Five questions to answer before building an AI agent

Most AI agent projects do not fail at the model. They fail at the definition. A capable model pointed at an ambiguous job produces ambiguous work, faster and at greater volume than the humans it was meant to help. Before any interface is designed or any platform is chosen, five questions deserve written answers. If any of them cannot be answered clearly, that is not a reason to write a better prompt. It is a reason to pause the build.

1. What exact job is the agent responsible for?

Name the job the way you would for a new hire: the inputs it receives, the output it produces, and the business outcome that should improve. "Handle inbound inquiries" is not a job. "Acknowledge each inquiry within minutes, gather the four qualifying facts we always need, and hand a summarized, prioritized inquiry to the coordinator" is a job, and it implies its own measures: response time, qualification completeness, and coordinator time saved.

If the job cannot be described this concretely, the agent will inherit the ambiguity. Ambiguity in a person is coached over weeks. Ambiguity in an agent ships to every customer on day one.

2. Which sources may it use?

An agent answers from whatever it can reach, so the source list is a design decision, not an afterthought. Decide which documents, systems, and pages are approved; who owns keeping each one current; and what the agent should do when the answer is not in an approved source. The honest options are to say so, or to route the question to a person. Improvising from general knowledge is how a wellness clinic's agent starts giving quasi-clinical advice and a law firm's agent starts citing law it was never given.

Stale sources deserve equal attention. An approved-but-outdated price list is more dangerous than a missing one, because the agent will state it with confidence.

3. Which actions may it take without review?

Separate what the agent may do on its own from what it may only draft. Reading, summarizing, and preparing are usually safe to automate. Sending, publishing, promising, booking, charging, and changing records are consequential actions, and each one needs an explicit decision: autonomous, draft-for-approval, or off-limits. The right split is specific to the business and its risk tolerance, but the worst version is the unexamined default, where the agent can do whatever its integrations happen to permit.

A useful discipline is to write the permission as a sentence a manager could enforce: "It may send the standard scheduling links unprompted. It may draft, but never send, anything that mentions price."

4. What conditions require a person to step in?

Every real workflow has exceptions, and the escalation paths deserve design before launch, not after the first incident. Typical triggers include emotional or distressed language, questions with clinical, legal, or financial consequences, a customer asking for a human, contradictory information in the sources, and repeated failure to resolve the same request. For each trigger, name where the handoff goes, what context travels with it, and how quickly a person must pick it up. An escalation that lands in an unwatched inbox is not a control; it is a delay with paperwork.

5. Who owns the outcome?

Finally, name the person who reviews quality samples, handles exceptions, watches cost, and reports whether the business outcome is actually improving. Not a committee, and not "the vendor": a named person inside the business with the authority to pause or change the system. Agents drift: sources age, edge cases accumulate, usage patterns shift. Without an owner, drift is discovered by customers.

The pattern behind the questions

All five questions are one question in different clothes: is this agent part of an operating system, or is it a substitute for designing one? An agent with a bounded job, approved sources, explicit authority, designed escalation, and a named owner is a system component that can be trusted with more over time. An agent with none of those things is an unsupervised intern with perfect confidence and infinite patience. Answer the five questions first, and the build that follows is smaller, safer, and considerably more likely to earn its keep.