The first AI agent project should have a finish line that the business can recognize. “Help the sales team” is too broad. “Prepare an account brief from approved CRM records before a scheduled meeting” is specific enough to evaluate.

In its guidance on building effective agents, Anthropic distinguishes predefined workflows from systems that let a model choose its next steps. That distinction is useful when choosing a pilot: a process does not need more autonomy simply because an AI system can provide it.

Start with the outcome and work backward

Write down the deliverable, the person who will use it and the evidence that makes it acceptable. For an account brief, that could mean the correct customer, current contract status, recent interactions and links back to the relevant records.

Then identify the inputs. If employees already struggle to find reliable information, an agent may inherit the same problem. A pilot should expose those gaps early instead of compensating with plausible assumptions.

Use four selection questions

  • Can completion be checked? A reviewer should be able to distinguish an accurate result from a persuasive one.
  • Can access be limited? The initial workflow should operate with the smallest practical set of systems and records.
  • Can mistakes be corrected? Prefer work that can be inspected before it creates an external commitment.
  • Does it happen often enough? Repetition gives the team opportunities to learn and compare performance.

This is a proposed selection framework, not a universal ranking of use cases. A support organization and a software consultancy may reach different conclusions using the same questions.

An illustrative choice between two pilots

Suppose a small services business is choosing between autonomous lead outreach and preparing meeting briefs. Outreach includes recipient selection, wording, sending and follow-up. A mistake can reach someone outside the business before a reviewer sees it.

A draft meeting brief can remain inside the team until approved. That makes it a more manageable first experiment if the immediate goal is to learn whether the agent can use company information reliably. It does not mean outreach automation has no value; it means the learning objectives and consequences differ.

Define the pilot before choosing the product

Record a baseline for the current process: time spent preparing the result, time spent checking it and the common errors. Run the proposed agent on comparable cases and record the same measures. Include difficult inputs, such as a missing customer record or conflicting information.

Assign an owner who can pause the workflow and decide what happens when it gets stuck. Set an explicit expansion rule, such as demonstrating reliable drafts before allowing any external action.

A good first project should produce either a useful workflow or a clear reason not to expand it. Both outcomes are more valuable than a polished demo with no agreed measure of success.