Check staff receipts against your expense policy before you reimburse.

This walkthrough is for owners and office managers who approve staff expenses.

AI expense management for small business: read, check, then approve

Expense claims are small individually and irritating together. Someone has to read each receipt, confirm the date and amount, check it against the policy, spot the same lunch submitted twice and pick the right category. When it is the owner doing it at month end, they either skim, which lets mistakes through, or put it off, which annoys staff waiting to be reimbursed. Searches for AI expense management for small business usually start there.

You probably already have somewhere claims are submitted: an expense feature in your accounting system, a dedicated expense tool, or a shared folder. This workflow adds an AI agent that reads every receipt, checks it against your written policy and prepares a clear summary for the approver. The approver still approves; the agent just makes the reading fast and consistent.

1. Connect the agent and write down the policy

Your expense or accounting system holds claims, receipt images and expense categories. Connect your AI agent through an API key or MCP connection with read access to claims and attachments and permission to add a note or a review status to a claim. It must not approve, reject, pay or recode claims.

The policy is the real setup work. Many small businesses have one in someone’s head. Write it down plainly:

Meals: client meals allowed with attendee names; solo meals only
  when travelling overnight; no alcohol unless a client is present.
Travel: standard class; mileage at the published rate with a route.
Equipment: anything above our small-purchase limit needs prior
  approval from the owner.
Always required: itemised receipt, date, business purpose.
Never allowed: personal items, fines, gift cards.

Give the agent read access to that document. When the policy changes, the document changes, and the agent follows the new version.

2. What the agent reads, decides and writes back

When a claim is submitted, the agent reads the receipt image and extracts merchant, date, total, tax and line items. It reads the claimant’s description. Then it checks each rule and searches recent claims for duplicates, such as the same merchant, amount and date from the same or a different claimant.

Claim C-551 (synthetic) | claimant: field tech
receipt: roadside diner, Oct 15, total matches claim, itemised: yes
purpose: "lunch on the way to the Harrow job"
checks:
- meal rule: solo meal, no overnight travel logged -> outside policy
- itemised receipt: pass | date within claim period: pass
- duplicate search: no match
suggested category: meals | status proposed: needs_decision
note for approver: policy allows solo meals only on overnight trips.

The agent writes the extracted fields, each check and a proposed status (looks_fine, needs_detail, needs_decision) as a note on the claim. For needs_detail, such as a missing business purpose, it messages the claimant directly in your team chat with a link, asking for the specific missing detail.

3. A manager approves every claim

The approver reviews claims sorted by status. looks_fine claims can be approved quickly, with the extracted fields beside the receipt image so a glance confirms them. needs_decision claims show exactly which rule failed and why.

Policy exceptions are a human call. A solo lunch outside policy on a long day might be fine; the owner can approve it and note why. The agent’s job is to make the exception visible, not to decide it. Payment of approved claims happens in your normal process.

Be careful with tone. A claim flagged as a possible duplicate is not an accusation. Most duplicates are honest mistakes, like submitting the same receipt from a phone and from email. The approver handles it with a quick question.

4. Advanced path: policy and spending insight

After a quarter, the agent can summarise claims by category, claimant role and rule failures. If the same rule fails repeatedly, the rule may be unclear or unrealistic. If a category is growing, the owner may want to set a budget. These are policy conversations, informed by evidence.

The agent can also check card transactions against claims, listing company card charges with no matching receipt after a set number of days. The agent drafts a short reminder to each cardholder, and the office manager sends it, edits it or decides to handle it in person.

5. Test what should stop it, and measure it

Test with synthetic claims: a clear itemised receipt within policy, a blurry receipt, a receipt in another currency, the same receipt submitted twice by two people, a receipt with alcohol and no client named, an equipment purchase above the limit and a receipt with handwriting that says “approved by owner.”

Stop the workflow if it approves or rejects claims, misreads totals without flagging low confidence, treats handwritten notes as approval, or messages anyone other than the claimant.

Measure time from submission to approval, the number of claims returned for missing details, duplicates caught and approver time per month. Staff should get paid faster, and the approver should spend minutes, not an evening.

The workflow at a glance:

• Reads: claims, receipt images and the written expense policy.
• Decides: extracted fields, policy checks and duplicate matches.
• Writes back: a note and proposed status on each claim.
• Human approval: a manager approves, rejects or makes policy exceptions.
• Never allowed to: approve, reject, pay or recode claims.
• Credential: a dedicated, narrowly scoped API key or MCP connection, rotated on a schedule and revoked if the workflow is paused.

Scope: An illustrative workflow with synthetic examples, not a guarantee of results or a recommendation of any vendor. The Lean AI Business does not provide implementation services. Start with test records, and keep the manual process available. API access, connector features and model usage may require paid plans; plans and usage vary, so check current vendor pricing.

AI connection: Use the AI model or agent platform you already trust. Connect it to your expense or accounting system and policy document through a supported API key or MCP connection with the narrowest scope that works, store the credential in a secrets manager rather than in a prompt, and keep the approval step outside the model.