Scope an AI agent that can work in production

Replace ‘we need an AI agent’ with a workflow someone can demonstrate. For example: when a client brief arrives, extract requirements, find approved inventory information, and draft a plan for a media planner to review. Define where the workflow starts, where it ends, and what the reviewer receives.

Record the current volume, handling time, quality standard, and recurring exceptions. Separate observed baseline data from estimates. A useful first build should make one job easier to complete and easier to check. A broad promise to automate a department does not tell a builder what done means.

Some jobs need a fixed workflow with a model handling one step. Others need an agent to choose tools and steps as it learns more. Anthropic distinguishes these patterns and recommends increasing complexity only when it improves the outcome. Ask a prospective builder to explain why the workflow needs agentic decisions.

Write down the ordinary path, the ambiguous path, and the failure path. Consider a copilot that drafts for a person when the action requires judgment or carries a meaningful cost. Consider multiple agents only when separate responsibilities create a measurable benefit after accounting for coordination, latency, and cost.

List each system, the records the agent needs, the authoritative source, and the permission owner. Separate reading, drafting, and changing records. Define which actions can happen automatically and which require a person to review the exact proposed change. Approval should attach to an understandable action such as sending a client email or changing a budget.

Include stale records, contradictory instructions, unavailable tools, and untrusted text in the scope. Ask how the system checks permissions outside the model, prevents duplicate actions when work retries, and records what changed. A model’s confident answer does not verify that an external update succeeded.

Collect representative examples with permission to use them. Include routine work, ambiguity, missing data, permission failures, and cases where the correct result is to stop. Describe the expected business outcome and any unacceptable action for every case. The CSV template provides a blank structure and clearly labeled illustrative cases.

Anthropic’s evaluation guidance distinguishes a run’s conversation and tool activity from the resulting state. Apply that distinction when reviewing a build: a convincing explanation is weaker than a correctly updated record or a draft that satisfies the brief. Combine automated checks with human review where quality requires judgment, and repeat trials when behavior varies.

Choose launch thresholds for task success, unacceptable errors, human rework, latency, and cost with the business owner. Keep a separate evaluation set for the release decision, and record the system version tested. A demo using three handpicked examples cannot establish reliability across a working queue.

Request separate estimates for discovery, implementation, integrations, evaluation, and ongoing support. Recurring cost should include model use, tools, hosting, human review, monitoring, and maintenance. Ask what happens to unit cost when a case needs several attempts or a provider changes behavior.

Estimate cost per accepted task, including rework, and compare it with the current process. Treat time saved as capacity until you can show how the business uses it. Document who handles failures, who can change the system, and who owns the code, data, accounts, and operating instructions at handoff.

Begin with representative data in a controlled environment. Move to draft-only or shadow operation where appropriate, then a limited live scope with a named owner. Review the results before expanding volume, integrations, or write permissions. Each expansion changes the job the system is trusted to do.

The worksheet ends with a release decision, rollback route, support owner, and review date. Those are part of the product you are commissioning. They give your team something concrete to operate when the build team finishes.