Separate an AI service incident from its innovation portfolio
Decompose the situation, select domain-appropriate actions and define how each component will be sensed and reclassified.
30 minutes
Scenario
A regional insurer is launching an AI-assisted claims service. A duplicate-payment defect is active, an integration latency problem has several technical explanations, customer response to new explanations is uncertain, and password resets follow a validated procedure.
- Your role
- Service owner responsible for restoring safety while preserving disciplined learning
- Method
- Cynefin Framework
Evidence pack
Immediate harm
Twenty-three duplicate payments were issued in 40 minutes; a payment hold and manual approval can stop further loss.
Technical diagnosis
Latency correlates with two integration paths, but architecture and vendor specialists disagree about the mechanism.
Customer behaviour
Three small cohorts respond differently to explanation style; effects vary by claim type and channel.
Routine work
The access-reset procedure resolves 97% of cases and has defined escalation for the remainder.
Guardrail
No experiment may delay a statutory claim deadline or remove phone access.
Authority
The incident lead may hold automated payments for four hours; extension requires the accountable executive.
Constraints
- Contain duplicate payments before full diagnosis.
- Do not experiment across statutory deadlines or accessibility channels.
- Keep expert disagreement visible.
- Time-box emergency authority and define the exit.
Case steps
Work through each prompt using the evidence pack. These guided cases support self-directed practice; server-scored attempts are not available yet.
Separate the situation into components, assign a provisional domain to each and state the evidence and uncertainty.
Choose a response, owner, limit and stopping rule for every component.
Define signals and dates for sensing outcomes, exiting crisis mode and reclassifying the components.