At month end, it is tempting to drop invoices, receipts, and supplier bills into AI and ask for an expense spreadsheet. That is also where risk starts: a decimal is missed, a date is read incorrectly, tax is mistaken for the total, or the same e-receipt is uploaded twice. AI can accelerate organisation, but it should not decide whether a charge is correct, payable, or compliant with a company policy.
A safer model uses AI OCR for entry and routing: extract fields, then send conflicting, unclear, or high-value documents to a human review queue. This workflow suits freelancers, small teams, and operations groups handling purchases, travel, or supplier documents. Use ChatGPT or Gemini only with redacted and approved files; personal identifiers, banking information, and sensitive contracts require tools and storage approved by your organisation.
Start with field rules, not text recognition
One bill may show an order date, issue date, payment date, and due date; a pre-tax amount, tax, discount, shipping, and grand total. Without defined columns, a model will select the most plausible number and it becomes difficult to audit later.
Create a common header with an exact meaning for each field:
- document_id: source filename plus a sequence; do not use a supplier name as the unique identifier.
- supplier / document_number: supplier and invoice number; leave blank and flag when unreadable.
- issue_date / due_date: record only the stated issue and due dates, not a mixed-in payment date.
- currency / subtotal / tax / total: retain currency with the total; never infer unclear tax.
- expense_category: choose only from a defined list; use “pending classification” when uncertain.
- evidence: page number or source excerpt for quick review.
- confidence / review_reason: record why a human should look, rather than letting the model merely self-score.
Keep filename, upload date, and original-file link in a controlled location. The table is an index, not the only evidence store.
Deduplicate and classify before extraction
Duplicate booking is often more common than an OCR typo. An e-invoice, email attachment, and phone photo may represent one expense. Calculate a file hash before upload; for different files with similar content, create candidate duplicate groups from supplier, number, currency, total, and date. AI may flag similarity but must not delete records automatically.
Also separate invoices, receipts, statements, quotes, order confirmations, and payment proof. They establish different facts. A quote is not evidence of an amount payable, and a payment screenshot does not necessarily contain tax information. Classify first so later fields and validation rules do not blur together.
Using these document images or OCR text, output JSON following the specified field sheet. Attach evidence (page number and source excerpt) to every field. Populate a value only when it is explicit in the source; use null and explain review_reason when it is unreadable, ambiguous, or inapplicable. Do not calculate tax, exchange rates, payment status, reimbursement eligibility, or accounting treatment. First classify the file as invoice, receipt, statement, quote, order_confirmation, payment_proof, or other.
Use explicit exceptions instead of trusting confidence
A confidence score of 0.92 does not make a transaction safe to book. Reproducible exception rules are more useful: total does not equal subtotal plus tax and shipping, allowing documented discounts; missing currency; blank number; implausible date; the same supplier and total recurring in a short period; a quote classified as an invoice; an amount over the review threshold; a blurry, cropped, or incomplete multi-page file.
These rules focus human time on genuine risk. Complete low-risk records can enter a pending-confirmation queue; exceptions need a reason, owner, and outcome. Do not overwrite original facts merely because a model supplied an answer.
What human review should check
Reviewers need not type the whole document again. Open the original and check only highlighted fields and business plausibility: supplier existence, number match, whether the amount is truly a total, whether the tax rate is stated, business purpose, payment state, and duplication. For foreign currency, retain the original amount and exchange-rate source; do not let AI guess a rate from a current web page.
Three outcomes are enough: confirmed, needs supporting material, or rejected/excluded. Record a brief reason for each. When a supplier asks later, a reimbursement is returned, or reconciliation does not balance, the team can return to the original and its review record—not ask what AI was thinking.
Protect privacy and permissions
Documents often contain names, addresses, phone numbers, tax IDs, bank fragments, and internal project names. Redact fields not needed for a trial sample; give shared sheets minimum access; do not paste originals into public chat, shared prompt libraries, or unapproved third-party services. If policy requires local processing or data residency, use the organisation’s permitted OCR and document-management workflow.
For payment approval, tax filings, and legal retention, AI-created structured output is supporting material only. Retention periods, tax treatment, and approval requirements belong to finance or qualified professionals.
A weekly operating rhythm
- Collect through one intake point with a naming rule and original-file location.
- Preprocess: hash deduplication, document classification, rotation, and clarity checks.
- Extract with a fixed JSON schema and retained evidence.
- Validate totals, dates, identifiers, currencies, and duplicate rules.
- Review only exceptions, high-value documents, and poor-quality files.
- Archive outcomes, link originals, and reconfirm before export to accounting or expenses software.
The best AI OCR result is not a fully unattended expense ledger. It is a worksheet where every row has a source, every exception has an explanation, and people inspect only the important items. Define fields, exceptions, and access rules first; automation then removes repetition without moving mistakes into the books faster.