AI accountant: Prepare the work, keep the judgement

An AI accountant can prepare much of the paperwork your accountant needs without taking over the decisions that belong to a qualified professional. It can collect receipts, match them to transactions, suggest expense categories and flag entries that need a closer look. The result is a more organised set of records for your accountant to review, rather than a pile of documents that still needs sorting.

The distinction matters for a small business. A receipt might look like a straightforward software expense until someone notices that it includes a personal purchase. A payment might match an invoice but belong to a different accounting period. An AI agent can prepare the evidence and point out the uncertainty. A person should decide how to handle it, and the accountant should apply their professional judgement where accounting or tax treatment is involved.

What an AI accountant can prepare before the review

The most useful work happens before anyone opens the accounts for a detailed review. An agent can gather documents from an authorised source, extract the information on each receipt and compare it with available transaction records. It can then assemble a working file that explains what appears to match and what still needs attention.

For example, imagine a month of card transactions with receipts scattered across an inbox and a shared folder. The agent can identify the merchant, date, total and currency on each receipt, then compare those details with the transaction list. When the merchant and amount agree, it can suggest a match. When the amount differs or the receipt is missing, it can leave the entry open for review.

• Match receipts to transactions using the available merchant names, dates and amounts.
• Suggest expense categories based on the description and the business's existing rules.
• Flag duplicate receipts, missing documents and transactions that have no clear supporting evidence.
• Prepare a list of unresolved questions for the bookkeeper or accountant.
• Organise the supporting documents so a reviewer can trace a suggested entry back to its source.

These are proposed matches and classifications, not proof that an expense is allowable or that a transaction has been recorded correctly. A good preparation file preserves the source document and makes uncertainty visible. It does not turn an educated guess into an approved accounting entry.

Suggested categories need evidence, not confidence

Expense categorisation is a useful example of where an AI accountant can save preparation time while still needing supervision. A monthly invoice from a familiar software provider may fit an established software subscription category. A receipt from a department store is less clear. It could represent office supplies, equipment or a personal purchase. The merchant name alone cannot settle the question.

The agent should record why it proposed a category and identify the evidence it used. If the description is vague, the receipt is incomplete or the transaction differs from previous purchases, it should ask for clarification instead of quietly choosing an answer.

This is where a short question list becomes valuable. Rather than asking the owner to inspect every transaction, the agent can bring forward only the entries that need human input. A question such as "This receipt includes both office supplies and a personal item. Which amount belongs to the business?" gives the owner a concrete decision to make.

The answer should be recorded with the relevant transaction so the reviewer can understand the decision later. Where the issue concerns tax treatment, depreciation, deductibility or another professional accounting question, the agent should leave the matter for the accountant rather than present its suggestion as settled advice.

The open questions are part of the deliverable

A preparation job is incomplete if it returns a neat spreadsheet but hides the entries that could not be resolved. The handover should make it easy to distinguish a likely match from a confirmed fact, and an ordinary missing receipt from a question that needs professional judgement.

A useful review pack might contain a transaction list, links to the available receipts, suggested categories, the reason for each uncertain match and a separate queue of questions. Each unresolved item should state what is missing and who can answer it. For example, the owner may need to explain the purpose of a purchase, while the accountant may need to decide how a particular asset should be treated.

This separation keeps the review focused. The owner answers questions about what happened in the business. The accountant interprets the records and applies the relevant rules. The agent does the clerical preparation that connects the two.

A bookkeeper template can help define this work before the first run. The Delagents bookkeeper template provides a starting point for specifying the task, its boundaries and the result the agent should return. The brief should state which records the agent may inspect, which categories it may suggest and which questions must remain open for a person.

Why filing and payments stay with a person

Preparing records and acting on them are different responsibilities. A suggested category can be corrected before it enters the books. A submitted tax return or an outgoing payment has consequences outside the preparation workspace, and reversing the action may be difficult or may require further work.

For that reason, an AI accountant should not independently submit tax returns, move money or pay an invoice just because its preparation looks complete. A person needs to review the relevant information, resolve outstanding questions and authorise the action. The same principle applies to changes in accounting records that could overwrite existing work.

Human approval works best when the reviewer can see exactly what is about to happen. For a payment, that means seeing the recipient, amount and purpose before authorising it. For a filing, it means reviewing the actual return and its supporting figures, not simply approving a message that says the task is finished. Approval should be tied to the specific action, not treated as a general permission for the agent to proceed.

There is also a limit to what approval can fix. If the underlying records are incomplete or the reviewer cannot understand how a figure was derived, the right choice is to send the work back for clarification. A deadline does not turn uncertain information into reliable information.

Set a narrow brief and a clear stopping point

Before delegating accounting preparation, decide what a finished deliverable looks like. "Sort out the accounts" is too broad. It leaves the agent to infer the period, the sources, the expected format and the decisions it is allowed to make.

A better brief might ask the agent to prepare the previous month's card transactions for accountant review. It can match receipts from an approved folder, suggest categories using the supplied chart of accounts and return a list of exceptions with source references. It must not alter the official ledger, contact suppliers, pay invoices or submit a return. Any ambiguity about the treatment of an expense belongs in the question list.

Give the task a budget cap and require the agent to stop when the requested preparation is complete or when it reaches a decision outside its authority. The review should show the work produced, the questions still open and any actions awaiting approval. An activity log can help the owner trace what the agent accessed and what decisions were approved.

Delagents is built around this pattern: a person briefs an agent, the agent prepares a deliverable and work that affects something outside the workspace waits for sign-off. Its approvals inbox presents the proposed action, the cost so far and a deadline so a person can approve, change or reject it. The workspace is in preview, and the point of the workflow is to keep preparation separate from authority to act.

Start with one month of records

Choose one completed accounting period and a manageable set of transactions. Gather the receipts and transaction export your accountant already uses, then write a brief that asks for proposed matches, suggested categories and an exception list. Keep the original records untouched. Have the agent explain uncertain matches rather than resolve them by assumption.

Before sending the pack to your accountant, check a sample of the proposed matches against the original receipts, review every flagged exception and confirm that the agent has not made changes or taken actions outside the brief. Ask your accountant whether the format contains the evidence they need and which recurring questions could be answered earlier in the process.

Use that feedback to refine the next brief. The first goal is not to automate the whole accounting process. It is to hand over a cleaner set of records, with fewer avoidable questions and a clear account of what still needs a human decision. Start with last month's transactions and make that preparation pack the agent's only deliverable.