Blog · 27 September 2026 · 9 min read

Using AI for bookkeeping safely

A practical guide to using AI for receipt entry, coding and checks while keeping people responsible for evidence, approvals and donor reports.

AI can read receipt images, suggest expense categories, compare lists and draft explanations. This can reduce repetitive work when you are preparing a donor report from a bank statement, a spreadsheet and paper documents. But AI can also misread an amount, choose the wrong date or confidently fill a gap with something that is not on the document.

The safe approach is to use AI as an assistant, not as the bookkeeper, approver or final reviewer. Give it limited tasks, keep the original evidence and require a person to verify every fact that affects the accounts or donor report. This guide explains what you can delegate, what you must check and how to set up a simple review process.

Give AI narrow, reversible tasks

AI is most useful when the task has clear boundaries and its output can be checked against a source. Reading fields from a receipt is a suitable example. AI can propose the supplier, date, currency, total and receipt number, while the image remains available for review.

It can also turn information into a consistent format. For example, receipts may show dates as 3/2/25, 03 February 2025 or 2025-02-03. AI can copy the date and propose one standard spreadsheet format. A person should still confirm what the original date means, especially where day and month order is unclear.

Another useful task is suggesting categories. If a receipt says “20 exercise books” and “10 pens”, AI might suggest “training materials”. Treat this as a suggestion. The correct heading depends on the approved budget, accounting policy and purpose of the purchase, not only on the words printed on the receipt.

Ask AI to show uncertainty instead of hiding it. A useful instruction is: “Copy only what is visible. Leave unclear fields blank. Mark uncertain text for review. Do not guess missing values.” A blank cell is safer than an invented supplier or amount.

Keep each task reversible. You should be able to correct the spreadsheet without changing the original document. Never allow an automated process to overwrite receipt images, bank statements, approved budgets or prior reporting files.

• Extract visible receipt fields into draft rows.
• Standardise date and currency formats after review.
• Suggest, but do not decide, expense categories.
• Flag possible duplicate entries or missing fields.
• Draft a list of questions for the person who made the purchase.

Keep judgement, approval and payment with people

AI should not decide whether an expense is eligible under a grant. Eligibility may depend on the agreement, budget period, procurement rules, activity plan, location, tax treatment and written donor approval. These facts may not appear on the receipt.

It should not approve transactions or payments. Approval shows that an authorised person accepted the purpose, amount and supporting evidence. Allowing a tool to generate or apply an approval weakens that control. The person who incurred an expense should also not use AI to create the appearance of independent review.

Do not let AI invent explanations for unclear transactions. Suppose a bank line says “Transfer 0148, 6,000” and there is no clear supporting document. AI might suggest that it was venue hire because another venue payment had a similar value. That is not evidence. Record the transaction as unresolved until a person identifies it and provides the relevant support.

AI should not create receipt details from context. If a faded receipt clearly shows a total of 1,800 but the tax amount is unreadable, do not ask the tool to calculate and present a tax figure as though it appeared on the receipt. Any calculated field must be labelled as a calculation, with its method visible.

Final decisions belong to named people. Your process should identify who entered or reviewed the transaction, who approved it and who signed off the report. AI can assist those people, but it cannot accept responsibility for the result.

• Do not let AI authorise payments or expenses.
• Do not use it to sign approval forms or imitate signatures.
• Do not accept a category only because the tool sounds confident.
• Do not ask it to fill gaps with likely values.
• Do not submit an AI-produced report without human review.

Verify every field that affects the report

Start with the original source. For each proposed transaction row, compare the entry with the receipt, invoice, bank statement, cash record or other evidence. Check the supplier, transaction date, amount, currency and reference. If a field is not visible, mark it as missing or unclear.

Pay particular attention to digits that are easy to confuse. A tool may read 1,750 as 7,750, or mistake a handwritten 4 for a 9. Currency symbols and abbreviations can also be missed. An amount may be correct but assigned to the wrong currency, which can materially change the report.

Then verify the accounting meaning. Confirm that the category matches the approved budget heading and the actual purpose. Check whether the date falls inside the reporting and grant periods. If foreign currency is involved, confirm the exchange rate source and calculation required by your organisation or donor.

For example, imagine an AI-generated row showing: “12 June, Bright Office, stationery, 4,500.” The receipt actually says “Bright Offset” and includes 3,000 for participant manuals plus 1,500 for delivery. The total is right, but the supplier and description are wrong. Depending on the budget, the correct category may be training materials rather than stationery. A total-only check would miss both errors.

