Every large enterprise now has AI pilots. Very few have AI in the P&L. The gap is rarely the model. It is the distance between a proof-of-concept that works on a curated dataset and a capability that runs inside the operating model — with an owner, controls, and data that can be trusted every day.
We see the same pattern across banking, energy, retail and manufacturing: ambition in the boardroom has outrun data maturity on the ground. Pilots succeed because a small team hand-carries the data. Scaling fails because nobody owns the pipeline, the definitions, or the decision the model is supposed to change.
What a “successful” pilot usually hides
A pilot is declared successful when the model performs. Nobody asks the questions that decide whether it will survive contact with the organisation:
- Which decision does it change, and who owns that decision? A churn model with no retention owner is a report.
- Where does the data come from on a Tuesday in March? In the pilot, an analyst pulled an extract. In production, it has to arrive on schedule, with lineage, from systems that change.
- Who is accountable when it is wrong? Model risk, bias, drift and explainability are not compliance decoration; they are the conditions under which a regulated business is allowed to act on a prediction.
- What does it cost to run? Inference, monitoring, retraining and support are recurring; the pilot budget was one-off.
A pilot that cannot answer these is not 80% of the way to production. It is closer to 20%.
The four gates
The fix is unglamorous, which is why it is skipped. Four things, built once for the portfolio rather than re-invented per pilot, separate the AI that reaches the P&L from the AI that reaches the slide deck.
1. A decision owner
Start from the decision, not the model: pricing, credit, replenishment, maintenance, onboarding. Put a business owner on it, with the authority to change how the decision is made and a number they are measured on. If no one wants to own the decision, stop — the model has no home.
2. A governed data path
Build the pipeline, definitions and access controls once, as a data product, so the third use case inherits what the first one paid for. This is where most of the money goes and where most firms under-invest. It is also the difference between “AI-ready” as a slogan and as an architecture: a semantic layer everything reads from, lineage that can be audited, quality checks that run before the model sees the data.
3. A sequenced portfolio
Rank candidate use cases on value at stake, data readiness and risk — then sequence them so the quick, safe wins fund the foundations the harder ones need. A portfolio that starts with the most exciting use case usually starts with the one furthest from its data.
4. Assurance built in
Monitoring, controls, human-in-the-loop points and model risk belong in the build, sized to your regulators rather than to a framework poster. Retro-fitting assurance to a pilot is slower than building it in and, in our experience, is where the project is quietly cancelled.
Why returns on data are non-linear
Firms that stall at reporting get linear returns: a better dashboard, a faster close. The leap from reporting to decision-native operations — where the model’s output is the default action and a person handles the exceptions — is where advantage concentrates. It is also the leap that needs all four gates at once, which is why it is so rarely made.
The shortest credible path to the P&L therefore runs through the operating model, not through another pilot. The organisations that get there treat AI as a change to how decisions are made, funded and governed — and treat the data platform as the infrastructure that change runs on.
A practical starting point
If you have pilots and no production, three weeks of work answers the question honestly: which decisions the pilots were for, which of the four gates each one is missing, and what it would take to build the gates once for the portfolio. That is usually a shorter list than the number of pilots suggests — and a much shorter list than the number of models.