# Validating document digitization under GxP: vendor checklist

By Alyse Gonthier, PhD · Published November 3, 2026

Canonical: https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist

## In brief

For GxP document digitization, validation should demonstrate that the configured workflow produces data fit for its defined use. The evidence comes from testing representative records against field-specific acceptance criteria, covering extraction errors, missing data, expert review, and delivery. Supplier documentation supports this work, while your team remains accountable for the process it puts into use.

## Key takeaways

- Validation evidence must cover the configured digitization workflow and its intended use, including model performance, review, and delivery of approved data.
- Acceptance criteria need to reflect both field type and the consequences of an error; a document-wide accuracy figure can hide critical failures.
- Changes to models, rules, document layouts, and downstream interfaces need assessment so the validation evidence continues to cover the process in use.
- Testing must detect missing records and values as well as incorrect extraction, and separately measure errors that escape review.
- Expert review supports control when decisions remain traceable to the source, reviewer, time, and reason for any correction.

## Contents

- [What is being validated](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#what-is-being-validated)
- [Step 1: Formally record the intended use](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#step-1-formally-record-the-intended-use)
- [Step 2: Set acceptance criteria by field type](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#step-2-set-acceptance-criteria-by-field-type)
- [Step 3: Test representative site records](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#step-3-test-representative-site-records)
- [Step 4: Verify review and approval](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#step-4-verify-review-and-approval)
- [Step 5: Control changes and ongoing performance](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#step-5-control-changes-and-ongoing-performance)
- [What evidence should be ready before routine use?](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#what-evidence-should-be-ready-before-routine-use)
- [A checklist for evaluating a digitization vendor](https://www.katalyzeai.com/blog/validating-document-digitization-under-gxp-vendor-checklist#a-checklist-for-evaluating-a-digitization-vendor)

## **What is being validated**

A value can be read correctly and still reach the wrong batch, lose its unit, or bypass required review. Validating digitization means testing the configured workflow from the source record to the data someone will use. The steps below describe a practical approach for executed pharmaceutical manufacturing records, with testing effort based on the consequences of a wrong or missing value.

For finished pharmaceuticals in the U.S., relevant requirements include 21 CFR Part 211, particularly § 211.68 on computerized equipment and § 211.188 on batch records. Assess Part 11 applicability to required records maintained electronically and relied on for regulated activities, and to relevant electronic signatures. In the EU, GMP Annex 11 calls for validation and data integrity controls based on a documented risk assessment.

## **Step 1: Formally record the intended use**

Before testing, document four things:

1. **Records in scope**: document types, sites, suppliers, formats, and languages.
2. **Fields in scope**: extracted values, definitions, units, and links to the source.
3. **Intended use**: the decision or process the output supports, its GMP significance, and the consequences of an error.
4. **Limits and responsibilities**: excluded records, fields, and uses, and who reviews and approves the output.

Keep the original record, scanned image, and extracted dataset distinct. Accurate extraction alone does not establish a [true copy](https://www.katalyzeai.com/blog/true-copies-scanned-batch-records-and-data-integrity) or authorize destruction of paper originals; retention and copy verification need their own documented controls.

## **Step 2: Set acceptance criteria by field type**

Set acceptance criteria before testing, by field type and consequence of error. A printed batch identifier may be easy to read but still critical if a mismatch attaches results to the wrong batch. The table identifies checks to turn into measurable criteria for the intended use.

| Field type | What to test | What can go wrong |
| --- | --- | --- |
| Printed form fields | Extracted value matches the source, including identifiers and units | A misread identifier can attach accurate results to the wrong batch |
| Handwritten values | Exact match and whether incorrect or unreadable values reach review | A plausible but incorrect reading may escape a confidence flag |
| Corrections | Correct amended value, with the original entry and amendment details accessible | Extracting a superseded value can reverse the meaning of the correction |
| Tables spanning pages | Row and column integrity across the page break | A shifted row or column can attach values to the wrong batch or parameter |
| Signatures and initials | Signature or initials remain visible and linked to a verified identity where needed | Reading a signature image does not authenticate the signer |
| Margin notes and free text | Relevant text is retained with its source and associated entry | An omitted note can remove the explanation for an unusual value |

*Extraction checks and failure modes to address in field-specific acceptance criteria*

Specify how each check will be scored, what counts as an error, the allowed error rate, and when review is mandatory. Define any permitted normalization, such as date formatting, so a correct transformation is distinguishable from a transcription error. Evaluate raw extraction and approved output separately, including values missed entirely and wrong values that escaped review.

