Table of contents
At a glance
How do you validate OCR for GxP-relevant documents?
Define the information the OCR workflow should recognise, assign and pass on. Test against an independent reference set covering relevant document types, layouts and failures. Assess characters, fields, relationships within tables and downstream actions separately. Confidence scores do not replace evidence of suitability. Control corrections, human review, model changes and the link to the source document.
OCR, Extraction and Separation of Professional Judgment
Optical character recognition converts document content into machine-readable text. Beyond this, document processing can assign fields, interpret tables, identify document types, and execute rules. Some methods employ trained models; others rely on rule-based processing. Determine the actual implementation architecture. The term "OCR" alone does not imply that each component employs artificial intelligence, nor does it indicate that the overall system is suitable for a GxP process.
Describe the intended effect. Does recognition support only a search, generate a draft for human review, or automatically pass values to a LIMS? These usage scenarios have different consequences for errors. A misidentified search term may trigger a different control than a misentered limit value. Define which fields are decision-critical and specify when processing must halt or be escalated for review. The system boundary includes the required follow-up actions.
Sources: [1]
A well-known source document is still not an independent reference standard.
Create a controlled reference dataset with independently verified target information. Document document type, origin, version, and expected values. Consider the acceptable input qualities in operational use: resolution, oblique orientation, shadows, multi-page files, overlaid stamps, handwriting, or varying supplier layouts. Explicitly exclude features that are not supported. A system must not silently treat an input outside the approved scope as a valid normal case.
Separate examples for the development of examples for the final performance test. A model can process well-known templates without mastering new relevant variants. Even subsequent night-raining based on the supposed test set changes its independence. Maintain the relationship between model level and reference set used. Data protection and usage rights of the documents are clarified separately; a genuine production document is not only due to its usefulness for a test to a freely usable training source.
Significance and field mapping are distinct verification steps
First check individual characters and units, including O and 0, I and 1, minus signs and decimal separators. Then check assignment. A correctly recognised number can belong to the wrong analyte, batch or column. Compare value, unit, parameter, row, column and document identity against a reference to the extent needed for the subsequent decision.
Consider empty and ambiguous fields. An unfound result must not be automatically interpreted as zero. Several similar batch numbers on one page need a unique assignment rule. A document composed of multiple sources must not be considered a single certificate without testing. Define the uncertainty the system detects and how it becomes visible. Only a plausible-looking overall expression cannot prove these controls.
Example: three correct character strings, one incorrect statement
A supplier certificate lists batch A0O17, water at 0.05% and residual solvent at 0.50%. In this exercise, OCR recognises the digits correctly but assigns 0.50% to water. The character match rate can look excellent while the result is wrong. The test must demonstrate that the parameter, value and batch remain correctly linked.
Add a truncated unit, a follow-up page, and an additional certificate with a similar number. The intended treatment will be predefined. Where a secure linkage is absent, a specialist review may be required. The following work aid presents individual review suggestions, not universal OCR acceptance thresholds. Technical errors are generated in controlled reference documents; actual released records may be altered without detection.
| Reference case | Possible undetected error | To be evaluated reaction |
|---|---|---|
| A0O17 with similar features | Confusing the digit zero with the letter O | Correct identity or observable clarification |
| Two similarly formatted result rows | Correct value at the wrong parameter | Relationship between field and table context |
| Cut unit | Plausible number with incorrect interpretation | Defined rejection or independent review |
| Second page with release information | Incomplete document interpretation | Completeness and correct page linking |
| Dual submission of the same document | A second technical process occurs | Appropriate rule for duplicates and events |
| Corrected OCR value | Original and editing not discernible | Received source and traceable correction |
Assess confidence values within our own process
Manufacturers may provide different confidence levels for document types, words, or fields. Microsoft describes a separate evaluation using representative documents, upon which a decision is made regarding automatic processing or manual review. Do not treat a cited threshold as a regulatory requirement. A technical confidence value is also not an independent proof that your specific extraction case was correctly processed.
Assess the actual misclassification events on your test set. Which incorrect values still receive a high score? Which correct values are unnecessarily flagged for review? Investigate this by relevant field and document classes. A common average may mask a poorly detected rare threshold. Changes to layouts or model versions may alter the validity of a previously established threshold. Therefore, release criteria define the scope of evaluation and the conditions under which the decision logic is applied.
Key figures relate to error sequence.
Define the unit you count, such as a character, word, field, document or complete business transaction. At 99% field accuracy, 10,000 evaluated fields include 100 incorrect fields. This alone does not show whether decisions are adequately protected. A wrong telephone number and a wrong release status count equally but can have very different consequences. Assess critical error types individually.
For automatically captured information, it is particularly relevant to identify which erroneous values the overall system accepts without detection. A high detection rate alone does not adequately address this question. Document the denominators, exclusions, and cases already manually corrected. Otherwise, the metric may reflect the corrected final state rather than the original OCR performance. Acceptance criteria are derived from usage patterns and error sequences. A general minimum detection rate for all GxP OCR applications cannot be established from these examples.
