I keep having the same conversation.
A leader shows me the agent their team built. It's connected to email, the CRM, the ticketing system. It can do a lot. Then I ask one question: what's its job?
The answer is usually a list of tools. Sometimes it's a shrug and "it helps with stuff."
We'd never hire a person that way.
What we think we're doing
Here's what actually happens. We spend weeks picking the model and wiring up the connections, then about ten minutes on the instructions. The prompt reads like a wish. Be helpful. Be accurate. Handle the inbox.
Then the agent does something nobody expected, and we decide the technology isn't ready.
The numbers match the pattern. In June 2025, Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, pointing to rising costs and unclear business value (Reuters). That's a forecast, not a count. But read the second reason again. Unclear business value is what a missing job description looks like from the boardroom.
McKinsey's 2026 State of AI survey tells the same story from the other side. Among companies with more than $1 billion in revenue, 40% have scaled agents in at least one function, up from 27% a year earlier. Smaller organizations held flat at 22% (McKinsey). Plenty of agents get built. Far fewer get a real job.
Most agents don't fail because the model is weak. They fail because nobody told them what the job is.
That's no knock on the teams building them. They're under pressure to show something working by the end of the quarter, and a demo doesn't need a job description. Production does.
The same problem we've always had
What I've learned the hard way, with people long before software, is that unclear roles produce two kinds of failure. Some people freeze and do too little. Others fill the gap with their own judgment and do too much.
We never fully solved this for people. Gallup has reported that only about half of workers strongly agree they know what's expected of them at work (Gallup, 2016). It shouldn't surprise anyone that we're repeating the mistake with software.
Agents lean hard toward the second one. An agent with no boundaries will always find something to do.
I coached football for 15 years. A defense rarely breaks down because eleven players lack effort. It breaks down because one player left his assignment to go make a play that wasn't his. Great effort, wrong gap, touchdown the other way.
An agent without a job description is that player on every snap.
So before you write the prompt, write the job description. The same document you'd hand a new hire on day one. If you can't write it, you're not ready to deploy.
What goes in it
-
One outcome it owns. A single sentence that finishes "this role exists so that..." If you need the word "and" twice, you have two agents.
-
Decision rights. What it can do alone, what it prepares for a human, and what it never touches. My own inbox agent drafts replies. It never sends. That one line matters more than anything else in its instructions.
-
An escalation path. When it's unsure, what does it do, and who does it go to? "Use your judgment" isn't a path.
-
What done looks like. How would you know, on a Friday, that it had a good week? If you can't score it, you can't coach it.
-
A named manager. A person, not a team. Someone who reviews the work, corrects it, and answers for it.
Here's the part people miss: the tools aren't on that list. Tools are equipment. The job description is the assignment, and equipment without an assignment is how people get hurt.
Then coach it
The first version will be wrong. It always is with people too.
So sit down with it after the first month the way you would with a new hire. Where did it wait when it should have acted? Where did it act when it should have asked? What did you assume it knew that you never wrote down?
That last question usually stings a little. Most of what goes sideways traces back to something that lived in my head and never made it onto the page.
This isn't about controlling the agent. It's about giving it what every good employee wants: a clear lane and someone paying attention.
Read your agent's instructions back as if they'd been written for you. Would you know what to do on Monday?
Worksheet: Download the one-page Agent Job Description (PDF)
Sources: Reuters, June 25, 2025: Over 40% of agentic AI projects will be scrapped by 2027, Gartner says. McKinsey, August 25, 2026: The state of AI in 2026: On the road to ROI (1,719 respondents, fielded May 4 to June 8, 2026). Gallup, September 27, 2016: Do Employees Really Know What's Expected of Them?