The operating model your AI capability deserves

AI in CC · ~7 minute read

Most AI failures are operating-model failures, not model failures. The capability exists; the operating model that keeps it running honestly does not.

What the operating model covers

Six components. Named ownership — a person, with deputy. Observability — usage, quality, drift, fairness, with alerts the owner reads. Governance — change-control tiers, fairness review cadence, annual retirement review. Human-AI roles — who does what, where the HITL sits, what the escalation looks like. Frontline integration — training, TL briefing, feedback channels. Risk & compliance — bias monitoring, vulnerable-customer pathways, data protection, regulatory engagement.

These don’t happen by accident. They happen because someone is responsible.

Ghost ownership and what it costs

A capability without a named owner runs until something breaks. Then everyone discovers no one was watching, no one knows the data flows, no one has the cadence for fairness review, no one is reading the model’s output. The capability becomes the problem of whoever raised it last.

Ghost ownership is the single most common operating-model failure. It is also the cheapest to fix — name an owner before deployment.

Building the operating model before deployment

A useful checklist before any AI capability goes live: owner named; deputy named; observability operational; governance forum scheduled; HITL boundary documented; frontline training delivered; risk review completed; retirement criteria written down.

A capability that passes this checklist is one the operation can sustain. A capability that doesn’t is technical debt with marketing.

The principle

Deploy AI; operate the operating model. The model is the deployable; the operating model is the discipline.

Operations that grasp this distinction get the value of AI. Operations that don’t end up with capabilities that look productive in the demo and produce incidents in the queue.

The AI operating model Six components ▸ Named ownership + deputy ▸ Observability the owner reads ▸ Governance with change-control tiers ▸ Human-AI roles documented ▸ Frontline integration ▸ Risk & compliance integrated Failure modes ▸ Ghost ownership ▸ No observability ▸ Ad-hoc change control ▸ Frontline not engaged ▸ HITL skipped under cost pressure ▸ No retirement criteria Model is deployable; operating model is the discipline

The closing principle

The capability is bought; the operating model is built. Most AI failures are operating-model failures; most successes are operating-model successes that happen to have a model attached.

See also