Document processing: rules, OCR or a language model?

Start by checking which data your team receives in a defined structure. A parser and business rules may cover an order with agreed fields. Scanned pages require text recognition. Consider testing a language model for a specific extra step when staff need to map varied wording to a business field. Compare the methods on the same documents before choosing one.

Make the choice for each task. You might extract line items with a document model, validate product codes with rules and ask a colleague to resolve an unclear delivery request. The order team checks each step against its own incoming documents.

Choose the first method from the input

EVORI created this decision aid; it is not a product benchmark. Microsoft documents text, table and structure extraction for its layout model. That description does not establish accuracy on your documents. Check the supported formats and outputs in the documentation for the specific tool you plan to test.[1]

An order with an unresolved delivery date

Synthetic example: Page 1 of an order states “Product B-17, 12 units”. A follow-up email says “Please deliver before our summer shutdown.” The product catalogue includes B-17. The start date and year of the shutdown are missing.

The order team defines expected results before comparing methods:

A language model might classify the phrase as a delivery request. Your colleague still needs to establish the date. Do not let the model fill in that date or turn a classification into a confirmed delivery commitment. This example remains a draft order.

Measure fields and complete orders separately

Work with a person responsible for the process to select typical inputs and known exceptions. Record the verified expected value, source location and allowed treatment of missing information for each field. Set aside some documents while configuring rules and prompts, then use those documents for the subsequent check.

Illustrative calculation, not an EVORI measurement: Your test contains 20 orders with five required fields each, and your team knows all expected values. The method returns 98 of 100 fields correctly. The two incorrect fields occur in two different orders. That leaves 18 of 20 orders correct across all required fields, or 90 percent. If both errors occurred in one order, 19 of 20 orders would be correct, or 95 percent. Field accuracy would remain 98 percent. Report both rates, the error types and the small test size. An order with five correct fields still needs the business approvals defined for the process.

Also record the minutes spent reviewing and correcting each order. A method with fewer field errors may require more searching if it provides no source references. Compare that work on your test documents; this example cannot establish a general time saving.

Check confidence against your own documents

Some document tools return estimated confidence for individual fields or words. Microsoft notes that not every field receives such a score.[2] A language model's written statement of certainty has a different origin from a confidence model trained for that purpose.[3]

Choose a review threshold using documents with verified answers, then check it on documents held out from configuration. Separate document types and critical fields. Microsoft's Content Understanding guidance describes this approach and limits its transfer to other documents and model versions.[3] Do not adopt a value such as 0.95 as a blanket approval rule for your business.

A template for comparing methods

The method comparison CSV template contains the five fields from the example. Copy the five rows for each method and record its version. Add the actual value and evidence reference; note discrepancies alongside the actual value. Enter total review time once per order and method so that summing the column does not count it five times. Preserve the “synthetic example” label when you add your own cases.

For the missing delivery date, enter “unresolved, ask the sender” as the actual value. Mark the result correct if the method identifies the gap and the required question. Mark an invented date incorrect. Unresolved in the result column means your team has yet to check the output. Handling a missing value correctly still leaves the order incomplete. Record open orders alongside the error rates.

EVORI created this template from the synthetic example. It contains no customer data or measured results. The person responsible for the process uses the checks to decide which fields the team may accept and which orders must remain drafts.

For the operational workflow, read Document and order processing. Use the pilot planning checklist to scope a first project.

Sources

1. Microsoft Learn: Document layout analysis, checked 3 November 2026.
2. Microsoft Learn: Accuracy and confidence, checked 3 November 2026.
3. Microsoft Foundry: Grounding and confidence in document workflows, 14 October 2026, checked 3 November 2026. Vendor account; its customer metrics do not establish EVORI results.