## **Step 3: Test representative site records**

Build a test set that covers the document versions, suppliers, languages, and scan conditions in scope. Include difficult cases, document the sample rationale, and compare the output with independently verified reference answers.

The test plan should:

- **Include** faint scans, double-sided pages, amended entries, missing pages, and unreadable fields.
- **Establish** reference answers independently of the system and verify them with a second person.
- **Keep** test records separate from those used to configure or tune the process.
- **Report **errors and omissions by field type, including incorrect values that were not flagged.

Reconcile incoming records and pages with processed outputs, exceptions, and rejected files. Challenge downstream delivery too: identifiers, units, source links, and approval status must survive export without silently losing a record.

## **Step 4: Verify review and approval**

Test whether the review workflow catches and resolves errors before data is approved. Reviewers need the source page, extracted value, and relevant context, with a traceable record of corrections or rejections, their identity, time, and reason.

For example, suppose a handwritten quantity cannot be read reliably. A useful test checks that it reaches a reviewer, stays out of approved output until resolved, and retains the source and decision history afterward. Also check unflagged values against independently verified records, since a history of resolved flags cannot reveal missed errors.

## **Step 5: Control changes and ongoing performance**

Assess changes to model versions, prompts, extraction rules, schemas, scanner settings, document layouts, and downstream interfaces before use. Document the impact and retesting needed, and keep version records so results can be traced to the configuration that produced them. Periodically review performance and exceptions to check that the evidence still supports the intended use.

As of November 2026, the proposed EU GMP Annex 11 revision and new Annex 22 remain consultation drafts. The current Annex 11 already addresses change control and periodic evaluation; the companion Annex 22 covers the proposed AI provisions.

## **What evidence should be ready before routine use?**

Bring together the intended use, risk assessment, approved acceptance criteria, and test results, with links showing which test supports each requirement. Record failures, their investigation and resolution, and any restrictions on use. A supplier package can contribute evidence, but it must cover the configuration and use your team is approving.

Identify the process owner and the people responsible for review, data approval, and change assessment. Confirm that the relevant procedures, training, and approval under your quality system are in place before routine use.

## **A checklist for evaluating a digitization vendor**

Alongside supplier quality, security, and support checks, ask for evidence specific to the configured digitization workflow:

- Can the vendor show **field-level performance** on comparable records, including omissions and errors that escaped flags?
- Can they demonstrate handwriting, corrections, margin notes, and cross-page tables on an **annotated record**?
- Can each value be **traced** to the source page and position throughout review and export?
- Can reviewers **correct or reject values**, and can unresolved exceptions be blocked from approved output?
- Does the **audit trail** preserve extraction, changes, and approval decisions, and can it be retrieved?
- Are model, prompt, rule, and schema **versions identifiable**, with change notification and impact assessment?
- Which validation documents and test evidence are **supplied**, and which activities remain the customer’s responsibility?
- Do exported **values retain** identifiers, units, approval status, and source links?
- Can the team **reconcile** every submitted record with an output or visible exception, including unreadable files?

Evaluate the answers against the same intended use and acceptance criteria as your validation plan. A demonstration can reveal gaps; retain the supporting records and test evidence rather than relying on the demonstration alone.

## Questions

### How should we evaluate the review workflow?

Test whether reviewers can identify and resolve errors using the source record, and whether unresolved values are kept out of approved output. Compare both flagged and unflagged values with independently verified reference records to measure errors that escape review. Retain the review decisions and test results as evidence of how the workflow performed.

