A&A INSIGHTS
Closing books faster: choose which exceptions to give to AI
For owners of small companies who want to close the monthly books faster: use Anthropic's Campfire case and its engineering essay on effective agents to separate exceptions into rule-shaped, interpretation-shaped, and decision-waiting shelves before choosing an AI product.
日本語で読むTHE STARTING POINT
When you consider AI to speed up the monthly close, treating all exceptions as one bucket risks having AI answer the ones that require judgement. Using Anthropic's Campfire customer page and its engineering essay "Building effective agents" as lenses, this article proposes classifying exceptions into three shelves — rule-shaped, interpretation-shaped, and decision-waiting — before selecting an AI product. A hypothetical ten-person company grounds the shelves in a concrete month-end.
"Close the books faster" is really about the shape of the exceptions
A&A perspective
When an owner asks for advice on closing the monthly books faster, the delay usually concentrates in "exceptions." Routine journal entries and bank matches are not the slow part; what remains on the accountant's desk are unmatched statement lines, name mismatches, entries that straddle a month boundary, and transfers whose supporting document is missing. Introducing AI to speed the close is really about deciding who handles this exception pile, in what order, and treating the whole pile as a single unit that AI can absorb tends to let AI answer the heavier judgement cases as well.
Hypothetical example
Imagine a hypothetical ten-person company. The monthly close currently takes about seven business days, and the last three days are spent by the accountant on "exceptions." That pile is a mixture: cases resolvable by a rule but numerous (fee allocation across categories, matching deposits from several accounts to receivables), cases that require interpretation to close out (multiple counterparties sharing the same remittance name, month-end invoices with annotated notes), and cases that must wait on a decision from the owner or the outside tax accountant (reserves whose estimate must be refreshed, potentially uncollectible accounts). While these three sit under a single "exceptions" heading, adding AI stalls at the same last column as before.
The Campfire case shows two shapes of AI involvement
From the sources
Anthropic's Campfire case page describes an accounting platform built on Claude, and highlights the average reduction in days-to-close for the monthly cycle together with the reduction in time spent on the bank reconciliation workflow. On the same page, automated reconciliation that parses bank statements and identifies discrepancies and an Ember AI chat interface for natural-language questions are described as distinct capabilities. The page also quotes a customer who noted that bank reconciliation had previously consumed roughly eighty percent of the team's bandwidth.
A&A perspective
What A&A takes from this case is not a target number, but that AI participates in two distinct shapes. The first is a machine match that lines up statement lines against internal records and flags rule-decidable discrepancies. The second is an explanation surface where a person can question why the facts were interpreted a given way. The case page itself describes Ember as an interface that shows the reasoning and the source data it consulted. Automated matching speed and an explanation surface are separate levers on the same close-time goal.
A&A perspective
For a small company, the important step is not to conflate the two shapes into a single "let AI handle exceptions." The first shape has straightforward rule surfaces once statement formats are stable, and either an AI or the rule engine of an existing accounting package can approach comparable output. The second shape requires that the words behind an interpretation be visible to the accountant or the outside tax advisor, immediately, at review. Copying only the headline numbers from a case like this typically drops these two preconditions along the way.
Sketch the branching before choosing an AI product
From the sources
Anthropic's engineering essay "Building effective agents" advises finding the simplest solution possible when building with LLMs, and only adding complexity when a simpler solution falls short. The same essay describes a routing pattern that classifies an input in advance and directs it to a specialized follow-up task, and treats returning to a person for further information or judgement as a normal part of the loop. The essay was published on December 19, 2024.
A&A perspective
Applied to a monthly close, this reads: rather than asking early on "AI or a person?" for the whole close, decide first which shelf each exception falls onto. Three shelves are enough. The first shelf holds rule-decidable exceptions — fee allocations, statement-to-invoice matches within tolerance — where the condition can be written down. The second shelf holds exceptions whose meaning must be interpreted before they can be closed out. The third shelf holds items that require a decision the accountant cannot make alone.
A&A perspective
Once the branching is sketched, product comparison narrows to "which shelves can this product handle?" A product that can fully automate the first shelf but produces answers on the second whose sources are hard to re-check gives you an effective gain that is limited to the first shelf. If you want to extend AI to the second, insist that the statement IDs it consulted and the words behind its interpretation be visible to the accountant. Without deciding the shelves first, the yardstick of the comparison ends up being defined by the vendor UI instead of by your company.
