What AI agents cannot do yet, and what that means for the operator

The honest version of the AI-agent pitch includes what the agents cannot do, because the operator deciding how to deploy them needs to know where the human stays in the loop. The list is shorter than the cynical version of the pitch suggests and longer than the marketing version admits.

The agents cannot make strategic decisions. Capital allocation, pricing strategy, market entry, accounting policy choice. Each of these requires understanding the company's specific situation and choosing among trade-offs the agent does not have visibility into. The agent runs the operations under the policy the human chose. The agent does not choose the policy.

The agents cannot make customer judgment calls. Whether to give a difficult customer a refund, whether to grant a contractual exception, whether to escalate a complaint to leadership. The agent gathers the context, drafts the response, and lands the package on the human's desk. The human reads, decides, and ships.

The agents cannot make compensation decisions. The band library is policy. The mapping from policy to a specific person's offer is judgment. The agent surfaces the policy, drafts the offer, and the human approves. The same applies to performance reviews, promotion calibration, and equity grants.

What the agents do well is everything else. The data entry, the reconciliation, the document chase, the ticket triage on the runbook-trivial cases, the deck assembly, the calendar tracking on compliance deadlines, the provisioning fan-out, the revocation fan-out, the cross-system sync. None of which is the part of the role the team was hired for. The operator who reads this list and concludes the agents cannot replace headcount is correct. The agents take the parts of the work people should never have been hired for in the first place.