10 guides for choosing and running an AI pilot
Ten guides for choosing, testing and assessing an AI pilot.
Fund an AI pilot when someone owns the workflow, you can describe its current cost or performance and a result could change your decision about investing further. If the team cannot name the owner or show how it will judge the result, defer approval until it can.
Working hands-on with current AI tools has changed where I look for opportunities. The starting point for an AI opportunity session is a workflow and an owner. These ten guides help with different decisions. Start with the one that is holding up your pilot.
1. Shadow AI
Start here if employees already use unapproved AI tools for the work under discussion. The guide to shadow AI separates a missing approved option from confusion about the rules. Ask employees to show you the task they were trying to finish and what the approved process made difficult. A survey from another company or market cannot tell you why your team chose its tools.
2. AI use case prioritisation
Use the prioritisation guide when you have several requests and need to choose which business problem to fund. Compare the result each team wants to improve and ask the workflow owner what they would change if the test succeeds. If nobody can act on the result, name an owner before comparing the proposal with those that already have one.
3. Current AI capabilities
The guide to testing current AI capability helps when a decision rests on a demo or an old chatbot trial. Pick a task the team does today and test it with the information the team could actually provide. Record what a person still has to correct or finish. The result applies to that task and setup.
4. AI business case
The business case guide helps you decide what the trial needs to measure before anyone estimates its return. For a support workflow, bring the current time spent on cases and say what the team would do with the hours it expects to recover. An estimate of hours saved leaves the financial effect open until you know what changes in the work.
5. AI experimentation
The experimentation guide is for teams that want people close to the work to try ideas within agreed limits. Decide what someone can try in their own workflow and what needs wider approval. A trial using sample material and a trial that touches live customer records raise different questions about who needs to be involved.
6. AI acceptable use policy
The acceptable use examples help when employees cannot tell which tools and data the trial permits. Name the tool and one type of input the team expects to use. Write down whether staff can proceed, need approval or must keep that input out. The wording needs to work for the employee making that choice.
7. AI impact assessment
Use the impact assessment examples to document what the system can access or change and who acts when it fails. Write the fallback in terms of the actual role and system. A published example can show how to record the answers. Your team still has to supply its own and check which formal assessment duties apply.
8. AI vendor assessment
Read the vendor assessment guide when the proposed pilot depends on an outside agent and a vendor's assurance. Compare the demo with the terms that apply to your company. Check what the vendor and its model providers can do with your data, including whether either can use it to train models. Mark the unanswered questions before relying on the agent for live work.
9. AI training for employees
The employee training guide helps you plan what people must practise before the pilot expands. Give participants a task from their job, then have them explain how they checked the output and applied the company's rules. This also gives you a way to assess a supplier's proposed exercises.
10. AI training and readiness
The training readiness guide helps managers decide who can take responsibility for the work when the pilot ends. A completion record can show who finished the assigned training. Before handing someone a task where mistakes affect other people, review a sample of their work and decide what they can do independently.
Sources
- Shadow AI: what the surveys count, and what to do about it
- AI use case prioritisation: settle the owner and the measure
- Your AI strategy may be based on an outdated experience
- AI business case: separate the evidence from the assumptions
- How companies can make room for AI experimentation without losing control
- AI acceptable use policy: five decisions, with real examples
- AI impact assessment templates: Canadian assessments and UK transparency records
- AI vendor assessment: what to get in writing before you sign
- AI training for employees: what it should cover
- What does completed AI training tell a manager?
Sources checked on 4 November 2026.