Choose an AI marketing use case for a measurable pilot
Start with recurring work that a marketing team can describe and measure. Examples include preparing product descriptions, classifying customer feedback, drafting campaign variants and answering internal product questions. For each task, identify the current input, output, responsible person and decision the output supports. This creates a useful comparison between candidate pilots and makes the cost of checking AI output visible before a tool is selected.
Describe the current work with actual examples
Collect representative inputs and the outputs the team currently accepts. Include ordinary cases, incomplete information and difficult exceptions. Measure the time spent preparing information, producing the result and correcting it. Identify where the result enters another system or reaches a customer. A task described as content creation is too broad for evaluation; drafting a product description from approved specifications for editorial review is much easier to test. Record the language, product categories and people involved so the pilot covers the intended working conditions.
Estimate value from the work being changed
Estimate how often the task occurs and what portion of the work AI could affect. Include preparation, review, correction, software and integration costs. Faster drafting has limited value when approval remains the bottleneck. For customer-facing use, define the commercial mechanism: improved product coverage, fewer unanswered questions or faster release of a campaign. Avoid converting every saved minute directly into revenue. State whether the team will use released capacity for additional work, reduce external production or change staffing, and who can make that decision.
Check information access and maintainability
Identify the source of product facts, brand rules and customer information. Check whether that source is complete, current and available under the organisation’s access rules. Record who updates it and how a change reaches the AI workflow. Product questions involving price, stock or delivery promises may require a live system connection. A static document can become inaccurate between updates. Use the smallest relevant information set and define what the system should do when required facts are missing or conflict.
Evaluate the consequences of a wrong answer
List errors that would make the output unusable or harmful to the customer experience. A fabricated product property, incorrect promotion date or unsupported delivery promise needs a different response from an awkward sentence. Assign the reviewer and the evidence they need to check the result. NIST’s AI Risk Management Framework organises risk work around governance, context, measurement and management. Apply those questions to the selected workflow: who owns the decision, what can go wrong, how it will be observed and what action follows a failure.
Compare candidates using evidence and explicit judgements
For each candidate, record frequency, current effort, data readiness, error consequences, review effort and integration dependencies. Add the evidence behind each judgement. A qualitative table is often sufficient to choose the first pilot. If a team uses scores, define each scale and explain the weighting before calculating a total. A high business-value score should not conceal an unresolved data or approval requirement. Select a task that can be evaluated end to end with a named owner and a practical route to daily use.
Define acceptance before generating the first outputs
Set aside examples for evaluation and preserve the current process as the comparison. Agree quality criteria, unacceptable errors, total handling time and the cost of a completed task. Record the model, instructions, information sources and review procedure for each run. Inspect disagreements between reviewers and revise ambiguous criteria before scaling. Finish with a documented decision to expand, change the scope or stop. The use-case worksheet supports selection; the linked pilot guide covers evaluation and the transition into a wider implementation programme.
Working files
Sources
Apply this to your business
Bring your business question and current figures. We will define the work, the data required and the decisions your team needs to make.