How to write an AI task contract your team can test

An AI task contract describes a job in terms another person can verify: its trigger, input, permitted actions, expected output and completion evidence. It is an operating specification, not a legal contract. Writing one prevents a vague instruction such as handle my follow-up from becoming an undefined delegation.

Use a concrete request

Consider a hypothetical weekly customer check-in. The task begins when a manager selects an account. The assistant reads the approved account notes, drafts a short update and saves it for review. It does not send the message or change the account stage unless those actions are separately authorized.

This is more testable than telling an agent to be proactive. It defines where work starts and which decisions remain with the manager.

Copy this task brief

• Trigger and owner: who requests the task, and what event starts it?
• Inputs and sources: which records may be read, and which version is current?
• Allowed actions: what can be prepared, changed or sent?
• Output and approval: what must the reviewer see before deciding?
• Completion and recovery: what proves success, and what happens after a failure?

Add an explicit stop condition. For this example, a missing account owner or conflicting contact identity means the assistant returns a partial draft with the gap marked. It should not guess a recipient to keep moving.

Turn each boundary into a test

Try an account with complete notes, one without a contact and one with an outdated policy attached. A passing system produces the expected draft in the first case and explains the limitation in the others. Check the destination record, not only the assistant's claim that it saved the draft.

Have the process owner review those outcomes. Technical correctness alone cannot decide whether a customer-facing draft is useful.

Keep the contract small and current

Version the brief alongside the workflow. When responsibilities change, update the contract before expanding permissions. A one-page contract that the owner understands is more useful than a long prompt nobody revisits. Its purpose is to make responsibility visible, especially when the person who configured the agent is unavailable.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.