Most organisations now have at least one AI pilot. Far fewer have AI running in daily operations. The gap is rarely the model — it is the work around it: a clear business metric, usable data, integration into real workflows, and someone accountable for keeping it accurate. Here is the playbook we use.
1. Start from a metric, not a technology
Pick a process where you can measure today's performance: average handling time, forecast error, manual review hours or first-time-right rate. Record that baseline before any build begins. If you cannot measure the problem, you will not be able to prove the solution.
2. Check the data before you promise anything
Spend the first one to two weeks on data: where it lives, how complete it is, who owns it and whether it can legally be used. For generative-AI assistants this means auditing the documents the assistant will answer from. Poor or outdated source content produces confident but wrong answers.
3. Build the thinnest useful version inside a real workflow
A model in a notebook is not a pilot. Put a basic version in front of a small group of real users, inside the tools they already use — the CRM, the claims screen, the support desk. Keep a human approving outputs at first. You will learn more in two weeks of real use than two months of lab testing.
4. Design for production from the start
- Security: run models in your own cloud tenancy or private endpoints, with role-based access and logging.
- Monitoring: track accuracy, drift, latency and cost per transaction on a dashboard someone owns.
- Fallbacks: define what happens when the model is unsure — route to a person, never guess silently.
5. Scale only what beats the baseline
Compare pilot results with the baseline from step one. If the gain is real, expand users and volume in stages, retraining or re-tuning as new data arrives. If not, stop — a fast, cheap "no" is a good outcome and frees budget for the next use case.
Planning your first or next AI initiative? Our AI Solutions team runs short readiness assessments that rank use cases by value and feasibility.