How to write a job description for an AI employee

September 28, 2026 · 5 min read

Most founders type a paragraph into a prompt box and call it a job description. Then they're surprised when their AI employee goes off in the wrong direction. The fix isn't a better AI — it's a better brief.

The prompt-versus-job-description gap

A prompt says: "You are a marketing assistant. Help me with marketing tasks." A job description says: "You own top-of-funnel content. You produce two long-form posts per week and one cold outreach sequence per month. You do not touch paid ads — that stays with the founder. When a task requires budget approval over $200, you flag it and wait."

These are not the same instruction at different lengths. They are different types of documents. A prompt sets tone. A job description sets authority, scope, and accountability. An AI employee operating from a real job description knows what it owns, what it doesn't, when to act independently, and when to stop and ask.

The result isn't incremental improvement. Founders who make this switch typically describe it as "night and day" — the AI stops second-guessing obvious decisions and stops silently overstepping on ones that actually matter.

The four things every AI job description must answer

You don't need a 10-page HR document. You need clear answers to four questions:

• What department does this role belong to? Marketing, Sales, Engineering, Finance, Design, Support, Operations, or COO/Chief of Staff. This isn't just taxonomy — it determines what resources the employee draws on and what other employees it can ask for input.
• What does this role specifically own? Verbs help. "Writes weekly SEO posts" is better than "content." "Monitors support inbox and drafts replies within 4 hours" is better than "customer support." Specificity tells the AI what proactive action looks like, not just how to respond when asked.
• What does this role explicitly NOT own? This is the section most founders skip and the one that matters most. If your AI Marketing employee doesn't know it shouldn't be revising the pricing page or contacting press, it will do those things with confident, helpful energy. Hard negatives prevent well-intentioned drift.
• When should it escalate vs. act? The escalation rule is the single biggest lever on how autonomous the employee feels. "Draft and publish blog posts independently; flag any post that names a competitor before publishing" is a real rule. "Always ask before doing anything big" is not — "big" is undefined and you'll become the bottleneck.

Scope creep is a job description problem

When founders complain that their AI employee "does too much" or "goes off script," the root cause is almost always a missing boundary, not a model problem. The AI is doing what any capable employee does when given a vague mandate: it fills the white space.

The solution is to write scope negatively as well as positively. "You own X, Y, Z — you do not own A, B, C" is a complete scope statement. "You own X, Y, Z" leaves A, B, and C as implicit gray zone.

Escalation rules: the difference between a tool and an employee

A tool executes instructions. An employee exercises judgment — but within a defined authority band. The escalation rule defines that band.

Good escalation rules are specific and testable. Examples:

• "Send outreach emails to the approved list without asking; add new domains to the outreach list only after flagging them."
• "File engineering tickets and move them through the board; do not close tickets marked 'needs founder review' without explicit sign-off."
• "Respond to support inquiries under $50 in refund value independently; escalate refund requests above that threshold."

Each of these tells the employee where it can move fast and where it needs to pause. That's the structure that makes delegation feel safe rather than nerve-wracking.

Maintain it like a living document

The first version of your job description will be wrong in specific ways you can't predict yet. That's fine. What matters is that you treat it as a living document — when the AI makes a decision you didn't want, the reflex shouldn't be "the AI is bad," it should be "what was missing from the job description that made this seem like the right call?"

Most founders who iterate this way find that their job descriptions stabilize after three or four revisions. After that, the employee runs largely on autopilot with occasional updates when the business changes.

The companies that get the most out of AI employees aren't the ones with the cleverest prompts. They're the ones who've done the work of actually defining what the role owns.

Frequently asked questions

How long should an AI employee job description be?

One to two pages is ideal — long enough to cover scope, escalation rules, and hard boundaries, short enough that you'd actually maintain it. If you find yourself writing five pages, you're probably solving a clarity problem you haven't consciously faced yet. Start with one page and add specifics as real edge cases surface.

Do I need a separate job description for each AI department?

Yes. An AI Marketing employee and an AI Finance employee have completely different scopes, different escalation paths, and different definitions of "done." Sharing one vague description across departments is like hiring one person to run both your ad campaigns and your books. The overlap in how you manage them is real — the overlap in what they own is nearly zero.

What's the single most common mistake founders make in AI job descriptions?

Skipping the "what this role does NOT own" section. Without explicit boundaries, an AI employee will fill the vacuum with its best guess — which may be reasonable but is almost never what you wanted. Hard negatives are more valuable than soft positives when it comes to scope.