AI Expense Management: Receipts, Mileage, and Approvals

AI handles expenses in three places. Receipt apps read photos and emailed receipts and match them to card transactions. Mileage apps detect drives from the phone's location and let the driver mark each one business or personal. Approval rules pass small, in-policy claims automatically and send the rest to a manager. You still write the policy and review the exceptions.

For most small teams, the biggest gain isn't faster approval. It's matching every company card payment to a receipt without chasing people at month end, and replacing mileage claims reconstructed from memory with logs recorded at the time. That's where money leaks quietly, and where AI is strongest.

Follow me on Instagram@sagnikteaches

From crumpled receipt to coded expense

When someone photographs a receipt, the app runs optical character recognition (reading the text in an image) and then an AI model that understands receipt layouts. It extracts the merchant, date, total, tax, currency and payment method, and often the individual items. It then looks for a card transaction with a matching amount and date, suggests a category, and checks the result against your policy.

Connect on LinkedInSagnik Bhattacharya

Here's an illustrative result for a fuel receipt before anyone checks it:

Subscribe on YouTube@codingliquids

That last row is the useful one. The receipt was read correctly, but the claimant has a fuel card, so a fuel claim is either a mistake or a double claim. The AI didn't need to be clever to catch it; the policy rule did the work.

Receipt reading goes wrong in predictable places:

• Tips and handwritten totals. A restaurant receipt with a printed subtotal and a handwritten tip often extracts the subtotal, so the claim doesn't match the card payment.
• Foreign currency. The receipt shows the local amount; the card statement shows your currency plus a fee. The match only works if the app converts at the card's actual rate, which it gets from the card transaction, not the receipt.
• Summary card slips. A card terminal slip shows the total but not what was bought. Some tax rules and most policies want the itemised receipt too.
• Two receipts in one photo. The app reads one and ignores the other, or merges them. One receipt per photo.

Card-first or claim-first: pick the model that suits your team

There are two ways to run expenses, and the choice decides which tools make sense.

Tools fall into three groups. Your accounting software may have an expenses feature already, which is the first thing to check. Dedicated expense apps such as Expensify handle claims, receipts and approvals; Expensify priced its Collect plan at a flat $5 per member a month when it launched in 2025, so confirm the current figure on its site. Spend-management card providers combine company cards with receipt matching and policy controls, and which ones operate depends on where your business banks. Whichever you choose, check that it syncs categories and receipts to your ledger rather than just exporting a spreadsheet.

Mileage: automatic tracking and the classification problem

Mileage apps use the phone's motion sensors and location to detect when a drive starts and ends, record the route and distance, and ask the driver to classify it. MileIQ, for example, has a free plan with 40 drives a month and an unlimited plan at $11.66 a month billed annually or $13.99 month to month; you swipe each drive left or right to mark it personal or business.

Tracking is the easy part. Classification is where claims go wrong, for three reasons:

1. Commuting. In most tax systems, the journey from home to a regular workplace is private, even if the driver stops at a property on the way. Apps can't know which workplace counts as regular, so agree the rule with your accountant and write it into the policy.
2. Mixed trips. A negotiator drives from a viewing to the supermarket to another viewing. The app records one drive or three depending on the stop length. Drivers need to split or trim these, which is a one-tap job at the time and a guessing game a month later.
3. Backlogs. Classification left for four weeks turns into swiping 120 drives at once, and people stop looking. Weekly is the minimum; daily is better.

An illustrative week for one negotiator, as the app presents it before submission:

Mon  08:05  Home -> Office                  6.1 mi   Personal (commute)
Mon  10:30  Office -> 14 Elm Court viewing   4.3 mi   Business
Mon  11:20  Elm Court -> Office              4.5 mi   Business
Tue  09:10  Home -> 7 Mill Lane valuation    9.8 mi   Business (not via office: check rule)
Tue  10:40  Mill Lane -> Office              7.2 mi   Business
Wed  13:00  Office -> Supermarket            1.2 mi   Personal
...
Week total business: 86.4 mi   Personal: 31.0 mi

The Tuesday line is the kind of trip your accountant's rule should settle once: going straight from home to a valuation can count as business in some systems and not in others. Apply whatever per-mile rate your accountant confirms; the app just multiplies.

Turning your expense policy into rules an app can apply

Approval automation only works if the policy is specific enough for software. "Reasonable meal costs" can't be automated; "up to $30 per person for a client meal, itemised receipt required" can. A filled-in rule set for a small team:

Two design choices matter. Keep auto-approval limits low at first and raise them once you trust the data. And give every rule a "goes to manager" condition, so nothing falls through as neither approved nor flagged.

An estate agency with nine negotiators: one month of expenses

The agency in this illustration has nine negotiators and two valuers who all drive their own cars to viewings and valuations, plus four company cards used for parking, key cutting, signboard fitting and client coffees.

