Transformation
The pilot that never ships
A pattern we see in almost every enterprise: an AI pilot that succeeds on its own terms and never reaches production. The cause is usually decided before the pilot starts.
Ask a large company about their AI programme and you will often hear about a pilot that went well. Ask what happened next and the conversation gets vague.
We have now seen enough of these to be fairly confident about the cause, and it is not the technology.
Pilots are designed to succeed, not to transfer
A pilot gets a hand-picked dataset, an enthusiastic team, and a suspension of the normal rules. Procurement is waived. Security review is expedited. The three people who understand the process are all in the room.
Production has none of that. It has the messy data, the sceptical team, the full security review, and the person who has done this job for eleven years and has opinions. A pilot that succeeded by removing those conditions has not proved it can survive them. It has proved the opposite is untested.
The three questions to ask before the pilot
Who owns this in production, and have they agreed? Not the sponsor: the operational owner whose team will run it and whose number it will affect. If that person is not in the pilot, you are building something that will need a new home at the worst moment.
What is the integration path? If the pilot runs on an export and production requires a live connection to a system that takes nine months to get access to, that timeline is the real timeline. Better to find out in week one.
What is the threshold for going ahead? Agreed in advance, in writing. Otherwise the result gets interpreted after the fact by whoever has the strongest view.
Run the pilot in production conditions
Our preference is to skip the sandbox. Take a narrow slice of the real workflow, with real data, real permissions, and real users, including one who thinks the whole thing is a waste of time. Ship it small.
It is harder in week one and dramatically easier in month six, because there is no transfer step. The thing that worked is the thing that is running.
The one that works
The pilots that reach production share a shape: narrow scope, real conditions, an operational owner with skin in the game, and a stated threshold. None of that is about the model. All of it is decided before any code is written.