Which AI ideas get done

AI demand intake is the way a company collects requests for AI solutions and decides which of them get built. Prioritization is the second half: choosing the few that are worth scaling, so that they do not stay pilots.

Why does a company need one way in for AI ideas?

In one week in November 2026, I gave four talks about AI, and the questions people asked sat at four different steps of the same path. Students asked about language models. A bank asked how to track the value of AI. A sales team asked which documents they may upload to AI. Managers in an executive program asked how to keep AI engineering documentation up to date.

Instead of tackling every question separately, people need a clear framework and a guided path, depending on where they are in their AI journey. A company should give every AI question its single, right next step.

What helps at the start of demand intake?

In demand intake, it has helped me to first collect the AI requests that already exist: who asked, and which problem it should solve. It showed us quickly where the value is. So I would recommend starting with such a collection as well.

What is worth checking before an AI pilot is scaled?

In January 2026, I put together a checklist for the moment after the first demonstration to users and before the decision to scale. It has eight points:

• Strategic pull: the idea aligns with a business priority, and the owner of that priority confirms it.
• Executive accountability: a senior sponsor is named and has committed budget or time.
• Steering: decision rights are defined, and so are the criteria to scale, pause or stop.
• User pull: a committed user champion exists, and users are asking for the solution.
• Data: access is granted, the data owner cooperates, and a first data review is done.
• Guardrails: security and privacy controls are defined and can be implemented.
• Measures: a baseline is captured, so that adoption and impact can be compared before and after.
• Operating model: ownership after launch is clear, and the capacity for further iterations is protected.

The odds of a pilot are often visible before scaling starts.