Skip to content

AI pilots that never go live: start with returns, not technology

Many AI experiments get stuck in the pilot phase. That is rarely down to the technology, but to the choice of process. Start with one well-defined task with a clear return.

The problem

Companies are experimenting with AI on a large scale, but few AI pilots become a permanent part of the work. Gartner expected at least 30% of generative AI projects to be abandoned after the pilot phase, due among other things to poor data, inadequate risk controls, rising costs or an unclear return.

Why it keeps happening

  • The technology comes first. "We need to do something with AI" instead of "this process takes too much time".
  • Starting too broad. An assistant that has to do everything ends up doing nothing well enough.
  • Not reliable enough. A demo works, but in real work the mistakes are too expensive.
  • Not embedded. The application sits next to the work instead of inside it.

Our approach

  1. Returns first. Which process demonstrably costs time or money? What is it worth if it becomes faster or better?
  2. One well-defined task. For example, answering questions about your own documents, or extracting data from invoices. Small enough to get right.
  3. Make it reliable. Let code do the calculations, have answers checked, report doubt honestly, and let people make the decision.
  4. Put it into the work. The application goes into the systems people already work in, not into a separate tool.
  5. Only then expand. It is less about the code and more about good choices: what do we build, and what do we not build?

What to watch out for

  • Measure before and after. Without a baseline, you cannot demonstrate a return.
  • Choose data that is in order, or put it in order first.
  • Keep costs predictable, for example with local AI at fixed hardware costs.

Would you like advice on AI pilots that deliver a return? Request a free advice conversation.

Sources

Updated on

AI

Personal data in the AI cloud: who is watching?

A conversation with an AI service in the cloud is not private by default. A case in Canada shows what a provider can see and share. What does that mean under the GDPR, in the United Kingdom and in Switzerland?