How to Write an AI Agent Job Description Before You Buy the Tool

AI conversations often start with a product name. Should we use ChatGPT? Should we try Claude? What is everyone else using?

Those are fair questions. They are just early.

The first question is about the work. What keeps repeating, and what kind of work is it? A task might need judgment, accuracy or just consistency. AI fits consistency best, and that is where a small team loses the most energy when nobody owns the step.

One company already hires AI the way it hires people. On The AI Why podcast (August 6, 2026), OneDigital's chief product officer, Vinay Gidwaney, described an internal staffing process for AI coworkers. "Everything starts with a job description," he said. "So we write a job description for the AI."

A small team can borrow the habit with one page.

Here is the six-line template. Fill in each line for one workflow.

A prompt asks for an answer. A role sits inside the business. It knows what starts it, what it can see, where it stops and who reviews the result.

If a line is blank, the tool has to guess. When output drifts and nobody owns the review, people blame the model. Usually the job was never written down.

Keep the job description in its own document, apart from any tool. Models change. When you switch, the rules, examples, boundaries and corrections come with you.

The operating environment says how the role runs inside your business. Write down five more things:

Here is what that looks like for a consultant's discovery-call follow-up. The trigger is a call marked complete. Access is the call notes and the approved follow-up templates, read only. Escalation is any mention of price, scope changes or a complaint. Evaluation is a few past follow-ups you were happy with. Correction is a running list the owner updates when a draft misses.

The correction log matters most over time. Each fix the owner records stays with your business, whatever model runs the role next year.

Gidwaney sorts generative AI into three kinds of work, according to the same episode:

OneDigital started with co-workers. Gidwaney said most companies "do that in reverse order. They start with the agents." He added that the setup an agent needs, from clean data to deciding where people step in, is hard to get right.

For a founder, coach or small agency, the safer first role helps a capable person do better work:

A person still decides when something is unusual. That is the design, and it is how a role earns more responsibility. The episode's title names the same path for an AI coworker: intern, then apprentice, then full time.

Some calls should not move to AI on a busy afternoon. Put these in the Boundaries line of every job description:

AI can still help with each one. It can draft the reply, find what's missing and suggest the next step. The owner makes the call.

Our tagline says it in four words: Automate Consistency, Not Decisions. AI should remove repetition, not responsibility.

Harvard Business Review published research on May 6, 2026, titled "Research: Why You Shouldn't Treat AI Agents Like Employees," by Matthew Kropp, Julie Bedard, Emma Wiles, Megan Hsu and Lisa Krayer. HBR's published summary says some organizations are placing AI agents on org charts as employees. In a large-scale experiment, that human framing reduced individual accountability, increased unnecessary escalation, lowered review quality and left employees less sure of their roles, without improving adoption. The summary closes on redesigning workflows, roles and governance so "humans remain clearly accountable."

We read that as a guide to which half of the idea to keep.

Keep the parts that make work checkable: a written purpose, clear inputs and outputs, boundaries, a named owner and real examples the role must pass.

Drop the parts that blur who is responsible: the human name, the box on the org chart and any sense that the tool answers for the result.

A person owns the outcome. The job description makes their review quick enough to happen.

"Hours saved" is the easiest number to report and the easiest to game. The OneDigital guests made the same point on the podcast. Set a target for hours saved with AI, one of them said, and "you're going to get a certain type of behavior out of that. People will cut corners."

Ask what the team did with the time that came back:

Pick one of those before the role goes live. Then check it after a month.

Copy this for one weekly workflow.

If you can't fill in Boundaries or Owner, hold off on the next tool. You would be choosing software before choosing the work.

Write the AI agent job description before you buy the tool. Six lines define the role and five more define how it runs. Start with a co-worker, keep promises, pricing and disputes with a person, and measure what the freed time produced.

What is an AI agent job description? A short written role for one AI task. It states the purpose, inputs and outputs. It also names the decisions the AI may never make, the person who checks its work and the real examples it must pass before it takes on more.

How is an AI agent job description different from a prompt? A prompt asks for one answer. A job description defines a role inside the business: what starts it, what records it can see, where it must stop and who reviews the result. The same job description can move to a new model when you switch tools.

Should I start with an AI agent or an AI assistant? Start with an assistant, or what OneDigital's Vinay Gidwaney calls a co-worker: AI that helps a person think, prepare or review. Move to a background agent only after the role passes its tests and the owner trusts its output on real examples.

Should AI agents be treated like employees? Borrow the structure, not the identity. Harvard Business Review research published May 6, 2026 found that treating AI agents as employees reduced individual accountability and lowered review quality. Keep the written role and the named human owner. Skip the human name and the org-chart slot.