Before approving AI-extracted invoice data, compare each field with the original document, recalculate line amounts and totals under the document’s rules, and check for duplicates. Leave unresolved values in a review queue with a named owner. A high confidence score or a balanced total is not enough to authorise a ledger entry or payment.
This method suits a small business receiving digital or scanned invoices and looking to reduce rekeying. It focuses on checking extraction, using a simple fictional example. It does not determine applicable taxes or whether an expense is acceptable for accounting purposes.
When a plausible extraction needs a closer look
The supplier looks familiar, but its name resembles another company. A date is readable, yet its day/month order is unclear. An amount belongs to the balance due rather than the current invoice. Or the same data arrives in a second PDF.
Start with the original and the raw extraction result. Retain processing warnings and logs where available. A difference between screens does not establish that the model caused it: a conversion rule or field mapping may have changed the value afterwards.
For example, Microsoft Document Intelligence’s invoice model can return structured fields and line items. Producing that structure is a technical step. Your process must still check what those values mean for the record.
Define the pilot and its boundaries
Begin with one document family, such as straightforward supplier invoices in one currency, without deposits or credit notes. Route other documents to an identified review queue. Agree on required fields, reviewers and stop conditions before connecting the output to an ERP.
You need a complete, readable original, an up-to-date supplier directory, a register of processed invoices and a place to record corrections. Use fictional documents for initial testing. Before a pilot with real documents, confirm access, data destinations, retention and the service’s applicable terms.
Manual checks and the small local test described here require no paid service. A cloud extractor, storage and reviewer time can carry costs. Check current prices and limits for your actual configuration before a trial. This article assumes neither a price nor a productivity gain.
1. Keep the evidence beside each value
Build a review sheet with four entries for each field: text read, proposed value, location in the document and decision. For an invoice number, retain the exact text before removing spaces or other characters. Any useful transformation should remain explainable.
Tools provide different evidence. Google Document AI’s Entity reference describes extracted text, normalised values, confidence and page or text anchors. These are optional. If evidence is absent, record that it was not supplied and open the original; do not invent a location or score.
A normalised value, such as a reformatted date, can simplify integration. It does not establish that the meaning was understood correctly. Keep both the original result and the value retained after review.
2. Confirm supplier, number, date and currency
- Supplier: match the business to a known supplier record using the authorised evidence available. A similar name is not a certain match.
- Invoice number: check characters, leading zeros and possible letter/digit confusion. Do not substitute a purchase order number.
- Date: distinguish issue and due dates. A value such as
04/05/2026needs clarification when the format has not been established. - Currency: confirm a code or unambiguous indication in the record. The
$symbol alone does not always distinguish CAD from USD.
A missing field must not silently become zero, an empty string or the usual value without an explicit rule. Extracted bank details must not automatically update payment master data; that change belongs to its own verification process.
3. Recalculate without confusing invoice total and amount due
For a simple invoice, check quantity multiplied by unit price, then sum the line amounts. Compare the subtotal and stated tax amounts with the invoice total. Before applying that rule to other documents, define how discounts, charges, rounding, deposits and credits are handled. Do not fix a discrepancy by changing a number until the addition works.
Microsoft’s official invoice schema separates InvoiceTotal, AmountDue and PreviousUnpaidBalance. This distinction helps prevent a mapping error: the current invoice total is not necessarily the remaining balance to pay.
Illustrative example, with no tax rate implied: two units at 40.00 and one at 20.00 produce a subtotal of 100.00. The fictional source states a tax amount of 15.00 and a total of CAD 115.00. An extracted total of 175.00 needs review. However, a changed extraction showing three units, a 140.00 subtotal, 15.00 tax and 155.00 total balances internally while contradicting the original. Values must therefore be compared with an independent reference, not only added together.
For a programmed check, use a decimal representation and an explicit rounding rule appropriate to the case. The Python Decimal documentation explains decimal calculations and context limits. Our test constructs amounts from strings, without first converting them from binary floating-point values.
4. Search for duplicates in the right scope
Compare the identified supplier plus invoice number with retained records and the other documents in the batch. The same number from a different supplier is not automatically a duplicate. Conversely, renaming a PDF does not make it a new invoice.
The matching rule depends on how suppliers operate: numbering series, entities or years may require additional scope. Document the rule before normalising identifiers. A match flags a potential duplicate for review; it does not justify automatic deletion. Distinguish a corrected invoice from another copy too.
5. Use confidence to route review
Microsoft describes extraction confidence as an estimate between 0 and 1, which is not supplied for every field. It can help prioritise results. It does not confirm your supplier match, field mapping or payment authorisation.
Do not adopt a universal threshold simply because it looks high. Evaluate representative documents against a verified reference, then measure missed errors and cases sent for review. Microsoft’s transparency note calls for evaluation suited to the use case and appropriate oversight. Our demonstration validates no production threshold.
Start with clear states: “needs checking”, “discrepancy to resolve” and “extraction verified”. The last state means the agreed checks have been completed; it does not approve the expense. Assign each exception to a person and retain the initial value, correction, reason and decision. A new extraction should reopen the affected checks instead of silently replacing a reviewed result.
What the local demonstration actually tested
On October 27, 2026, a local Python 3.9.6 test using Decimal checked eight entirely fictional structured datasets. All eight assertions passed: two cases passed the technical checks and six were routed to review. On November 11, 2026, a new check of the same eight datasets, using a script reconstructed from the documented rules, confirmed these outcomes.
- The simple 115.00 case passed.
- A quantity of three paired with a line amount of 80.00 was flagged.
- A missing currency and an ambiguous date were flagged separately.
- An already known supplier/number pair was flagged; the same number from a different known supplier was not.
- A total of 175.00 was flagged despite a synthetic confidence score of 0.99.
- The internally balanced 155.00 case was flagged against the reference, even though its arithmetic checks passed.
We ran no OCR or AI model and uploaded no invoice. The reference was a manually specified dataset; no visual comparison with a PDF or image was tested. Scores and errors were injected; they do not measure Microsoft, Google or any other provider. The test excludes rounding, discounts, deposits, credits, previous balances, currency conversion, fraud and tax rules. It triggered no ERP entry, payment or external communication.
A checklist before connecting extraction to your workflow
- The original is complete, readable and linked to its result.
- Supplier, number, date and currency have verifiable evidence.
- Values have been compared with the document and calculations repeated under the case’s rules.
- The invoice total is distinguished from the amount due.
- The register and batch have been checked for potential duplicates.
- Each exception has an owner and a recorded decision.
- Extraction verification remains separate from accounting approval and payment.
Our user acceptance testing grid can help record these criteria. The guide to choosing rules, AI or human review helps separate fixed checks, interpretation and decisions that require approval. When checks need to connect with existing systems, our custom applications and integrations service in Quebec and across Canada can help scope the pilot and exception handling.
Technical sources rechecked on November 11, 2026. Initial demonstration on October 27; the same cases were checked again on November 11, 2026.

