Choose a repetitive task with clear boundaries.
Your first step should not be to automate the entire business. Choose a specific task, such as categorising enquiries, finding information in documents or drafting responses. Describe how the work is done today and where time is lost. Consider whether conventional software or rule-based automation could handle a simple task whose rules do not change.
Identify data sources and access limits.
List the files, records and systems the solution will need to access. Clarify who may use that information and whether it can be shared with an external service. Outdated documents and conflicting records can affect results. Using only the examples needed for the pilot keeps its scope clear and manageable.
Define what a successful output looks like.
Replace a broad requirement such as “give the right answer” with a sample task and an expected result. For classification, success might mean selecting the correct category; for document search, it might mean finding the relevant source. Include missing information, ambiguous requests and out-of-scope questions in your evaluation. Test varied scenarios so your quality assessment does not rest on a few successful examples.
Decide where human approval is needed.
Drafting a response and sending it to a customer automatically require different levels of authority. Identify approval steps according to the impact of each action. Plan what should happen when the system cannot find enough information and how an action can be stopped. Assigning someone to review early outputs helps reveal errors and gaps in scope.
Assess the full cost of your pilot.
Alongside model or service fees, include integration work, data preparation and human review time in your assessment. Compare the time required before and after the pilot. If quality and oversight requirements meet your objectives, expand gradually. Focus on tasks where the benefit can be measured, rather than using AI for every task.
Before you begin: a checklist
- Are the task and its exclusions clearly defined?
- Are data sources and access permissions agreed?
- Do you have sample scenarios for evaluating quality?
- Which steps require human approval?
- Have usage costs and review time both been included?