The Achilles heel of most AI programmes isn't the model

The model was never the constraint

Walk into most stalled marketing AI programmes and you'll find impressive infrastructure: a frontier model on tap, a vector store humming, a governance policy printed and signed. What you won't find is a single senior operator who can name — in one sentence — the commercial decision the system is meant to change.

That is the actual Achilles heel of enterprise marketing AI. Not compute. Not the model. Not even the data, though the data is usually a mess. The heel is capability: the ability of your people to turn a probabilistic system into a repeatable, defensible commercial motion.

Three capability gaps we see on every engagement

• Decision literacy — teams that can prompt a model but can't articulate which decision it's replacing, augmenting, or auditing.
• Review muscle — no defined human-in-the-loop step, so outputs either ship unchecked or bottleneck on one exhausted reviewer.
• Operating rhythm — pilots run in isolation from planning cycles, so nothing that works in the lab ever meets a quarterly plan.

What senior operators do differently

The programmes that compound share a pattern. They start with the decision, not the tool. They pick one revenue-adjacent decision — next-best-offer, media reallocation, lifecycle triage — and design the AI system around the review gate that decision already has.

They also stop pretending the org chart is ready. Someone owns the model's outputs the way a trader owns a position: with a mandate, a limit, and a rollback plan. That role rarely exists on day one. Creating it is the transformation.

The board test

A useful gut check: could your CFO explain, without notes, how the AI programme changed a number on the last quarterly report? If the answer is a hedge, you have a capability problem dressed as a technology problem.

Fix the heel and the model gets cheaper, the governance gets simpler, and the second round of funding becomes an easier conversation. Skip it, and every new model release just extends the runway on the same unresolved question.

Where to start this quarter

• Pick one commercial decision. Write the sentence: 'This AI system changes how we decide X.'
• Name the human reviewer, their limit of authority, and their SLA.
• Instrument the decision — not the tool — so you can prove lift or kill the programme in two quarters, not two years.