November 5, 2026 · Jermaine Barker
Agentic AI Needs a Job Description Before It Gets Access
Before an AI agent touches your systems, define its role, authority, and limits like you would for a new hire. A practitioner's guide to scoping agentic AI for production.

Most teams I talk to about agentic AI start with the same question: what can it do?
Wrong question. The right one is: what should it be allowed to do?
An agent is not a chatbot. It takes actions. It reads records, sends messages, updates systems, triggers workflows. That makes it closer to a new employee than a piece of software. And nobody hands a new employee admin credentials on day one with no job description.
Yet that is exactly what I see.
The pattern I keep running into
A team gets excited about an agent. They connect it to email, the CRM, a shared drive, maybe a ticketing system. It works in the demo. Everyone nods.
Then someone asks a simple question: what happens when it's wrong?
Silence.
In every engagement where this comes up, the technology isn't the gap. The scoping is. Nobody wrote down what the agent owns, what it can touch, and when it has to stop and ask.
Write the job description first
Before any agent gets access, I ask the client to write a one-page role document. Plain language. No jargon. It answers five questions.
1. What is the one job? Not five jobs. One. "Triage inbound member requests and route them to the right team" is a job. "Help with operations" is not.
2. What does it read? List the specific systems and data. If you can't list them, you aren't ready to connect them.
3. What can it change? Reading and writing are different risk classes. Start read-only. Earn write access.
4. When does it escalate? Define the triggers. Low confidence. Sensitive data. Dollar thresholds. Anything involving a regulated record. The agent should know when to hand off, and the handoff should be easy.
5. Who owns it? A named person. Not a committee. Someone who reviews its work weekly and can shut it off.
If you can't answer all five, the agent isn't ready for production. That's not a failure. That's the process working.
Why governance is what lets you say yes
I hear the objection often: governance slows things down.
It does the opposite. Leadership approves what it can understand and bound. A scoped agent with clear limits, an audit trail, and an owner is an easy yes. A vague "AI assistant with broad access" is an easy no, and it should be.
Guardrails are not the brake. They are what gets the project through the approval meeting.
This matters most in healthcare and the public sector. When the data is protected and the stakes are real, you need to show an auditor or a board exactly what the system can and cannot do. A role document gives you that on one page.
Start narrow, then widen
The teams that succeed with agents don't begin with a sweeping vision. They pick one workflow, one agent, one owner. They run it read-only or with human approval on every action. They measure it for a few weeks.
Then they widen the scope, one permission at a time, based on evidence.
That is how you ship something real in 90 days. Not by building the most capable agent. By building the most clearly bounded one.
What to log
Even a well-scoped agent needs a paper trail. At minimum, record:
- What it was asked to do
- What data it accessed
- What action it took
- Whether a human approved it
- The outcome
When something goes sideways, and eventually something will, this log is the difference between a quick fix and a long investigation.
The takeaway
Eighty percent of AI pilots never reach production. Many of the ones that stall are agents nobody scoped. They were impressive in a demo and impossible to approve.
Treat your next agent like a hire. Define the role. Limit the access. Name the manager. Review the work.
If you want a clear read on whether your workflows, data, and governance are ready for agents, start with the free AI Readiness Assessment. And if you already know the workflow you want to tackle, our ASCEND framework is built to take one scoped use case to production in 90 days.
Discipline over enthusiasm. Every time.