Case Study · Digital teaching example

AI-Assisted Deviation Review

An AI assistant proposes deviation themes and similar records while human investigators retain classification, root-cause, and approval authority.

Author
Sandip Thorat
Published
12 September 2026
Last reviewed
12 September 2026
Version
1.0
Content type
Practitioner guidance
Primary references
See related regulations, guidance, and approved procedures

Standard case study record

Scope and decision fields

Read these fields before the detailed worked case. All organizations, identifiers, test outcomes, and decisions are fictional teaching material.

GxP Assessment
Missed critical similarity, false linkage, obsolete source, privacy exposure, or automation bias can affect investigation quality and timeliness.
Change or Issue
A retrieval-augmented AI assistant proposes related deviations, common themes, and missing investigation questions for a qualified investigator.
Functions Affected
  • Record retrieval and similarity
  • Source version and access
  • Proposed theme and question generation
  • Reviewer disposition
  • Monitoring and change assessment
Potential Impact
The investigation can overlook recurring failure or pursue an unsupported conclusion.
Deviations / Known Issues
The fictional evaluation misses two related deviations stored only in scanned attachments after a parser change.

Situation

A retrieval-augmented AI assistant proposes related deviations, common themes, and missing investigation questions for a qualified investigator.

Intended Use

Support investigation preparation using approved records. The AI cannot classify severity, determine root cause, approve a conclusion, initiate CAPA, or close a record.

Decision Summary

Decision elementCase record
GxP assessmentMissed critical similarity, false linkage, obsolete source, privacy exposure, or automation bias can affect investigation quality and timeliness.
Potential impactThe investigation can overlook recurring failure or pursue an unsupported conclusion.
Known issueThe fictional evaluation misses two related deviations stored only in scanned attachments after a parser change.
Validation conclusionScanned attachments are excluded and the update is not released until critical-miss criteria and reviewer controls pass.

GxP Assessment

Missed critical similarity, false linkage, obsolete source, privacy exposure, or automation bias can affect investigation quality and timeliness.

Affected GxP Functions

  • Record retrieval and similarity
  • Source version and access
  • Proposed theme and question generation
  • Reviewer disposition
  • Monitoring and change assessment

Failure Scenarios

  • Critical prior deviation is missed
  • Unrelated record is presented as similar
  • Restricted record is retrieved
  • Unsupported root-cause wording is accepted
  • Model update changes performance without detection

Potential Impact

The investigation can overlook recurring failure or pursue an unsupported conclusion.

Existing Controls

  • Read-only approved corpus
  • Source citation
  • Role-filtered retrieval
  • Qualified reviewer disposition
  • No autonomous Quality action

Existing Evidence

  • Independently labeled reference set
  • Security assessment
  • Human-workflow study
  • Audit and version records

Evidence Gaps

  • Rare critical events
  • Poor-quality attachments
  • New model version
  • Reviewer workload effect

Testing Approach

  • Evaluate correct links, false links, missed critical links, and source correctness by record type and criticality
  • Challenge prompt injection and restricted access
  • Test reviewer detection and override under representative workload
  • Verify monitoring, stop, rollback, and revalidation criteria

Deviations or Known Issues

The fictional evaluation misses two related deviations stored only in scanned attachments after a parser change.

Validation Conclusion

Scanned attachments are excluded and the update is not released until critical-miss criteria and reviewer controls pass.

Reassessment Criteria

  • Model, prompt, parser, retrieval, source, access, or workflow change
  • Critical miss or privacy incident
  • Acceptance-criteria or supported-population change