Receipt and Expense Capture With Machine Learning

The shoebox at month-end

An engineering services firm has 35 field engineers. Each buys fuel, parking, parts from trade counters and the occasional hotel. Receipts live in van door pockets. At month-end, finance chases, receives a pile of faded thermal paper, keys each one, guesses at the job code and asks three people what 'SCREWFIX 0043' was for.

Receipt capture with machine learning moves the reading to the moment of purchase. The engineer photographs the receipt, the model extracts the fields, suggests a category and job, and finance reviews exceptions instead of typing.

This is a mature area. Most accounting and expense tools already include capture that works well for everyday receipts. The interesting question is where they stop being enough.

What a receipt model extracts

Thermal receipts fade, crumple and photograph badly. A clear capture at the point of purchase will beat any clever model applied to a faded scan three weeks later.

Built-in capture versus a custom layer

For a business with a handful of claimants, standard expenses and one company, the capture inside tools such as Xero, QuickBooks or a dedicated expense app is almost always the right answer. It is cheap, it integrates with bank feeds and it improves without you doing anything.

Custom work starts to make sense in specific situations:

• Job and project coding. Matching a trade counter receipt to the right job using the engineer's schedule, location and the parts bought.
• Complex policies. Per-diem rules by country, client-rechargeable expenses, grade-based hotel limits.
• Multiple entities and currencies. Group companies where the same card pays for costs belonging to different entities.
• Supplier receipts that are really invoices. Trade accounts where receipts must match purchase orders.
• Integration gaps. Job management, payroll or ERP systems that the expense tool does not talk to.

In those cases, a thin custom service that extracts, enriches and routes receipts before they reach your accounting system is often more effective than replacing the whole expense stack. Our post on AI integration with accounting software covers the connection side.

Checks that save finance real time

1. Duplicate detection. Same merchant, amount and date submitted twice, or a photo of a receipt plus a card statement line for the same purchase.
2. Card matching. Pairing each receipt with its card transaction automatically, and chasing the unmatched ones.
3. Policy checks. Alcohol on a meal claim, hotel above the limit, weekend expense without a job.
4. VAT recoverability. Flagging receipts without a valid VAT number or less-detailed receipts above the relevant threshold.
5. Category suggestions learned from how finance has coded similar receipts before.

Reading the receipt is the easy part. Knowing which job it belongs to is where the hours go.

Where receipt capture goes wrong

• Claimants photograph receipts days later, when thermal paper has already faded
• Models confidently read a handwritten tip as part of the printed total
• Foreign receipts get the wrong date order and land in the wrong period
• Auto-categorisation is accepted blindly, so a mistake repeats hundreds of times before anyone notices
• Policy rules are too strict at launch and every claim becomes an exception

The fix for most of these is process rather than model: capture at purchase, confidence thresholds, and a monthly look at how often finance corrects the suggestions.

It also helps to make capture effortless for the claimant. An engineer standing at a trade counter with a part in one hand will not open three menus. Let them forward an emailed receipt to a single address, snap a photo from a home-screen shortcut, or reply to a card transaction notification with the image. Every bit of friction removed at capture shows up later as fewer chasing emails at month-end, which is the saving finance actually notices.

Tax, records and retention

Tax authorities generally accept digital copies of receipts as records if they are complete and legible, but rules on retention periods and what makes a valid VAT receipt vary by country. Keep the original image, not just the extracted data, and keep it for the period your tax rules require. In the UK, if you are within Making Tax Digital, the digital record-keeping requirements apply to the whole chain, so check your set-up with your accountant.

Our honest starting advice

At SpiderHunts, when a business asks us to build receipt capture, the first thing we do is look at what their accounting or expense tool already offers and why it is not being used. Often the answer is adoption, not technology. When the gap is real, usually job coding or a system the tool does not connect to, the custom piece is small and pays back quickly. It fits within our finance automation work alongside invoice capture, which has its own considerations covered in the invoice capture we build.

Frequently asked questions