AI can help turn invoices, order forms and other documents into usable fields. Give it a limited job: propose the information, show where it came from and pass the result through checks before another system relies on it.
Reading a total from a PDF is a different action from approving a bill or changing a supplier's bank details. Keep those decisions outside the extraction step.
Define the fields before choosing a tool
Take one document type first. For a fictional supplier-order workflow, the required fields might be the supplier reference, order number, document date, currency, line quantities and total.
For each field, decide what to do when it is absent or unclear. A blank order number should remain missing, with a reason for review. It should not become a plausible number invented to satisfy a required field.
Agree the meaning of similar values. Is the requested total before or after tax? Is a date the issue date, delivery date or payment due date? Include the currency with every amount. A neatly populated record can still contain the wrong kind of information.
Keep the evidence beside the proposal
Microsoft's document-processing integration lets a workflow use extracted fields and their confidence scores. Check the selected model's actual outputs and limits rather than assuming every field has the same evidence available.
Ask your builder to retain a link to the approved original and, where supported, the page or highlighted passage behind a field. The reviewer should be able to compare the proposed value with the document without downloading another loose copy.
Treat the score as a reason to investigate, not an approval. A critical field can need checking even when the system reports high confidence. Choose review rules from checked examples of your documents and the consequences of an error.
Add checks that don't need AI
Use ordinary validation after extraction. Your workflow can check whether:
- the supplier reference matches an existing permitted supplier.
- the order number has the expected format and belongs to that supplier.
- the currency is one the business handles.
- quantities and amounts are valid numbers.
- totals reconcile under the agreed rounding rules.
- the document has already been processed.
A matching total does not prove every line is correct. A duplicated order can also have perfect arithmetic. Keep those checks separate, and decide which failures require a person.
If the order cannot be matched confidently, create an exception rather than selecting the first similar record. Names alone are a weak basis for connecting documents to business accounts.
Try an awkward document on purpose
The fictional workflow's acceptance test could include a clear one-page order, a scan with a faint decimal point, a credit document, a missing order number and two files containing the same order.
Write the expected outcome before running the test. For example, the duplicate should point to the existing record. The missing reference should create a review item. A credit should not quietly enter the ordinary purchase flow as a new positive amount.
Use an isolated test environment and fictional or appropriately approved documents. Confirm that downstream emails, accounting entries and payment functions cannot reach live recipients or accounts during testing.
Review fields, then release the record
Keep “extracted”, “checked” and “posted” as different states. Record which fields a reviewer changed and whether those corrections reveal a recurring document layout problem.
The system should prevent an incomplete review from posting to the destination. If posting fails, retain the checked result and the destination reference so a retry does not create a second entry. When the result of an earlier attempt is uncertain, inspect the destination before trying again.
Any requested bank-detail change needs the business's separate verification procedure. Accurate reading of an instruction does not establish that the instruction is genuine or authorised.
Judge the whole process
Measure time from document arrival to a usable, checked record. Include uploading, reviewing, correcting and dealing with failed transfers. Count errors by field, because a wrong reference and a wrong amount create different problems.
A useful first integration often produces a well-prepared draft record with visible evidence. It doesn't need to make every financial decision to remove repetitive typing.