An example: capturing the text is only the start
Illustrative example with invented data
Order R-104. Please send 12 boxes of C-18 and 4 units of F-02 to our Lyon site. Delivery by 06/07. Thank you.
Scroll the table to compare all columns.
| Field | Captured value | Required check |
|---|---|---|
| Reference | R-104 | Confirm the customer and whether this order already exists. |
| Product C-18 | 12 boxes | Match the catalogue code and pack size. Do not turn boxes into individual units without an approved rule. |
| Product F-02 | 4 units | Check the product code, unit and permitted quantity. |
| Destination | Lyon site | Match a specific delivery location: a city name is not an address. |
| Requested date | 06/07 | Clarify day, month and year where the context does not make them certain. |
Keep the captured values in the draft. Flag uncertain lines for clarification, with a team member approving corrections before the order proceeds.
Define five steps
Choose the inputs
Specify which inboxes, attachment types and customers belong in the first workflow. Collect authorised samples, including incomplete messages and amendments to existing orders.
Prepare the fields
Separate customer, reference, product lines, quantities, units and destination. Keep a link back to the original message and document.
Match against approved records
Use customer records and the destination system’s catalogue. An abbreviation or customer product code must not become a new item just because it looks similar.
Give exceptions to the team
Show the source value, proposed value and reason for stopping. Decide who may correct and approve it; do not hide an exception inside a generic status.
Confirm the destination result
Check the supported import format or operation. Record the returned reference; a timeout does not prove the order is absent. Check the result before retrying.
Before you turn it on
- Every customer and delivery site has a verifiable identity.
- Codes, variants, units and packs have approved mappings.
- Missing mandatory destination fields stop the draft.
- Duplicates, forwarded messages and amendments follow distinct rules.
- Missing prices, discounts or commercial terms are never invented.
- Each failure has an owner and a clear way to resume the work.
- Testing includes a valid order, an incomplete one, a duplicate and a destination-system failure.
When orders arrive from several European markets
Message language and data format are separate questions. Add these checks to the same workflow, without copying the automation for every country.
| Detail | Rule to agree |
|---|---|
| Language and codes | Read the source language while preserving product codes and references. Translating a description does not authorise a code change. |
| Decimals and dates | Interpret commas, points and dates using a verified format. Keep the source value and stop ambiguous cases. |
| Units and packs | Distinguish units, boxes, cartons, weights and measurements. Convert only with factors approved by the team. |
| Currency | Preserve the received amount and currency code. An amount without a currency needs clarification; never apply an implicit conversion. |
| Delivery | Identify country, address and customer site separately. Do not infer destination from the email’s language. |
| Team handoff | Record the handoff date, time and time zone. Decide who owns the case and who approves an exception. |
Extraction confidence does not approve an order
For example, AI Builder can return extracted values and confidence scores. These can help direct document review; they do not replace checks against the catalogue, pack sizes and commercial rules. Evaluate drafts the team accepts and the time needed to correct them, rather than counting captured fields alone.
References: Microsoft Learn — Use the document processing model in Power Automate · Microsoft Learn — Interpret confidence score for tables and table cells
Practical questions
Does the business system need an API?
That depends on the required operation. Check supported imports, mandatory fields and confirmation behaviour first. An import file may be enough for an initial workflow; the existence of an API does not guarantee every operation.
Can we start without writing to the business system?
Yes. The first output can be a draft the team checks and imports. That lets you correct rules and mappings before authorising automated writes.