AI Incident Response Plan: A Template for Rapid Response

It's 2 AM. Your AI system just made a decision that affected thousands of customers. Something went wrong. The board is asking what happened. Regulators want answers. And your team is scrambling to figure out what to do.

This is not the moment to create your incident response plan.

AI systems create new categories of incidents—model failures, data leakage, biased decisions, adversarial attacks—that traditional IT incident response doesn't fully address. You need an AI-specific incident response plan ready before incidents occur.

This guide provides a template and framework for building that plan.

Executive Summary

• AI incidents differ from traditional IT incidents in detection, investigation, and remediation approaches
• Categories include: Model failure, data breach, bias incidents, security attacks, governance violations, and third-party AI failures
• Response must be fast but also careful—rushing can cause additional harm
• Roles and responsibilities must be clear before an incident occurs
• Documentation is essential for investigation, regulatory response, and improvement
• Notification requirements vary by jurisdiction and incident type—know your obligations
• Post-incident review prevents recurrence and improves capability
• Regular testing ensures the plan works when needed

Why This Matters Now

AI incidents are inevitable. The question isn't whether your AI systems will experience problems—it's whether you'll be prepared when they do.

Several factors make AI incident response urgent:

AI failures can scale instantly. A traditional software bug might affect one user at a time. An AI model problem can affect every decision the system makes—potentially thousands before anyone notices.

Detection is harder. Traditional systems fail obviously (error messages, downtime). AI systems can fail subtly—making increasingly bad decisions while appearing to function normally.

Regulatory expectations are rising. Regulators expect organisations to have AI incident response capabilities. In some jurisdictions, AI-related breaches have specific notification requirements.

Reputational stakes are high. AI incidents—especially those involving bias or privacy—attract media attention and public concern in ways traditional technical failures don't.

What Constitutes an AI Incident?

An AI incident is any event involving AI systems that:

• Causes or threatens to cause harm (to individuals, the organisation, or third parties)
• Violates laws, regulations, or organisational policies
• Compromises data security or privacy
• Results in significant business impact
• Creates reputational risk
• Represents unexpected or unexplained AI behavior

AI Incident Categories

AI Incident Response Plan Template

Section 1: Purpose and Scope

Purpose This plan establishes procedures for responding to incidents involving artificial intelligence systems used by [Organisation Name].

Scope This plan applies to:

• All AI systems owned or operated by the organisation
• AI systems provided by third parties that process organisational data
• AI systems used by employees in the course of their work

Objectives

1. Minimise harm from AI incidents
2. Ensure rapid, coordinated response
3. Meet regulatory and contractual notification obligations
4. Preserve evidence for investigation
5. Learn from incidents to prevent recurrence

Section 2: Roles and Responsibilities

AI Incident Response Team (AI-IRT)

Escalation contacts:

• Executive Sponsor: [Name, contact]
• Board notification: [Process]
• Regulatory notification: [Process]

Section 3: Incident Classification

Severity Levels

Classification Criteria

Consider when assessing severity:

• Number of people affected
• Type of data involved
• Harm potential (financial, physical, reputational)
• Regulatory implications
• Business continuity impact
• Media/reputational risk
• Reversibility of harm

Section 4: Response Procedures

Phase 1: Detection and Initial Assessment (0-2 hours)

Phase 2: Containment (2-4 hours for Critical/High)

Containment Options:

• Disable affected AI system
• Route to manual processing
• Apply input/output filters
• Reduce AI authority (human approval)
• Isolate affected data
• Revoke access

Phase 3: Investigation (4-48 hours)

Phase 4: Notification (as required)

Phase 5: Remediation (varies)

Phase 6: Post-Incident (1-2 weeks after closure)

Section 5: Communication Templates

Internal Escalation Template

SUBJECT: [SEVERITY] AI Incident - [Brief Description]

SUMMARY:
An AI incident has been identified requiring [immediate/urgent/routine] attention.

DETAILS:
- System affected: [Name]
- Discovered: [Date/Time]
- Severity: [Level]
- Current status: [Investigating/Contained/Remediating]

IMPACT:
- Users/customers affected: [Number/scope]
- Data involved: [Type]
- Business impact: [Description]

CURRENT ACTIONS:
[List of actions being taken]

REQUIRED DECISIONS:
[List any decisions needed from escalation recipients]

NEXT UPDATE: [Time]

Contact: [Incident Commander name and contact]

External Notification Template (General)

SUBJECT: Important Notice Regarding [Service/System]

Dear [Recipient],

We are writing to inform you of an incident affecting [description].

