Start With One Real Workflow, Not an AI Roadmap

Most companies do not need to be persuaded that AI can draft, research, summarise, or analyse. Their teams already use chat tools for some of that work.

The more useful question is where the work stops. Usually, a person still has to copy the result into another system, check the latest context, decide whether it is safe to act, and follow up when the first path fails.

Chat is already useful

That is good news. A company does not need to start from a blank page or a generic transformation programme. It can start from a task where people already see value in AI but still carry the operational burden around it.

Find the human router

Look for the person who moves information between a chat tool, email, WhatsApp, a spreadsheet, a CRM, or an internal system. They are often doing work that is valuable precisely because it requires judgement, but parts of the loop may be repeatable.

Ask what arrives, what they check, what they change, who they notify, and what makes them stop. That reveals the real workflow more clearly than a feature request.

Define one useful action

Choose one action with a visible outcome. It might prepare a complete customer update for approval, reconcile a request against the right source material, or route a valid exception to the right person with the relevant context.

The action should be narrow enough to test, but valuable enough that someone would notice if it worked reliably.

Keep the decision boundary human

The first system does not need to make every decision. In fact, it is stronger when it makes the boundary explicit: what it can prepare, what it can execute, and what must be approved.

This keeps the project useful without pretending that the organisation has delegated judgement it still needs to own.

What a diagnostic should produce

A serious diagnostic should leave the company with a shared view of the workflow, the systems involved, the available context, the action boundary, exception cases, and a smallest credible build.

That is enough to decide whether to build. It also prevents an AI project from becoming a vague promise with no operator, no proof point, and no way to learn from the result.