Document AI: verifying extracted fields before they enter Odoo
On November 20, AWS announced general availability of Amazon Textract Insights in Northern Virginia and Oregon. The service can split a mixed packet, classify its documents and extract structured fields. Each value comes with a confidence score and a reference to its location on the page. Before automating ERP data entry, check whether that reference lets reviewers find the right supporting document quickly.
1. Distinguish extracted values from inferences
AWS lets teams define fields read directly from a page and fields inferred from its content. For a supplier invoice, treat the printed amount and the suggested expense category as two different results. Recreate cases where two amounts look alike or a reference appears on several documents. Check that the reference points to the correct document and measure errors by field before setting review thresholds.
2. What this changes for Belgian and French companies
An SME can start with one invoice type and measure correction time per case. A mid-market company or public body can build a test set with scans, multilingual files and ambiguous attachments. For workflows that require processing in Europe, check availability of the specific feature before using real files. AWS announced Textract Insights in two US Regions and describes cross-Region inference within the geography where the request originates.
3. Underside analysis — keep the evidence with the ERP draft
In Odoo Enterprise, attach the document ID, page number and extraction rule version to each proposed value. Give the agent permission to prepare the draft and reserve approval for an authorised accounting role. If the same corpus feeds a RAG system, apply document access controls to indexed excerpts. To compare a cloud service with local extraction, for example on a managed Mac, use the same synthetic files and record where originals, results and logs reside.
4. Base the decision on observed errors
Include a duplicate invoice, an unreadable total and a missing reference in your tests. Route those cases for review and check that retrying after a correction preserves the link to the source without creating a second draft. Track error rates for critical fields, review time and cost per accepted case.
Before automating data entry, define which fields need human approval and have the team that handles exceptions verify the links to source documents.
Assess a document AI workflowRead the official AWS announcement · Read the AWS technical overview