The headline number is well known by now: 95% of corporate AI projects deliver no measurable ROI. Most do not reach production. Some are killed before pilot completes. Some land but cannot be operated by the team they were built for. The pattern is consistent across sectors, geographies, and consultancy models.
The reason, in our experience across the operational economy, is rarely the model. The model is usually fine. The reason is the gap between a boardroom conversation about AI and a specific operational problem with a measurable outcome. That gap is where projects die.
Where the gap actually sits
The hyperscalers — Google, AWS, Microsoft Azure — provide the infrastructure. The model providers — Anthropic, OpenAI — provide the intelligence. Both have done their job. What sits in between is the translation work. Translation from a sector-specific operational problem to a working system on the operations team’s dashboard. Translation from “we need an AI strategy” to “the system answered 7,400 calls last month and books appointments correctly 94% of the time.”
That translation work is engineering, not strategy. It involves data plumbing, integration with legacy systems, careful prompt and tool design, observability, on-call procedures, and a handover that lets your team own the system after we leave. None of it is glamorous. All of it is what makes the difference between a pilot and a production system.
What actually works
Three patterns that consistently move AI from boardroom to production:
- Scope to the outcome, not the feature. Pick the operational outcome you are willing to measure: cost per call, MTBF on a specific equipment class, OEE on a specific line. Reverse engineer the system from there.
- Production from day one. The first artefact is a working agent in front of a real user, not a slide deck or a Jupyter notebook. Pilots that “graduate to production” rarely do.
- Hand it over. The system is owned by your team after we leave. Not by us. Documentation, tests, and a deployment story your engineers can read are the artefacts that determine whether the work survives the engagement.
The bet we are making
Most AI work in market today is a demo wrapped in a deck. The bet we are making, deliberately, is that the customers who will outperform are the ones who treat AI as a production engineering problem and not a strategy exercise. The translation gap is real, the failure rate is real, and the small set of organisations who close it consistently are the ones that pull operational and cost advantage that compounds.
If that is the bet you want to make, the fastest way to find out where AI sits in your operation right now is the half day scoping workshop. No obligation, costed brief at the end, your team in the room.