Most AI failures are operating-model failures
AI operating model, adoption, and change management
The question: What organizational design makes AI stick?
Pilots die after launch for organisational reasons: no ownership, no incentives, no training, no measurement cadence. The operating model — who identifies, builds, approves, monitors, scales, retires — is the difference between AI as capability and AI as theatre.
What the lesson covers
An AI operating model answers seven questions: who spots opportunities, who builds, who approves (and against what criteria), who monitors, who scales, who retires systems, and who is accountable when something goes wrong. If any answer is "unclear", incidents will clarify it expensively.
Three archetypes: centralised (one AI team — control and standards, but distant from the business), federated (each unit builds — proximity, but duplication and uneven governance), and hub-and-spoke (a central platform/governance hub + embedded spokes in functions) — the pattern most scale-ups converge on. The hub owns platforms, standards, evaluation infrastructure, and training; spokes own use cases and outcomes.
Adoption is an engine, not an email: training by persona (executives need judgment, managers need workflow redesign, users need protocols), champions inside teams, incentives aligned (if saved time just means more tickets, adoption stalls), manager routines that use the new outputs, and visible metrics. McKinsey's high-performer pattern: C-suite alignment, roadmaps, workflow redesign, and adoption tracked like a product launch — not tool access.
AI literacy is now also a legal requirement: EU AI Act Article 4 obliges providers and deployers to ensure staff AI literacy appropriate to their role. Treat it as a capability programme with role-based curricula — which, usefully, is exactly what drives value anyway.
Measurement cadence makes it stick: weekly usage-and-friction reviews in teams, monthly workflow-metric reviews with managers, quarterly portfolio reviews with kill/scale decisions. Connect every use case to a process KPI and a P&L line; retire what does not move them.
Key points
- The operating model answers who identifies/builds/approves/monitors/scales/retires — before incidents do.
- Hub-and-spoke balances central platforms and standards with functional ownership of outcomes.
- Adoption engine: persona-based training, champions, aligned incentives, manager routines, visible metrics.
- AI literacy is an EU AI Act obligation (Art. 4) — and the highest-ROI training you can run.
- Cadence: weekly friction, monthly workflow metrics, quarterly kill/scale.
Framework — AI Operating Model Canvas
strategy · portfolio · platform · data · governance · talent · adoption · measurement. Fill all eight boxes for your organisation; the empty box is where your next AI failure is scheduled.
The lab
Design an operating model and adoption plan you could present to a leadership team.
Open this lesson, its lab and its quiz
Sources and further reading
- The State of AI — rewiring organisations — McKinsey
- EU AI Act Article 4 — AI literacy — AI Act Explorer