Start with a real piece of work.
Choose a recurring activity with a clear owner, a measurable baseline and enough volume to learn from. Understand the work before selecting a tool: who does it, what information is needed, where judgement matters and what happens when something goes wrong.
An AI business case is also a workforce design question. It should explain what the person will do differently, what a manager must check and whether adjacent teams can use the output.
An illustrative calculation, not a savings promise.
Suppose a team handles 200 requests a week. A pilot suggests that drafting assistance could save five minutes per request after allowing for review. That is roughly 16.7 hours of potential weekly capacity—not automatically a reduction in cost.
The case still needs to test those assumptions on representative work, include rework and oversight, and establish how the capacity would be used. Faster customer responses, backlog reduction or more time for complex cases may be more credible than claiming cash savings. These numbers illustrate the reasoning only; they are not a measured client outcome.
Specify the pilot's decision rules.
Agree what would justify progressing, adapting or stopping before the trial begins. Otherwise a compelling demonstration can become an investment commitment without enough evidence.
- Value: the business outcome and baseline, not just the tool's speed.
- Quality: the standard an output must meet and the cost of correcting mistakes.
- Responsibility: who reviews, overrides and owns the final decision.
- Adoption: the skills, workflow and manager support required.
- Economics: licences, integration, review time, training and ongoing support.
- Gate: the evidence needed before wider rollout, including unresolved risks.
Connect the pilot to an operating plan.
A focused review can produce a prioritised opportunity, a task-and-responsibility map, a value hypothesis and a pilot plan with success measures. Running the pilot or building a production system is a separate scope.
Bring business, technology and people leaders into the discussion early. A technical improvement only becomes useful business change when the surrounding work can absorb it.
The question is not simply whether AI can perform a task. It is whether the resulting way of working is useful, accountable and worth extending.