What happened:
[Brief, factual description without speculation]

What information was involved:
[If applicable]

What we are doing:
[Actions taken and in progress]

What you can do:
[Any actions recommended for recipients]

For more information:
[Contact details, FAQ link]

We take this matter seriously and are committed to [resolving the issue/protecting your information].

Sincerely,
[Name, Title]

Section 6: Specific Incident Playbooks

Playbook A: AI Model Failure/Degradation

1. Confirm model is producing incorrect/degraded outputs
2. Document examples of failures
3. Assess scope (what decisions/outputs affected)
4. Implement containment (fallback model, human review, disable)
5. Investigate cause (data drift, model drift, training issue)
6. Determine affected decisions that may need review
7. Plan remediation (retrain, rollback, replace)
8. Test before restoring
9. Monitor closely after restoration

Playbook B: AI Data Breach

1. Confirm data exposure and scope
2. Preserve evidence (logs, access records)
3. Contain breach (revoke access, isolate system)
4. Assess regulatory notification requirements
5. Identify affected individuals
6. Prepare and execute notifications
7. Investigate root cause
8. Implement preventive measures
9. Document compliance

Playbook C: AI Bias Incident

1. Document specific evidence of bias
2. Assess scope (how many decisions, over what period)
3. Determine impact on affected individuals
4. Contain (add human review, adjust thresholds, disable)
5. Investigate root cause (training data, model design, implementation)
6. Assess remediation for affected individuals
7. Develop bias mitigation measures
8. Test for bias before restoring
9. Implement ongoing bias monitoring

Section 7: Testing and Maintenance

Plan Testing

• Tabletop exercise: Quarterly
• Simulation exercise: Annually
• Contact verification: Monthly

Plan Updates

• Review after each incident
• Annual comprehensive review
• Update when AI systems change
• Update when regulations change

Common Failure Modes

1. No Plan at All

Creating a plan during an incident guarantees poor response. Build and test the plan before you need it.

2. Generic IT Plan Only

Traditional IT incident response doesn't address AI-specific issues (model behavior, bias, explainability). AI-specific procedures are essential.

3. Unclear Ownership

"Someone will handle it" means no one handles it effectively. Clear roles and contacts must be designated.

4. Untested Plan

A plan that's never been exercised will fail when needed. Regular testing reveals gaps.

5. Missing AI Expertise

Incident response teams without AI/ML expertise will struggle with AI-specific investigation and remediation.

6. Poor Documentation

Inadequate documentation during response creates problems for investigation, compliance, and improvement.

7. Rushing Remediation

Pressure to restore service quickly can lead to incomplete fixes and recurrence.

Implementation Checklist

Plan Development

• Identify AI systems in scope
• Define incident categories relevant to your AI use
• Establish severity classification criteria
• Define roles and assign individuals
• Create escalation paths
• Develop phase-by-phase procedures
• Create communication templates
• Develop incident-specific playbooks
• Document notification requirements by jurisdiction
• Establish evidence preservation procedures

Preparation

• Train AI-IRT members
• Distribute plan to responders
• Establish communication channels
• Create incident logging system
• Verify escalation contacts
• Prepare response toolkits

Testing

• Conduct tabletop exercise
• Review and incorporate learnings
• Schedule regular testing

Maintenance

• Assign plan owner
• Schedule regular reviews
• Establish update triggers

Metrics to Track

Response Effectiveness

Response Quality

Taking Action

The time to build your AI incident response capability is before you need it. An incident at 2 AM is not the moment to figure out who's responsible, what to do, or who to notify.

Build the plan. Assign the roles. Test the procedures. When the incident comes—and it will—you'll be ready.

Ready to strengthen your AI incident response capability?

Pertama Partners helps organisations build comprehensive AI incident response plans tailored to their systems and regulatory environment. Our AI Readiness Audit includes incident response capability assessment.

Book an AI Readiness Audit →

Disclaimer

This guide provides general information about AI incident response planning. It does not constitute legal advice. Notification requirements and compliance obligations vary by jurisdiction and should be verified with legal counsel. Organisations should adapt this framework to their specific context and regulatory environment.