Check relationships between records as well as individual fields. A valid-looking receipt may belong to a different payment. Confirm that dates, references and amounts make sense together. If one payment covers several documents, or one document was paid in parts, record that relationship clearly rather than forcing a false one-to-one match.

Finish with control totals. Add the verified transaction rows and compare them with the relevant report totals. Review any difference rather than inserting an adjustment simply to make the figures agree. A balanced total is useful, but it does not prove that every entry is valid or correctly classified.

• Compare extracted values with the visible source.
• Confirm the currency, not just the number.
• Review category, purpose and grant period separately.
• Check calculations with spreadsheet formulas or a calculator.
• Investigate differences and unsupported adjustments.

Protect financial and personal information

Receipts and bank records can contain names, phone numbers, account details, addresses, signatures and payment references. Before uploading documents to an AI service, find out where the data is stored, who can access it, how long it is retained and whether it may be used to train models.

Use only tools approved by your organisation. A free personal account may not provide the controls required by your donor, local law or internal policy. If nobody has assessed a service, do not assume it is safe because it is well known.

Share the minimum information needed for the task. If you only need help choosing a category, you may be able to provide a short transaction description without uploading the full bank statement. Redact unrelated account numbers, balances, identity details and signatures where appropriate. Keep enough information in your controlled records to support the transaction.

Do not paste passwords, one-time codes, private keys or online banking credentials into an AI chat. Do not ask a tool to log in to a bank account unless your organisation has deliberately approved a secure integration with proper access controls. Bookkeeping assistance does not require giving a general chatbot control of project funds.

Limit access inside the organisation too. Store source documents and reviewed files in an approved location. Use individual accounts where possible, remove access when roles change and avoid sending complete financial files through informal messaging groups.

If records must remain in a particular country or region, confirm that the service can meet that requirement before use. When the answer is uncertain, work with redacted extracts or keep the task offline.

• Use an organisation-approved service and account.
• Upload only information needed for the task.
• Remove unrelated personal and banking details.
• Never share passwords or authentication codes.
• Keep original files in your controlled storage.

Build a simple human review trail

A safe workflow separates extraction, review and approval. One person may prepare draft rows with AI assistance. A reviewer then checks the rows against source documents and records corrections. An authorised person approves the completed report according to the organisation’s normal process.

Add simple status columns to the working sheet, such as “AI draft”, “checked against source”, “query open” and “approved”. Also record the reviewer’s initials and review date. These labels make it clear which entries are ready and which still need attention.

Keep the AI output as a draft, not as the master record. Save the verified version separately or use version history. If a value changes, record the reason where it would help another reviewer understand the correction. For example: “Amount corrected from 8,600 to 3,600 after checking receipt image.”

Use targeted review questions. Ask whether the source is readable, whether all required fields are present, whether the purpose is supported and whether the category follows the grant rules. A general instruction to “check everything” is less useful than a short list of defined checks.

Consider a small trial before using a new tool for a full reporting period. Test it with copies of varied documents, including handwritten receipts, faint printing, several currencies and multi-line invoices. Note which fields it handles poorly. The purpose is not to prove that the tool is accurate in general, but to learn where your own review must be strongest.

Document your local rules in a short procedure. State which tools are permitted, what information may be uploaded, which tasks may be automated and who performs each check. Include what staff should do when the document is unclear or the AI suggestion conflicts with the budget.

For a practical example, suppose there are 40 receipt images. AI prepares 40 draft rows. The preparer checks that every image has a row and marks seven uncertain fields. A reviewer opens each image, corrects two amounts, changes three categories and leaves two transactions as open queries. Only the 38 resolved rows move into the report. The remaining two wait for evidence or a documented decision by the responsible person.

This process may feel slower than accepting the first output, but it preserves the distinction between assistance and evidence. The source document supports the transaction, the human review supports the data entry, and the authorised person remains responsible for the report.

• Label all AI-produced entries as drafts.
• Record who checked each row and when.
• Keep unresolved items out of the approved total.
• Save corrections and reasons where useful.
• Review the process when the tool or reporting rules change.

Questions

Use AI for limited, checkable work such as draft extraction, formatting, category suggestions and exception lists. Keep people responsible for evidence, coding, approvals, privacy and the final report. Honest Record can automate matching and checks, but its results should still be reviewed against the source documents and the grant rules.

More guides

• Add budget versus actual without confusion
• What donors check in a small NGO financial report
• Preparing grant records for a donor audit
• How to report a payment when the receipt is missing