A practical way to turn a broad AI ambition into a first build that someone can use.

A miniature retail district with a purposeful route linking storefronts and a field-planning notebook.

A thought to take with youWrite down the decision, its owner and the evidence they need before choosing a model.

Find the moment that matters

“We want to use AI” opens a conversation, but it does not yet define a project. A more useful starting point is a recurring decision: which store needs a visit, which request needs escalation, or which document answers a colleague’s question.

Choose one moment and follow it through the business. Who makes the decision? What do they look at? What happens when the information is missing? The answers give a team something concrete to design around. They also make it easier to recognise when a simpler integration or interface would solve the problem.

Follow the evidence

Take a field manager preparing the next day’s visits. The useful output may be a brief with a small set of priorities, the reasons behind them and links to the supporting reports. A fluent summary alone would leave the manager with the same verification work.

Sketch that brief before deciding how to generate it. Identify which inputs already exist, which need cleaning and which depend on local knowledge. Keep unknowns visible. The design should make it possible to inspect a recommendation and change course.

Define what can happen next

Reading information and changing a business record are different responsibilities. Decide what the system may prepare, what it may execute and where a person needs to approve. Put the owner of the decision into this discussion early.

For the first build, a reviewed recommendation may be more useful than an automatic action. It lets the team test whether the information is relevant while keeping the final decision in a familiar workflow. The boundary can change when there is evidence to support that change.

Test usefulness with real work

Agree on a small set of representative tasks and difficult cases. Include missing records, conflicting information and questions the system should decline to answer. Ask the people doing the work to review both the output and the experience of reaching it.

Choose measures that match the decision. Time spent preparing a brief, the usefulness of the supporting evidence and whether the next action is clear are different signals. Avoid treating an attractive demo as proof that the whole workflow is ready.

A useful first release should answer one business question well enough to inform the next investment. That gives the roadmap a direction grounded in use, rather than a list of technologies waiting for a purpose.

From a question to a working system

Have a similar decision in front of you?

Talk it through with us