Start with the work, not the model

A useful AI initiative begins with a business task that is slow, repetitive or hard to scale. Naming a model or a tool before understanding the work tends to produce a demonstration without a clear owner or a reason to keep using it.

Choose a task where you can describe the current steps, the people involved, the information used and what a good result looks like. Examples might include finding approved guidance, sorting incoming documents or preparing a first draft of recurring commentary. The example is an opportunity to investigate, not a claim that every organisation should automate it.

Check whether the information is ready

Most AI use cases depend on the quality, permissions and context of the information around them. Before a pilot, ask where the source data lives, who may access it, whether it is current and whether it contains personal, confidential or regulated information.

If the source material is fragmented or unreliable, improving data access and governance may be the more valuable first step. A smaller, well-scoped set of approved information is often easier to evaluate than connecting every system at once.

Keep a person in the right part of the loop

Generative systems can produce fluent answers that still need checking. Decide which outputs may be used as suggestions and which require review before they are shared, stored or acted on. Make it easy for staff to see the source, correct an answer or escalate an uncertain case.

The level of oversight should reflect the consequence of an error. A draft internal summary and a decision that affects a customer or employee deserve different controls.

Run a small, measurable pilot

A pilot should answer a business question. Select representative examples, agree a baseline and define acceptance criteria before comparing options. Measure usefulness and failure modes as well as time or volume.

Record who owns the pilot, which information it can use, how feedback is reviewed and what would make the team stop. If it works, plan for support, access changes, monitoring and ongoing cost before broadening access.

  • Name one process owner and the people who do the work.
  • Use approved data and a clear evaluation set.
  • Include incorrect, ambiguous and out-of-scope examples.
  • Decide what needs human review and who handles exceptions.
  • Document a safe stop condition before expanding.

A sensible first conversation

An AI readiness discussion should leave you with a prioritised opportunity, the information and controls it depends on, and a small next step. Sometimes that next step is a pilot. Sometimes it is a data-quality fix, a workflow redesign or simply clarifying the process first.

That is a useful outcome: it helps the business invest attention where it can make a practical difference without assuming that AI is the answer to every problem.

All insights