AI implementation · 28 September 2026 · 4 min read

Before you add AI, map the decision.

A team asks whether AI can speed up a process. My first question is: which part of that process is actually a decision?

Take an incoming customer request. An AI tool might summarise the message, find the relevant policy and draft a reply. Those are different activities from deciding whether to grant an exception. If the boundary is vague, people may treat a draft as an approved answer or assume someone else checked it.

So I map the path before choosing a tool. What information arrives? Where does it come from? What rule or judgement is applied? Who has authority to make the decision? Who carries out the next action? What happens when the information is incomplete or the case is unusual?

Make ownership visible

One person may own the AI system, another the customer process, and a third the final decision. That can work well if the handovers are explicit. It becomes difficult when a tool produces an answer that nobody knows they are responsible for reviewing.

For a pilot, I define what the system may suggest or do, when a human must intervene and where an exception goes. The escalation route needs a named owner and a time expectation; otherwise an automated step can simply move a queue from one place to another. I also decide what to record: relevant input, source material, version or configuration where it affects the result, review, action and outcome.

This record has a practical purpose. If an important decision is questioned later, the team should be able to reconstruct what led to it. It helps them fix mistakes, compare outcomes and improve the workflow.

Test the narrowest useful change

Choose a single step with a clear owner and measurable outcome. For instance, test whether a tool can prepare a complete case summary for a human decision. Compare time to decision, missing information and the number of cases sent back for clarification. Review examples where the tool failed as closely as those where it worked.

Only then decide whether to expand the tool's role. A useful implementation connects the tool to the work people already do and makes the decision process clearer, rather than hiding it behind a faster interface.

All insightsTell me about it