### Does FDA’s Computer Software Assurance guidance govern drug manufacturing digitization?

No. FDA’s February 2026 Computer Software Assurance for Production and Quality Management System Software addresses medical device production and quality management systems. A risk-based approach to drug manufacturing digitization still needs to be grounded in the requirements applicable to that use.

### What accuracy level does a regulator require?

There is no universal extraction-accuracy percentage for every GxP use. Define the measure, denominator, and acceptance threshold for each relevant field or failure type, justify them against the intended use, and test both extraction and the controls that produce approved output.

### What happens when the vendor updates the model?

Assess its impact through change control before use and retest as justified by that assessment. Retain version and test records so the quality team can reconstruct which configuration produced a result.

### Can we rely on a vendor's validation package on its own?

No. Supplier evidence can support qualification and testing, but the customer remains accountable for the configured workflow, its intended use, and its operation within the quality system.

## Sources

- [Computer Software Assurance for Production and Quality Management System Software](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/computer-software-assurance-production-and-quality-management-system-software), U.S. FDA, February 2026
- [21 CFR Part 211](https://www.ecfr.gov/current/title-21/chapter-I/subchapter-C/part-211), U.S. Code of Federal Regulations, November 3, 2026
- [Part 11, Electronic Records; Electronic Signatures: Scope and Application](https://www.fda.gov/media/75414/download), U.S. FDA, August 2003
- [Data Integrity and Compliance With Drug CGMP: Questions and Answers](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/data-integrity-and-compliance-drug-cgmp-questions-and-answers), U.S. FDA., December 2018
- ['GXP' Data Integrity Guidance and Definitions, Revision 1](https://assets.publishing.service.gov.uk/media/5aa2b9ede5274a3e391e37f3/MHRA_GxP_data_integrity_guide_March_edited_Final.pdf), MHRA, March 2018
- [EU GMP Annex 11: Computerised Systems; EudraLex Volume 4 adopted guidance list](https://health.ec.europa.eu/medicinal-products/eudralex/eudralex-volume-4_en), European Commission, January 2011
- [Consultation on revised Chapter 4, Annex 11, and new Annex 22](https://health.ec.europa.eu/consultations/stakeholders-consultation-eudralex-volume-4-good-manufacturing-practice-guidelines-chapter-4-annex_en), European Commission, October 7, 2025

## Read next

- [The complete guide to pharmaceutical document digitization](https://www.katalyzeai.com/guides/the-complete-guide-to-pharmaceutical-document-digitization): Pharmaceutical document digitization is the process of reading GxP records such as executed batch records and certificates of analysis into a defined schema, with every value linked back to the page it came from. A scan by itself is a picture, not data. Done well, digitization produces structured data that reviewers can check against the source, that meets the regulatory expectation for accurate and complete records, and that other work can build on. This guide covers how digitization works, what OCR is and how to evaluate its accuracy, and how to build trust with digitization technology. 
- [How to digitize paper batch records without losing the source](https://www.katalyzeai.com/blog/how-to-digitize-paper-batch-records-without-losing-the-source): Digitizing a paper batch record means reading every value on the executed record, including handwritten entries, corrections and signatures, and loading into a defined schema, with each value linked to the location it came from. The steps are to define the schema, capture a good image, extract, check against rules, have an expert review flags, and deliver the approved data. Keeping the link to the source at every step is what makes the result usable in regulated work.
- [How to digitize certificates of analysis from suppliers](https://www.katalyzeai.com/blog/how-to-digitize-certificates-of-analysis-from-suppliers): Digitizing a supplier certificate of analysis (CoA) means extracting the material, lot, tests, specifications, results, and methods into one schema, with each value linked to the certificate it came from, whatever layout the supplier used. The steps are to define one schema for every supplier, map each supplier's terms to it, extract, check the results against the specification, and review anything flagged. Structured certificate data makes it practical to trend a supplier's results across lots and to support the verification of supplier data that 21 CFR 211.84 requires.

[See How It Works](https://www.katalyzeai.com/product/document-intelligence)