Test AI on bank reconciliation: four steps from a review list
You can trial AI on monthly bank reconciliation if it only drafts a review list and staff make the final call.
Key points
- Try one matching task: bank vs invoices
- AI drafts a review list, posts nothing
- Staff confirm each match on the list
- Decide to keep or stop by count and time
Contents
Try one matching task: bank lines against invoices
A&A perspective
If you want to try AI on your monthly accounting, start with payment reconciliation alone. Payment reconciliation means comparing the bank statement with your list of invoices to see which invoices have been paid.
Hypothetical example
For example, imagine a B2B company with about 20 employees, where the person who keeps the books spends several hours on this every month. Until the matching is done, the monthly figures the owner wants to see are not final.
From the sources
Anthropic, which makes conversational AI similar to ChatGPT, publishes accounts of companies that use its AI on its own website. One of them is Campfire, which provides accounting software to fast-growing tech companies in North America.
From the sources
According to the account, traditional accounting software had reached its limits with rule-based automation, leaving accounting teams bogged down by manual tasks. Campfire then set the order of its AI projects based on usage data and the time spent in its software.
From the sources
In a December 2024 engineering article, Anthropic recommends starting with the simplest solution possible when building AI systems. Complexity is added only when it is needed.
A&A perspective
Automate & Augment (A&A), the company writing this article, sees two conditions for the first task to try. It takes real time, and its answers can be checked against the bank statement. Payment reconciliation meets both.
A&A perspective
If you widen the trial to the whole close or to creating invoices at the same time, you cannot tell where it worked. That is why the first trial is limited to this one matching task.
A&A perspective
You prepare two files: the invoice list exported from your accounting software and the bank statement as a CSV. A CSV is data exported in a form you can open in a spreadsheet.
A&A perspective
The statement contains payer names and amounts. Before handing it to an AI tool, check that tool's terms of use for how the data you give it is handled.
AI drafts a review list and posts nothing
A&A perspective
Do not let AI put its matching results into the books; have it build a review list for the person who checks them. Each line of the list shows the candidate payment for an invoice, the reason it was chosen, and the original line of the bank statement.
From the sources
The Campfire account introduces AI features built into its accounting software. One is automated bank reconciliation, which parses bank statements and identifies discrepancies.
From the sources
The other is a chat feature that lets users ask questions about their financial data in plain language. It explains its logic and reasoning, shows its work, and provides source data. The account calls this a critical requirement for accounting professionals.
From the sources
Campfire's goal was faster monthly closes for its customers while maintaining accuracy and auditability.
A&A perspective
A&A applies the same condition to payment reconciliation at a small company. If the list shows only the AI's answers, staff cannot tell whether they are right. With the reasons and source lines beside them, staff can compare line by line and decide.
Hypothetical example
For example, one line of the list could look like this. Next to the invoice number, customer, due date and amount, put the date, payer name and amount of the candidate bank line. The last column gives the reason it was chosen, such as "amount and name match".
Hypothetical example
You will also see payments under a different name and partial payments. For example, the invoice is addressed to a company, but the payer name on the statement is an employee's personal name. Keep such lines as candidates and have the AI write in the reason column what differs.
A&A perspective
Shape the instructions you give the AI to fit this list. What you ask of the AI stops at "list the candidates and reasons." That way, even if the AI's answer is wrong, the books do not change.
| Payment line | Reason column (written by AI) | What staff do |
|---|---|---|
| A payment whose amount and name match the invoice | Amount and name match | Compare with the source bank line and, if it matches, mark it paid in the accounting software |
| A payment under an employee's personal name | Payer name differs from the customer | Contact the customer to check |
| A payment of only part of the invoice | Amount is less than the invoice | Contact the customer or wait for the next payment |
Staff confirm each match before it is marked paid
A&A perspective
Staff, not AI, confirm that an invoice is paid, after looking at the list. The list the AI builds only gives staff the order in which to check and the clues for checking.
A&A perspective
Staff compare each reason with the original bank line and mark only the lines that match as paid in the accounting software. For lines with a different name or a partial payment, staff decide whether to contact the customer or wait for the next payment.
A&A perspective
Lines where the amount and the name match only need the reason checked against the source line. Staff spend their time on the lines with a different name or a partial payment.
A&A perspective
If no one in your company keeps the books, first decide on one person who will confirm. It can be you, the owner, or the tax accountant's office that keeps your books.
A&A perspective
A&A reads a person confirming at this step as the way to keep, in your own company, the auditability the account talks about. If you add columns to the list for who confirmed and when, the list records who marked each line paid and on what reason.
A&A perspective
This account does not tell you whether the approach fits Japanese tax and bookkeeping rules. The Campfire story is about a North American company and its customers, and it does not mention Japanese rules.
A&A perspective
A&A does not infer how Japanese tax rules apply from an overseas case, and recommends checking first with your tax accountant or the company behind your accounting software. Keep the trial to the review list, before anything is marked paid.
A&A perspective
Also, this account is an introduction published on the AI provider's own site, not something a third party has verified. Whether a small company in Japan would go the same way has not been checked either.
Decide to keep or stop by pending items and hours
A&A perspective
After a one-month trial, decide whether to keep going or stop based on two numbers: how many items are still pending and how long matching took. A pending item is a line where the AI proposed a candidate but staff could not mark it paid.
From the sources
Anthropic's engineering article says the key to success is measuring performance and iterating on implementations. It also says to add complexity only when it demonstrably improves outcomes.
A&A perspective
A&A uses this "measure before you expand" idea to decide whether to keep the reconciliation trial. Two things you can count from everyday work are enough.
Hypothetical example
Here are hypothetical numbers. Say a company spent 10 hours a month on matching, and in the trial month had 3 pending items and spent 6 hours, including checking the list.
Hypothetical example
In this case, if the 4 hours saved multiplied by the staff member's hourly rate is more than the monthly cost of the AI tool, that is a reason to keep going. Staff clear the 3 pending items by contacting the customers.
Hypothetical example
On the other hand, if many lines remain pending and staff go through the statement again from scratch, the hours do not fall. Stop for that month, or first tidy up the customers who often pay under a different name, then try another month.
A&A perspective
The account also includes numbers that show results, but A&A does not use them as a benchmark for your company. It does not say which period or which comparison they come from. Your benchmark is your own numbers from the month before the trial.
A&A perspective
If your accounting software's automatic matching already clears most payments, stopping may take less effort than testing AI for the few left over.
Summary
A&A perspective
Try one matching task first: the bank statement against the invoice list.
A&A perspective
AI drafts a review list and posts nothing to the books.
A&A perspective
Staff confirm each match on the list before an invoice is marked paid.
A&A perspective
Decide to keep or stop by the pending count and the matching time.
Monthly bank reconciliation can be trialled without letting AI mistakes into the books, as long as the AI drafts a review list and staff keep the final call. After a month, compare with your own numbers from the month before the trial and decide to keep or stop.
FAQ
Do AI matching results comply with Japanese tax rules?
This article's case does not tell you. It is about a North American company and does not mention Japanese tax or bookkeeping rules. Check with your tax accountant or the company behind your accounting software.
Is AI worth trying if my accounting software already matches payments automatically?
It is worth trying if many lines are left over after automatic matching each month. Put only the remaining lines in the list and have the AI set out candidates and reasons. If most are cleared automatically, stopping may take less effort.
Can a company with no bookkeeper try this?
Yes. First decide on one person who confirms that invoices are paid. It can be you, the owner, or the tax accountant's office that keeps your books.
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
Accessed 2026-10-27 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-27