How to verify a supplier certificate of analysis before material use

Three drums arrive against a valid purchase order. The supplier's certificate of analysis is clear, complete-looking and easy for the document system to read. There is one problem: the drums are labelled lot A, while the certificate names lot B.

An extraction tool can perform perfectly and still support the wrong decision. It may capture every field from the PDF accurately, but the certificate does not yet establish anything about the lot received.

Certificate-of-analysis verification should therefore connect the document to the physical receipt and the approved material requirements. Extraction is one step. Identity checks, technical review, required testing and authorised disposition remain separate steps with separate evidence.

Decide what the certificate is expected to support

A certificate of analysis, or COA, reports information about a specified material batch. Its role in the receiving and testing process depends on the applicable requirements, material and approved quality procedure.

For APIs and intermediates, ICH Q7 section 11.4 addresses certificate content, including identification, tests, acceptance limits and results. Section 7 addresses incoming-material controls. This does not mean every supplier certificate permits a purchaser to skip its own testing.

Before designing automation, QC should answer a concrete question: what decision can this certificate support for this material and source? Possibilities include confirming document completeness, supporting a defined supplier-assurance process, or supplying information for comparison with in-house testing. The software should implement that approved role.

Do not let an application's “COA verified” label combine document authenticity, field extraction, specification comparison and material release into one unexplained status. A reviewer needs to see which checks passed, which remain open and which are outside the application's scope.

Use three comparisons rather than one document score

The first comparison is between the certificate and the physical receipt. Check the material name or controlled alias, supplier and manufacturer identifiers, batch reference, and the quantity or container information relevant to the process. Investigate discrepancies before treating a similar-looking identifier as an exact match.

The second comparison is between the certificate and the approved specification. A supplier's limits may differ from the purchaser's requirements. A document containing the word “complies” does not resolve a difference in test method, unit, basis or specification version.

The third comparison is between the evidence and the approved acceptance workflow. Which tests or reviews are required? Who is authorised to decide disposition? Are there open deviations or source restrictions? A complete certificate cannot answer those questions on its own.

This separation also makes measurement useful. You can distinguish an extraction error from a supplier-document discrepancy or an unfinished laboratory task. Combining them into one “automation failure” rate conceals the work needed to improve the process.

Follow the lot mismatch through the workflow

In the illustrative three-drum example, the receiving record should retain the observed lot A identifiers and the original lot B certificate. The system flags that the supplied evidence does not match. It should not change A to B, strip a character from the identifier or choose the closest match to make the transaction pass.

The supplier provides another certificate naming lot A. Retain the first document and the communication explaining the replacement, subject to the approved record procedure. Link the corrected document to the same discrepancy rather than creating an unrelated receipt that hides the sequence.

QC then evaluates the replacement evidence. A matching batch number resolves one problem; it does not establish that the source, specification, required testing and material disposition are all acceptable. The review screen should make those remaining tasks visible.

Suppose one of the three drums actually carries a different lot number. That is a physical identity issue, not merely a PDF problem. The affected scope and any sampling or investigation must follow the approved procedure. Software can maintain the relationships and evidence, but it cannot infer a scientifically valid grouping from the purchase-order quantity.

Design the extraction record for review

For each extracted field, retain the original value, proposed normalised value and location in the document. A reviewer should be able to move directly from a result to the supporting page or table cell.

Useful fields include:

• Document identifier, received time and source channel.
• Supplier, original manufacturer and manufacturing site where available.
• Material identity, grade and batch number.
• Specification and method references.
• Test name, reported result, unit, basis and acceptance limit.
• Approval or authentication information shown on the source.
• Extraction uncertainty, detected conflicts and review status.

Preserve qualitative results such as “complies” as qualitative values. Do not manufacture a numerical result. Preserve symbols such as less-than signs, because removing one can change the meaning of a reported value. A system should not treat “not detected,” a blank field and zero as interchangeable.

An alias table can connect supplier test names to internal names, but the mapping needs controlled ownership. Similar wording does not establish equivalent analytical methods. Keep uncertain mappings in a review queue until the responsible specialist resolves them.

Control units and specification versions

Unit conversion is often presented as a simple calculation, but the scientific basis may matter as much as the unit. A percentage reported on a dry basis is not necessarily comparable with a result on an as-is basis. The application needs to preserve that distinction and route the comparison when it lacks an approved rule.

Where a conversion is authorised, record the original value and unit, conversion rule, resulting value and rule version. Do not overwrite the supplier's reported result. The reviewer must be able to reconstruct the comparison.

Use the specification applicable to the material and intended use. “Latest specification” is not always enough: a changed specification may have an effective date, transition arrangements or product-specific scope. A controlled mapping should explain which version applies to the receipt.

The same discipline applies to corrections. FDA's final data-integrity guidance discusses original records, metadata and review in the US drug-CGMP context. The engineering implication for this workflow is to retain the source and explain changes instead of storing only a detached final number.

Start with a bounded automation pilot

Choose a small set of materials and suppliers with representative document layouts. Include clean digital PDFs, scanned pages, multi-page tables and documents containing corrections. Obtain permitted, appropriately redacted samples for engineering work; avoid sending confidential records to an unapproved service.

Run the extraction and comparison beside the existing approved process. Classify errors by field and consequence. A wrong footer address is different from a missed minus sign, wrong batch or incorrect acceptance limit. A single overall accuracy percentage cannot show that difference.

The pilot should contain deliberate challenges:

1. A correct document for the wrong lot.
2. A result with a missing unit or changed reporting basis.
3. A specification revision that is not applicable to the receipt.
4. A repeated document submission.
5. A corrected certificate that must remain linked to its predecessor.
6. A document with an apparently complete table but a missing required test.

Agree acceptance criteria before judging the results. Measure missed material-identity mismatches, critical-field extraction errors, reviewer corrections and elapsed time to resolve an exception. Keep the human review required by the approved process even when the model reports high confidence.

Define the boundary with ERP and LIMS

An enterprise resource planning (ERP) system usually needs a receipt identity and material status. A laboratory information management system may own testing and reviewed results. A document service may own extraction and the evidence queue. Decide which system holds each authoritative record rather than copying unrestricted “approved” flags among them.

The integration should tolerate duplicate files and delayed messages. A repeated upload must not create another receipt. A later rejection or hold must not be overwritten by an older successful comparison message. Maintain an exception queue with an accountable owner when document, receipt and laboratory states disagree.

The inventory-status guide explains the downstream stock controls. The instrument-integration guide addresses the related problem of preserving result context across systems.

Summary and next actions

A COA workflow is useful when it establishes what the document says, which receipt it belongs to and what authorised decision it can support. Accurate OCR alone establishes none of those relationships.

• Select representative recent receipts and collect their COAs, specifications and receipt records.
• List the identity, specification and unit mismatches reviewers encounter; assign an owner to each exception.
• Agree which comparisons can be automated and which require an authorised review decision.
• Test deliberately incorrect documents before allowing extracted data to influence operational status.

Pair that work with supplier qualification. NextEdge Labs can help scope the document model, review interface and integrations; the material acceptance policy and disposition remain with the organisation's authorised quality team.