The Five Risk Categories
AI projects fail in different ways than traditional software projects. Mapping the failure modes to risk categories early lets you build mitigations rather than be surprised. The five categories that cover almost every incident we've seen:
- Model risk — the AI is wrong in ways that matter.
- Data risk — the data going in is bad, missing, or leaks out.
- Operational risk — the system breaks or is misused.
- Regulatory risk — you've violated a law or guidance.
- Reputational risk — the AI does something that ends up in the press.
Model Risk
Hallucination, bias, drift, performance degradation. Mitigations: grounded retrieval, evaluation suites, regular re-testing, human-in-loop on high-stakes decisions.
Data Risk
Sensitive data leaving your perimeter, training on private data, stale data driving wrong answers, PII in logs. Mitigations: BAAs and DPAs, no-training contract terms, data classification, log scrubbing.
Operational Risk
AI outages, rate limit hits, latency spikes, misuse by employees. Mitigations: multi-vendor abstraction, rate limit monitoring, internal AUP for AI use.
- • Who declares an AI incident, and on what trigger?
- • What's the kill switch?
- • How does the team communicate (internal and external)?
- • What's the rollback path?
- • Who does the post-mortem?
Regulatory Risk
GDPR, CCPA, AI Act in the EU, sector-specific rules (HIPAA, GLBA, FINRA), and an emerging patchwork of US state AI laws. Mitigations: legal review before launch, periodic audits, documented model methodology, transparency to customers.
Reputational Risk
AI doing something embarrassing in front of customers or in the press. Mitigations: tight guardrails, sample monitoring, fast incident response, an ability to pull the plug.
The Practical Framework
- Identify the use case's risk profile (low/medium/high on each category).
- Set mitigations proportional to the risk.
- Define monitoring that would catch a problem early.
- Define incident response in advance.
- Review quarterly.
The companies that have AI incidents in the news are not the ones that took risk. They're the ones that didn't plan for the risk they were already taking. The difference is documentation and rehearsal.
FAQ
What about insurance? Some carriers now offer AI-specific coverage. Useful for high-risk deployments.
Who owns AI risk? The AI champion at the operational level, the CEO at the strategic level.
How often to review? Quarterly for tactical, annually for strategic.