Part 1 — The Risk Taxonomy
The seven categories that actually derail AI projects
Most AI transformation failures are not model failures. Industry trackers commonly cite project failure or stall rates in the 70–90% range, which is worth treating as a direction rather than a precise figure — and the leading causes they point to are unclear governance, weak data foundations, and poor change management, long before anything to do with the model itself.
This playbook gives founders and COOs running an AI transformation a repeatable way to find, assess, and control the risks that derail these projects. It applies to anything that embeds AI into products, operations, or decisions — from a single automation pilot to a company-wide AI operating model — and it sits alongside standard project governance, meaning the steering committee and stage gates you already run, not in place of them.
The seven categories
Risk splits into seven categories, adapted from the NIST AI Risk Management Framework (govern, map, measure, manage) and ISO/IEC 23894.
| Category | Typical trigger | Early warning sign |
|---|---|---|
| Strategic & business case | AI chosen before the problem is defined | No agreed success metric after kickoff |
| Data | Fragmented, unlabeled, or low-quality sources | Data prep eats more than half the project timeline |
| Technical & model | Drift, hallucination, weak integration | Accuracy degrades post-deployment with no monitoring in place |
| Governance & compliance | No accountable owner, unmapped regulation | Nobody can answer “who approved this model” |
| Organizational & change | No change plan, unclear new roles | Low tool adoption weeks after rollout |
| Vendor & third-party | Single vendor, vague SLA or data terms | Vendor roadmap changes affect your commitments |
| Ethical & reputational | Untested edge cases, biased training data | Complaints or inconsistent outputs across user groups |
Every category here has an early warning sign before it becomes a failure. The work in the rest of this playbook is noticing those signs while they are still cheap to act on.