McKinsey - AI agents need performance management
Companies are deploying AI agents in live workflows and supervising them less rigorously than they supervise a graduate hire.
A person in a process has a manager, a review cycle, an escalation path, and a record of decisions. An agent doing the same task often has none of that. It runs, it acts, and the first anyone hears of a problem is when a customer complains, or a number looks wrong.
Why the gap exists is worth naming, because none of the reasons is irrational on their own:
- An agent looks finished from day one. A graduate visibly learns the job, so scrutiny feels natural. An agent's output reads as polished code or a confident answer, and that polish gets mistaken for competence.
- Ownership is undefined. Onboarding a graduate triggers a manager assignment, a review cycle, a probation period, almost by default. Deploying an agent usually triggers none of that, because no one's role description says "supervise this.
- Deployment outpaces governance. Agents go into production at the speed of a sprint, not a hiring cycle, so oversight structures don't have time to catch up before the next agent ships.
- Agents don't generate the signals managers rely on. A graduate asks questions, makes visible mistakes, and shows uncertainty, and all of that cues a manager to step in. An agent tends to fail silently, so the absence of visible friction reads as an absence of risk.
There is no accumulated intuition for AI reliability yet. Everyone in a management role has been a junior employee once and has some feel for where a graduate's judgment will fail. Almost no one has that same feel for an agent's failure modes, so oversight defaults to none rather than to a calibrated amount.
McKinsey's argument is that agents need performance management as a discipline. Clear objectives. Monitoring against them. A named owner. A way to catch drift before it compounds. This is not treating software as human. It is recognising that anything making decisions inside your business needs the same accountability structure you would demand of a person.
The harder organisational question sits underneath it. If an agent handles a task end-to-end, who is responsible when it fails? The team that deployed it, the function that owns the process, or the vendor. Right now, that answer is usually unclear, which means it is effectively nobody.
Before the next agent goes into production, name its manager.