Before. Each person reconstructed mileage from their diary at month end, which took about 45 minutes each: 11 people, a little over 8 hours of staff time. The office manager checked every claim at about 20 minutes each, nearly 4 hours. Around a third of company card payments had no receipt by month end, and chasing them took another 2 hours.

Setup. Everyone installed a mileage app on the unlimited plan (11 × $11.66, about $128 a month billed annually). Card payments flowed into the ledger's expenses feature, which sent the cardholder a reminder to photograph the receipt within 24 hours. The policy table above went in as rules, with the commute rule agreed with the accountant: first trip from home counts only when it goes straight to a property, not via the office.

After the first month. Staff time on mileage fell to about 10 minutes a person a month for classification. The office manager's review shrank to around an hour, mostly on the eight claims the rules sent to her. Missing receipts dropped from a third to a handful. The tracked business mileage came in noticeably lower than the previous months' diary-based claims, because hand-written logs had rounded every trip up and included some commutes. At an illustrative $0.55 a mile, even a few hundred miles a month is a real amount, which the agency hadn't expected to be the largest saving.

The complaint. Two negotiators found the app drained their battery and switched location off, so a week of drives went unrecorded. The agency now accepts manual entries for those gaps, but they always go to the manager for approval. That keeps the exception visible without punishing anyone for a phone setting.

A tour operator's guides: cash, currencies and daily allowances

Expenses look different for an illustrative tour operator whose guides lead multi-day group trips. Guides pay entrance tickets, group meals and drivers' tips, often in cash, in other currencies, with receipts in other languages.

AI helps in two ways here. Receipt reading copes well with other languages and currencies, and a photo taken at the end of each day is readable even when the guide can't read the receipt. And a daily allowance for the guide's own meals, instead of receipts, removes most of the small claims entirely. How a daily allowance is treated for tax differs by country, so that's one for the accountant.

The rules need to be different too. Cash claims can't be matched to a card, so the operator issues a cash float per trip and the app reconciles the claims against the float: float issued, receipts claimed, cash returned. Anything that doesn't balance within a small tolerance goes to the operations manager. Currency conversion for cash uses the rate at which the float was bought, which the operator enters once per trip.

Where expense AI gets fooled

• The same receipt twice. A photo in week one and the emailed version in week three. Good apps flag duplicates by merchant, amount and date; check that yours does before trusting it.
• Split bills. Two colleagues each claim half of one dinner, each with a photo of the same receipt. Detection needs to compare across claimants, not just within one person's claims.
• Personal items on a business receipt. A hardware store receipt with a work item and a garden hose. Itemised extraction shows it; header-only extraction doesn't.
• Fabricated receipts. Expense-audit vendors including AppZen have reported a sharp rise since 2025 in receipts generated by AI image tools, realistic down to creases and handwritten tips. Visual checks won't catch them. The defence is matching: a claim-first receipt with no card trail and an unusual merchant deserves a question.
• Rubber-stamping. A manager who approves everything the app sends them in one batch has turned the rules into decoration. Look at how long approvals take; ten claims approved in 20 seconds weren't read.

Using a chat assistant to review a month of claims

Even with rules in place, a second look at the whole month catches patterns no single rule sees: one person's parking every Saturday, a merchant that appears for only one claimant, mileage that jumps in the last week of the month. Export the month's expenses as a spreadsheet, replace names with staff codes, and use a business plan that doesn't train on your data.

Attached is last month's expense export (columns: staff_code, date, merchant,
category, amount, payment_method, receipt_attached, miles).
List anything worth a manager's question, grouped as:
1. Possible duplicates (same merchant and amount within 14 days, any claimant)
2. Claims outside working days or hours
3. Merchants that appear for only one person
4. Mileage totals that differ by more than 30% from that person's usual month
For each, give the rows involved. Don't accuse anyone; describe the pattern.

An illustrative reply might list two items: "S04 and S07 both claimed $86.40 at the same restaurant on 11 Sep" and "S09's mileage is 740 miles against a usual 420-480; 12 of the extra trips are to the same address". The first was a shared dinner claimed twice by accident. The second turned out to be genuine: S09 was covering a colleague's area for two weeks. That's the right outcome. The review raises questions; the manager answers them with context the data doesn't have. Check any totals the assistant quotes against the export before acting, because arithmetic on long exports is where chat assistants slip.

