The gap
Every organisation now has MLOps tools. Very few have an operating model that makes those tools actually work in a live business.
What is usually missing
- No single owner for a model once it is in production.
- Retraining decisions made in endless alignment meetings instead of automated triggers.
- No defined escalation path when a model starts degrading.
- Governance and compliance treated as a one-time checkbox.
- Handover from data science to operations is informal and incomplete.
What a working AI operating model includes
- Clear RACI for production AI: who owns the model, who monitors it, who approves retraining, who can roll it back.
- Automated runbooks for common failure modes (drift, data quality drop, cost spike).
- Defined retraining cadence tied to business events or performance thresholds, not calendar dates.
- Production readiness checklist that must be signed off before go-live.
- Post-go-live review process at 30/60/90 days to catch the issues that always surface after handover.
The blunt rule
If no one can answer “who gets paged at 2 a.m. when this model breaks?” the operating model does not exist.
The fix
Treat the operating model as a first-class delivery artefact, not an afterthought. Document it, automate what you can, and make it part of every AI project’s acceptance criteria. The difference between pilots that stall and systems that run for years is almost always the operating model.