Human verification as a separate controlled step
A review requires an appropriate presentation. The reviewer should be able to clearly see the extracted value at its source, including necessary context. Verify that highlighted locations actually direct to the correct page and line. A preview that displays only the already detected text may confirm a detection error rather than uncovering it. Define which fields must always be reviewed and specify when additional uncertainty triggers a further assessment.
Corrections must remain linked to the original document and the processing procedure. Verify permissions, reason for change, and the transmission of the corrected value. If the source document is modified, a previous release must not be inadvertently carried over to a new content. Train reviewers using actual error classes from the controlled test set. The presence of a human confirmation button alone does not constitute effective control or independent verification.
Sources: [4]
Source document, model, and follow-up action collectively obtained
The required reconstruction may include source document, extraction, model or rule version, corrections, and target record. Establish the primary record. An OCR-generated text file is not automatically a complete True Copy of a complex original document. Missing elements such as images, signatures, table relationships, or other required information may occur. Assess the suitability of a copy for the intended use separately from the actual recognition performance.
Changes to the model, pre-processing, field schema, or target mapping require an impact assessment. A new scan filter may remove weak features; a new supplier layout may alter mappings. Use a controlled regression dataset and supplement with relevant new variants. The system must be able to detect and process inputs outside the scope of the tested range. This results in a controlled document process rather than a collection of plausible-looking extraction results.
Project experience and collaboration
Evaluate OCR accuracy against the correct business result
With the process team and IT, we select an independent reference set from the document types you actually use. Layout changes, ambiguous characters and similar identifiers reveal what errors the OCR process and human review detect. Depending on the consequences of failure, targeted checks on critical fields may suffice or automated downstream processing may need tighter limits. We trace corrections to the value used at the destination and its source document. A high confidence score cannot replace that comparison.
Frequently asked questions about GxP validation of optical character recognition
Does an OCR accuracy of 99 percent meet GxP requirements?
A general rate is insufficient as justification. Define unit of measure, error classes, and consequences for the process. Specifically assess incorrectly accepted critical information and the impact of additional controls. Suitability is demonstrated for your specific scope and process; no universal release threshold is established from this example.
Does a high confidence value indicate a correct measurement result?
No. The score is an output of the employed method. Evaluate on an appropriate independent test set, assessing its relationship with correct and incorrect results. The key factors are actual field mapping and professional interpretation. A representative manufacturer-defined threshold is not a regulatory acceptance criterion.
Is the OCR text an exact copy of the scanned document?
Not automatically. Verify that all content and relationships required for the intended use are fully and accurately preserved. Plain text alone may lose images, layout, signatures, or table references. Therefore, the evaluation of a copy and the evaluation of recognition performance address different questions.
Sources: [4]
Do we also need to test poorly scanned documents?
Test the quality variants that may occur within the approved operating environment. Concurrently, define inputs outside this scope and establish procedures for their handling. An inappropriate scan may be rejected or investigated to ensure clarity. Crucially, it must not lead to an apparently reliable automated decision without proper oversight.
Can a human reliably detect each OCR error?
This must be evaluated based on the designed review process. The reviewer requires source data, context, clear criteria, and sufficiently distinct error representations. Representative error cases should be assessed. A formal confirmation step or comparison with the same erroneous OCR text does not constitute effective control.
When do we need new validation evidence following an OCR modification?
Evaluate the effect on the released use. Change of model, new layout, modified pre-processing and customized target mapping can change the statement. Select appropriate regression tests and new reference cases. The decision refers to actual changes and existing controls, not only the version number.
Sources and review date
Status: October 15, 2026. Regulatory classification and manufacturer responsibilities are separately documented. Examples, test matrices, and templates are standalone reference tools; they are customized for actual usage purposes.
- European Commission: EU GMP Annex 11, Computerised Systems (2011)Current version as per EudraLex inventory as of 15.10.2026. Application, risks, operation, and data lifecycle; specific testing proposals are derived independently.
- Microsoft: Interpret and improve model accuracy and confidence scoresField and document confidence, as well as internal assessment. Confidence is not a blanket assurance of error-free operation within own GxP processes.
- Microsoft: Transparency note for Document IntelligenceOwn representative evaluation and decision on Human Review. Example limit values are not regulatory acceptance limits.
- PIC/S: PI 041-1, Good Practices for Data Management and Integrity (2021)Inspection Guidance on Data Integrity; in particular electronic data, transmission, access, and review.
Sources checked: .
Compare OCR outputs with known originals
Based on a representative selection of your documents, we identify critical fields and known target values. Thereafter, detection, manual review, and error handling can be appropriately assessed.
- Computer system validation →
Bring your GxP systems into validated operation, from laboratory software, MES and eQMS to ERP. We define the scope, plan risk-based tests using GAMP 5 and hand the evidence to QA for release.

