revisit the split after each month's close, and tidy up your customer records so that some judgment exceptions move over to the rule-based side.
FAQ
Will accounting software with AI speed up my close even if I don't split exceptions?
It may for rule-based exceptions. If you hand over judgment exceptions as well, someone ends up rechecking every AI answer, and the close is unlikely to get faster. When you compare products, check which of the two kinds of exceptions they can handle and whether they can show their evidence.
What if a tax accounting firm handles all of our bookkeeping and closing?
Decide the split together with your tax accountant. If you turn last month's questions from them into a list of exceptions, you can discuss which ones a rule can settle and which ones cannot be settled without asking about the company's circumstances.
Is splitting worth it if we only have a few exceptions a month?
With only a few, it can be faster for a person to handle them without splitting. It is enough to decide on the split when the number grows in a month, or when an exception with the same payer name starts stalling the close every month.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Campfire accelerates accounting with Claude
Anthropic · Publication date not stated on page
Accessed 2026-10-22 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-22