Almost every large organisation now has a portfolio of AI pilots. Far fewer have AI systems that customers and colleagues rely on every day. The gap is rarely about model capability. It is about decisions that were never made.
1. Decide who owns the outcome
Pilots are often owned by an innovation team, while the process they would change belongs to an operations leader who was never asked. Before building, name the business owner who will run the system and be accountable for its results.
2. Decide how success will be measured
"It works in the demo" is not a success measure. Agree the metric that matters, such as handling time, accuracy or cost per case, measure today's baseline, and set the threshold that justifies going live.
3. Decide the path to production on day one
Model risk, security, data access and support models take time. Treat them as part of the build, not a gate at the end. A pilot designed for production from the start moves faster, not slower.
What this looks like in practice
- A named business owner and a written success measure before any code
- Risk and security reviewers involved in the first sprint
- A pilot that runs on real data, in a production-like environment
- A go or no-go decision at a fixed date, based on the agreed measure
None of this is glamorous. All of it is the difference between an impressive demo and a system that pays back.
