Most AI initiatives start with the wrong question. Asking “where can we use AI?” produces pilots. Asking “what capability do we need?” produces results.
Almost every leadership team we talk to has an AI initiative. Many of them started the same way: someone asked, “Where can we use generative AI?” The question sounds ambitious. In practice, it tends to produce a portfolio of pilots that impress in a demo and never reach operation.
The problem is not the technology. It is the starting point. When the tool is the starting point, every use case is a solution looking for a problem — and nobody owns the outcome.
The question is not “what can AI do?” It is “what does the organization need to be able to do?”
Start with the capability
A capability is something the organization can do repeatedly, reliably and at a known cost: closing a regulatory report on time with evidence, detecting a fraud pattern before it scales, onboarding a client in hours instead of days. Capabilities have owners, indicators and operating models. Tools do not.
When we work with a leadership team, we ask three questions before talking about any technology:
- What capability are you trying to build — and how would you know it works?
- What is preventing you from building it today: data, process, people, platforms or risk?
- What is the smallest piece of work that would prove value with real evidence?
Where AI actually fits
Take regulatory reporting. The capability is “every report validated, traced and approved on time.” AI can help: it can explain why a validation failed or draft the narrative of an exception. But the foundation is not a model. It is a catalog of rules, managed cut-offs and evidence that an auditor can follow. Put AI on top of a weak foundation and you automate the confusion.
That is why we treat AI as a multiplier of a capability, not as its substitute. Architecture, data quality and governance are what make the multiplier worth something.
A simple test for any AI pilot
Before approving the next pilot, check that it has:
- A business owner who will use the result — not only a technical sponsor
- A baseline, so improvement can be measured instead of claimed
- Access to the data it needs, with the right permissions
- A view of how it will be operated, monitored and paid for after the demo
- An evaluation method that includes its limitations
If any of these is missing, the pilot may still be interesting. It is just not yet a step toward a capability.
AI will keep changing faster than any roadmap. The capabilities your organization needs will change much more slowly. Anchor your strategy there, and let the technology serve it.