Short answer
AI supplier document extraction is a workflow where a model reads supplier quality and compliance documents, such as mill certificates, certificates of analysis, certificates of conformance and safety data sheets, pulls out the values that matter, checks them against your purchase order and specification, and files them against the receiving lot. In New Zealand it also helps keep safety data sheets current under WorkSafe rules. Quality staff release the goods.
- Extract to a fixed schema; a missing value stays empty, never guessed.
- Check values against the PO and spec, and documents against each other.
- A changed safety data sheet revision should trigger an inventory update and worker briefing.
- Extraction prices change: check dated model pricing before budgeting.
What gets extracted, and why it is not invoice capture
AI supplier document extraction reads the quality and compliance paperwork that arrives with materials and turns it into checked records. Accounts payable tools already capture supplier, amounts and due dates from invoices, and inventory systems are adding it too: Cin7 lists document recognition for invoices and purchase orders. Certificates are different. The values are heat numbers, chemistry, mechanical properties, batch results and hazard classes, and the question is whether the material meets the specification, not whether the price is right.
| Document | Fields to extract | Check that catches real problems |
|---|---|---|
| Mill test certificate | Heat number, grade, standard, yield and tensile strength, chemistry | Values within spec limits; heat number matches the packing list and the stamp on the material |
| Certificate of analysis | Batch, test methods, results, expiry | Each result inside the agreed range; batch matches the delivery |
| Certificate of conformance | Product, standard claimed, signatory, date | Standard matches the one on the purchase order |
| Safety data sheet | Product name, supplier, revision date, hazard classification, emergency contact | Revision newer than the one on file triggers an update |
How the workflow runs
- Inputs
- Certificates and data sheets from supplier emails, portals and delivery paperwork; purchase orders; specifications with numeric limits; the approved supplier list; the hazardous substance inventory.
- Systems
- The ERP or inventory system for POs and receiving lots, a document store, the quality register, and a vision-capable model API.
- Steps
- Identify document type; extract into a fixed schema with page references; validate units and ranges; compare against the PO and specification; cross-check documents in the same delivery; file against the lot; raise exceptions.
- Human review
- A QA inspector reviews every exception and a daily sample of clean extractions, then releases or holds the lot. An engineer decides concessions. The health and safety lead reviews every new or revised safety data sheet.
- Outputs
- Lot records with extracted values and source page images, an exceptions queue with the failed check stated, an updated certificate register with expiry dates, and an updated substance inventory.
- Measures
- Field-level extraction accuracy on a fixed test set, exceptions per supplier, minutes per document, and lots released without a complete certificate pack.
- Risks
- Misread decimals and units, guessed values, mismatched heat numbers, and altered or reused certificates.
The NZ rule for safety data sheets
WorkSafe’s guidance on safety data sheets says a business supplied with a hazardous substance must obtain a sheet when the substance is first supplied, including the first time in five years, and when it is first supplied after the sheet is amended. A current sheet, or a condensed version, must be kept with your inventory and be readily accessible to workers and emergency services, and workers need to be made aware when the sheet changes. Extraction helps by comparing each incoming revision date with the one on file and opening a task when it changes. It does not replace the person who reads the new hazards and briefs the team.
Illustrative example: a Christchurch steel fabricator
Illustrative example, synthetic numbers: a fabricator buying from 85 suppliers receives about 300 mill certificates a month, plus around 40 safety data sheet revisions a year. Assumptions: keying a certificate into the register takes 6 minutes; reviewing an extraction with the page image alongside takes 1.5 minutes; 10% of certificates raise an exception needing 10 more minutes.
- Before: 300 x 6 minutes = 30 hours a month of keying.
- After: 300 x 1.5 minutes (7.5 hours) + 30 exceptions x 10 minutes (5 hours) = 12.5 hours.
- Model fees: two pages per certificate at roughly 6,000 input and 800 output tokens, on Gemini 3.8 Flash at US$0.75 and US$3.75 per million tokens, is about US$2.25 a month. Google’s pricing page (updated 16 October 2026) says those rates double from 1 January 2027, to about US$4.50. Budget from the later figure.
The exceptions are where the value sits. A heat number on a certificate that matches nothing on the packing list is found at receiving, not after the steel is cut and welded.
What stays with a person
Releasing or holding a lot, accepting a concession, approving a new supplier and deciding whether a certificate needs independent verification all stay with qualified quality staff. The model has no permission to change lot status in the ERP.
Where extraction goes wrong
- Guessing to fill a field. If yield strength is missing, the model must return empty, not a typical value. Enforce a schema that allows nulls and reject invented values; see structured outputs and validation.
- Units and decimals. MPa versus ksi, commas as decimal separators, and values split across lines. Validate ranges per field.
- Stamps and scans. A receiving stamp over a printed value produces a confident misread. Show the cropped source image with each value; the trade-offs are in OCR versus vision models.
- Reused certificates. Identical values on two certificates for different heats deserve a question. Compare each new certificate against history.
Extraction output feeds quality inspection report triage when a lot fails, and the same document hygiene supports maintenance knowledge search. Our AI document processing service builds and runs workflows like this; the manufacturing overview shows the wider picture.
Sources
- Safety data sheets, WorkSafe New Zealand, accessed 19 October 2026
- Cin7 AI features, Cin7, accessed 19 October 2026
- Gemini API pricing, Google, updated 16 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.