Reshelve your own exceptions
A&A perspective
The reshelving work starts by rereading the last one or two monthly close records. Take each case the accountant marked as an exception and classify by the actual action that resolved it: a category was decided by rule (rule-shaped), several counterparties sharing one remittance name were split by interpretation (interpretation-shaped), the scope of a reserve was confirmed with the owner (decision-waiting). Even work that is called "reconciliation" in the log can land on different shelves depending on the action taken.
Hypothetical example
Continuing the ten-person example, suppose reshelving the last month's fifty exceptions produces thirty rule-shaped items, fifteen interpretation-shaped items, and five decision-waiting items. Now the accountant can decide whether the useful lever is reducing the thirty, or shortening the review of the fifteen. Attempting both at once tends to produce a side effect where rule-shaped items are handled by AI in an interpretive way, and their evidence is not preserved. The five decision-waiting items are removed from the AI conversation entirely; only the routing and reporting form is standardized for them.
A&A perspective
Reshelving is not a one-time task. Continue the classification each month and watch whether the rule-shaped shelf grows and the interpretation-shaped shelf shrinks. If the rule-shaped shelf swells seasonally, for instance around invoicing peaks near the fiscal year end, that creates a specific window when it is appropriate to widen the AI's scope. Conversely, if the interpretation-shaped shelf stays roughly the same each month, tightening internal category definitions and counterparty master data to move items between shelves is a prerequisite that should happen before adopting AI.
Match the review grain to the shelf
A&A perspective
For the first shelf (rule-shaped), sampling suffices for review. The accountant re-reads a defined percentage of AI-handled cases and confirms that the rule was applied correctly. The sampling rate depends on the materiality of the close and on volume; start high and reduce as confidence accrues. For the second shelf (interpretation-shaped), sampling alone is insufficient. Require that, for every case, AI produce the specific statement lines it consulted and the words behind the interpretation, so the accountant can read the reasoning and hand items back where warranted.
Hypothetical example
In the hypothetical company, the fifteen mixed-counterparty remittances go through AI, which each month emits a per-item list of "which statement lines were assigned to whom and why." The accountant reads each item's reasoning. If the implementation cannot emit that reasoning, the accountant ends up re-matching passbook data against the invoicing ledger anyway. During the first months, review every case; when the hand-back rate drops, transition to a sampling rate. Watching only the volume of cases misses the timing of that transition.
A&A perspective
For the third shelf (decision-waiting), the point is less "keep AI away" and more "standardize how the item is routed to the owner or the outside tax advisor." A single monthly form that records whose decision is required, which lines the item relates to, and by when the decision is needed makes the waiting pile visible even when it stalls the close. The set of items whose close-speed AI should improve and the set whose waiting pile you want to make visible are separate design goals. Blurring them tempts the owner to ask a short-circuit question — "AI is in, why is this still stuck?"
Four items an owner decides before choosing an AI product
A&A perspective
First, phrase the reason for a faster close in terms of a delayed decision, not just fewer days. The close is late, so a cash-management decision slips by a week; the close is late, so hiring decisions can only happen once per quarter. Aim for a close that is faster because a specific decision moves; then work back to which shelf must shrink to produce those days. Otherwise, three days saved end up as three days without a decision attached, and no one on the operating side is any less busy.
A&A perspective
Second, complete the reshelving on at least the last two months of exceptions. Third, ask each candidate AI product to write down which shelves it can handle and what review format it can emit, not what its standard UI looks like, but whether the accountant can read the reasoning behind interpretation-shaped items. Fourth, share on a single sheet the responsibility boundaries per shelf across the monthly-close accountant, the outside tax advisor, and whoever configures the AI product. Leaving the boundaries loose lets interpretation-shaped items ping-pong between the accountant and the tax advisor and can slow the close further.
A&A perspective
When A&A takes a monthly-close design engagement, we begin with the reshelving and then agree on the per-shelf handling before any AI product decision. The existing accounting software's rule engine handles what it can; AI is limited to the second shelf; the third shelf is made visible by a standard routing form rather than automated. A related article, "Explaining return costs with product data: Wayfair's return-driver analysis," applies the same reshelving idea to a different operating domain. Read overseas cases not as targets to import, but as tools to reshelve your own exceptions.
Whether AI can handle exceptions is the wrong single question for the monthly close. Split exceptions into rule-shaped, interpretation-shaped, and decision-waiting, compare AI products shelf by shelf, and the days you save can be tied back to a specific business decision. Sources like Campfire and Anthropic's essay are useful as tools to reshelve exceptions, not as target numbers to import.
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 2026-09-19 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-09-19
AI-assisted editorial production
A&A uses AI for research, writing, translation and editorial checks. Source facts, our analysis and hypothetical examples are labeled separately.
Editorial check: 2026-09-19