Why most AI pilots never reach production

Most organisations we meet have already run an AI pilot. Far fewer have one running in production. The pilot is rarely the problem — the gap between a demo and a dependable system is.

A demo optimises for the happy path

A pilot is judged on whether it produced an impressive answer in a meeting. A production system is judged on what it does at 2am when an upstream API returns malformed data, when a user pastes 40,000 words into a box sized for 400, or when the model confidently invents a policy that does not exist.

Those failure modes are not model problems. They are the ordinary engineering concerns of any system that touches real users: input validation, timeouts, retries, observability, and a clear answer to “what happens when this is wrong?”

Three questions worth answering before you build

1. What is the cost of a wrong answer? A drafting assistant and a system that approves claims sit at opposite ends of this scale and deserve very different guardrails.
2. Who reviews the output, and how quickly? If the answer is “nobody”, the system needs to be far more conservative than one with a human in the loop.
3. How will you know it degraded? Model providers update models. Your data drifts. Without evaluation running continuously, quality erodes invisibly.

Start with the boring parts

The teams that ship are the ones that build the evaluation harness before the clever prompt, and instrument the system before scaling it. It is less exciting than a demo. It is also the difference between a pilot and a product.