Use OCR when you mainly need readable text from an image. Use structured extraction when you need particular values in particular fields. Add business checks when the question is whether those values agree with other records or satisfy your process. Buying a stronger reader does not automatically solve the third problem.
Start with the decision someone makes after opening the file. “Find text in an archive” and “prepare a purchase record for approval” need different outputs. Write down that decision before comparing product demonstrations.
Separate three jobs
| Job | Useful output | Question to test |
|---|---|---|
| Read | Searchable text | Can staff find the original passage? |
| Structure | Named fields and table rows | Are values assigned to the correct labels? |
| Validate | Matches, conflicts and missing evidence | Does the record meet agreed requirements? |
A product may provide several of these capabilities. Ask to see them separately so an attractive summary does not hide missing evidence. An existing document store may already provide enough OCR for search, while your accounting or operational software may own the validation rules.
A fictional purchase request
Harbour Studio receives a scanned supplier quotation, Q-218. The page says “24 display stands”, “GBP 38 each”, and “delivery excluded”. A colleague wants a draft purchasing record for project P-74. OCR might return every phrase correctly but leave the colleague to identify the quantity and price.
Structured extraction could propose quantity 24, unit price 38, currency GBP and subtotal 912. That arithmetic is straightforward once the fields are right. The draft still needs the project reference, delivery arrangement and approval owner. Those are process questions, not character-reading questions.
Suppose the supplier later sends a clear digital PDF instead. The system should use the available text rather than automatically turning it into another image. The input format is part of the design. AWS advises against unnecessary document conversion or downsampling; that is a useful reason to preserve the source file.
A short buying checklist
- List the document families actually received, including scans and digital originals.
- Name the fields the next person uses, rather than requesting every possible field.
- Specify how a reviewer reaches the exact source page.
- Identify checks owned by another system, such as an approved supplier list.
- Decide what happens when the file is unsupported or ambiguous.
- Check the existing product and plan before adding a separate service.
Keep the document library and the reader separable where practical. You should be able to replace the extraction component without losing source files, accepted corrections or the relationship between a record and its evidence.
Exceptions that change the choice
Handwriting, rotated pages, photographs of tables and missing pages deserve their own examples. A system that handles a clean sample may still return a convincing but incomplete row set from a poor scan. The correct fallback can be a clearly labelled manual task rather than a guessed result.
Also test a perfectly readable document that belongs to the wrong project. This exposes the difference between extraction quality and workflow correctness. A successful read should not authorise an update to whichever customer record happens to be open.
Run a decision-focused pilot
Select a small, varied set that the team is authorised to use. Have a reviewer prepare the expected fields and missing-information decisions before testing. Compare the proposed output with that reference, recording errors separately for reading, field assignment and business checks.
Time the whole task: opening files, checking results, correcting them and completing the handoff. Include files that require manual work. Then choose whether native configuration, a narrow integration or a custom review screen addresses the remaining difficulty.
Clear Anvil’s workflow service can be scoped around that remaining step. The useful brief is “prepare these fields for this reviewer”, with an explicit exception path, rather than an open-ended request to add AI to documents.
Sources and further reading
AI-assisted research and writing, checked against the linked sources. Worked examples are fictional. Product features can change; confirm the details in your own account.