Case Study · Fictional teaching example

Agentic deviation triage: control actions as well as answers

An assistant summarizes a deviation and proposes a category. A proposed extension would create investigation tasks automatically. Approved severity, root cause, disposition, and closure remain decisions of authorized personnel.

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
The agent reads deviation records and can create draft tasks. Its access, action permissions, confirmation, transaction control, and recovery can affect the regulated investigation process.
Change or Issue
Adding tool use changes the risk from an incorrect suggestion to an incorrect or duplicate workflow action.
Functions Affected
  • Deviation summarization and triage proposal
  • Draft task creation
  • Human confirmation
  • Action logging, retry, and safe stop
Potential Impact
A critical investigation may be delayed, misdirected, duplicated, or exposed to an unauthorized recipient.
Deviations / Known Issues
A timeout creates duplicate tasks in the fictional execution. Action capability stays disabled while read-only recommendations continue under the approved pilot boundary.

Situation and intended use

An assistant summarizes a deviation and proposes a category. A proposed extension would create investigation tasks automatically. Approved severity, root cause, disposition, and closure remain decisions of authorized personnel.

What changes when actions are added?

An incorrect suggestion can mislead a reviewer. An incorrect action can also modify a workflow, notify the wrong people, expose information, create duplicate tasks, or leave a partially completed transaction. The assurance boundary must include tool permissions and failure recovery.

Proposed controls

Limit tools to approved actions and destinations. Require human confirmation where defined by the process. Validate structured action parameters. Prevent duplicate execution. Log action requests, approvals, actual outcomes, and errors. Provide a safe stop and an owned recovery process. Keep production approval and closure permissions outside the assistant’s access.

Worked tests

Challenge an ambiguous deviation, missing product information, duplicate submission, tool timeout, unauthorized destination, manipulated source instructions, and a retry after partial completion.

Completed test extract — fictional: “AG-T09 simulates a timeout after task creation. On retry, the agent queries the transaction state and does not create a second task for the same approved request. The review record identifies the original action and recovery result.”

Worked outcome

The assistant correctly summarizes most cases but creates duplicate tasks after timeout. The team may continue evaluating read-only suggestions under the approved pilot boundary, but action capability remains disabled until duplicate prevention and recovery are supported by evidence.

Monitoring and reassessment

Review incorrect categories, unauthorized action attempts, repeated failures, reviewer corrections, tool changes, permission changes, and incident patterns. A model update can affect action selection even when the tool code stays the same.

Challenge: Can a prompt stating “never close deviations” establish the boundary?

Worked answer: Use technical permissions and workflow controls that prevent closure. The prompt can communicate intent but should not be the only barrier to a consequential prohibited action.