A one-page expense policy you can adapt

EXPENSES POLICY (one page)

1. Who can spend: cardholders listed by the office manager. Others claim back.
2. Receipts: photograph within 24 hours, one receipt per photo, itemised where
   available. Card slips alone are not enough for meals.
3. Mileage: tracked in the app, classified within 7 days. Home to office is
   personal. Home straight to a property counts as business.
   Rate: as confirmed by our accountant each year.
4. Limits: parking up to $25; client meals up to $30 a head, no alcohol,
   attendees named; supplies up to $50 (above that, raise a purchase order).
5. Travel and accommodation: pre-approved only.
6. Approval: in-policy claims approve automatically. Everything else goes to
   the office manager within 3 working days.
7. Reimbursement: with the next payroll after approval.
8. Mistakes: tell the office manager. Deliberate false claims are a
   disciplinary matter.

Keep it to one page. Longer policies don't get read, and anything the app can't turn into a rule should probably be a manager's decision anyway.

Switching over without losing a month

The changeover is where claims go missing, so run it as a short project rather than an announcement.

1. Pick a start date at the beginning of a month. Old claims up to that date go through the old process; everything after goes through the app. Mixing the two in one month is how things get paid twice.
2. Set up with two volunteers first. A week of real use shows which rules are too tight and whether the phone app works on everyone's handsets.
3. Spend 15 minutes with each person. Install the app, photograph one receipt, classify one drive. People who have done it once in front of someone keep doing it.
4. Keep the old spreadsheet read-only for a quarter. You'll need it for the duplicate check when someone submits an old receipt through the new app.

For the estate agency above, the running cost of about $128 a month for mileage tracking plus the ledger's own expenses feature compares with roughly ten hours of staff and manager time saved each month, before counting the lower mileage bill. The sum works for most teams of five or more who drive for work; for two people with occasional claims, the free tiers and a monthly check are usually enough.

Checking it's working after the first month

Four numbers tell you whether the system is doing its job:

1. Receipts matched within seven days, as a share of card payments. Aim for nearly all of them.
2. Claims sent to a manager, as a share of all claims. If it's over about a fifth, the rules are too tight or the policy is unclear; if it's almost none, the limits may be too generous.
3. Time from submission to reimbursement. Staff notice this more than anything else.
4. Corrections after sync. How many expenses the bookkeeper recoded in the ledger. A steady number means the category mapping needs fixing at source.

If you'd like the approval side written up properly first, setting spending rules and approvals for business purchases has the template. Expenses are one of several jobs worth automating early; the finance tasks AI can automate in a small business puts them in order, and for vehicles on the company's books rather than staff cars, fleet admin for small businesses covers fuel cards and servicing.

Expense questions that come up once the app is live

Do I still need to keep the paper receipts?

Often not, but it depends on the rules where you're taxed and on your accountant's advice. Many tax authorities accept a clear digital image that shows the supplier, date, amount and what was bought. Ask your accountant what they need, check the image quality settings in your app, and keep originals for large purchases until you've had that answer in writing.

Can the app decide whether an expense is allowable for tax?

No. It can apply your company policy, such as a meal limit or a ban on alcohol, and suggest a category. Whether something is deductible for tax, or counts as a taxable benefit for the employee, depends on local rules and the circumstances. Agree the categories with your accountant once, and let the app apply that mapping.

What if staff don't want a tracking app on their own phone?

Mileage apps only need location while a drive is detected, and most let the driver pause tracking outside working hours. Explain what's collected and who sees it, put that in writing, and offer an alternative such as a manual trip log for anyone who declines. Forcing it on personal phones without that conversation usually ends in poor data rather than better claims.

Further reads

• Dext vs Hubdoc: Which Receipt Capture Tool Saves More Time? — Two receipt capture tools compared on time saved.
• AI Invoice Processing: Stop Typing Supplier Bills by Hand — The same capture approach for supplier bills.
• AI Receipt Capture and Bank Categorisation: A Bookkeeper's Setup — A bookkeeper's receipt capture and bank rules setup.
• How to Categorise Transactions With AI and Check Its Work — Check the categories the expense app pushes to the ledger.
• AI Payroll Checks: Catch Timesheet Errors Before Payday — Catch pay errors, including reimbursements, before payday.
• Can AI Do My Bookkeeping? What Still Needs an Accountant — Where expenses fit in the wider bookkeeping split.
• AI Bookkeeping Software vs a Human Bookkeeper: Accuracy and Cost — Which transactions software gets right and which need a person, three cost scenarios, and a 30-transaction audit to measure accuracy in your own books.
• How to Speed Up Month-End Close in a Small Business With AI — A day-by-day close calendar, four AI prompts with sample outputs, and a farm shop that went from a nine-day close to four without skipping checks.
• How to Track Profit per Project in a Service Business With AI — Build a weekly project margin review that includes your time, unpaid supplier bills and the cost of finishing each job.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: MileIQ pricing page; Expensify's announcement of Collect plan pricing; AppZen and industry reporting on AI-generated receipts (2025-2026). Checked September 2026.