A summary of my article in CIO.
I wrote an article for CIO about a mistake many companies make with AI agents, Pearl included. An agent gets access to company systems and is expected to figure things out, with nobody explaining the job, the priorities, or who to ask. When it can’t, the project stalls. Nobody would onboard a new hire that way.
The numbers back this up. Gartner found that while 40% of C-level executives say their organizations have operationalized AI, only 13% have achieved measurable and scalable results.
We learned this the hard way with Work Radar, our internal agent that surfaces missed tasks and summarizes what happened while you were on PTO. We connected it to Outlook, Teams, Confluence, Asana, and our HRIS. It had all the data and still flailed. It couldn’t tell strategic priorities from tactical noise, so a routine status update got the same weight as a critical business deliverable.
The fix was unglamorous: we onboarded it like a new employee. We gave it a single page covering personal context, operating level, and explicit priorities. After that, it stopped flailing.
Since then, we’ve run agents using five management principles:
-
Role — Assign governance and a supervisor. If no person owns the output, nobody owns the failure.
-
Remit — Before expanding permissions, weigh the agent’s sphere of influence against the blast radius.
-
Personality — Write down its permanent operating context instead of letting the model guess from the prompt. For Work Radar, that was the one-page brief.
-
Performance — Evaluate its work against concrete examples of acceptable and unacceptable output.
-
Teamwork — Decide how it works with other teams and other agents.
None of this is new. It’s the same management basics we apply to people, pointed at a new kind of coworker.
The full article in CIO goes deeper on each principle.