Your AI Agent Hasn’t Earned That Promotion

A great demo gets your attention. Authority takes more.

“Why don’t we just let it handle this?”

It comes up in the operating review, right after an impressive demo.

An AI agent has just produced a sharp account summary. It pulled together emails, CRM notes and support tickets, noticed that a customer’s security review has stalled, and drafted a sensible follow-up.

My answer starts with another question: handle what, exactly? Send an internal reminder? Update the forecast? Tell the customer that engineering will ship the feature by month-end?

Those are very different jobs. A good summary tells you almost nothing about whether the agent should own any of them.

Before an agent gets more authority, I want two things: evidence that it can handle the work, and explicit permission from the person accountable for the consequences.

The demo is the interview

We already know how to do this with people. An analyst who writes a strong investment memo doesn’t get to commit the company’s money. A capable engineer can diagnose an outage without unrestricted access to production. We promote people on their record, into a defined scope, and someone signs off on the decision.

With AI, the demo often does the job of the whole promotion case. One polished walkthrough on clean data, and the conversation jumps to switching it on.

Evidence is only half the test.

Engineering hasn’t agreed to that

Take that stalled security review. I’d happily let an agent gather the evidence and draft a status update, as long as it shows its sources. Sending the internal follow-up could come next.

Committing to the customer is a different kind of decision. The agent may understand the situation perfectly. It still doesn’t own the engineering roadmap, and neither does the account team that deployed it.

Good reasoning doesn’t give anyone, human or software, the right to commit resources that belong to someone else. That authority comes from whoever owns the resource, however well the agent performs. I suspect that’s where the expensive mistakes in agent rollouts will come from.

Fifty approvals before lunch

The opposite reflex is just as tempting: put a human in front of every action. It sounds responsible. It also dodges the question of whether that person has the time and context to make a real decision.

If an executive has to reopen six systems to check a recommendation, the approve button hasn’t saved anyone time. If fifty requests land before lunch, the reviewer’s available attention becomes a constraint you need to measure.

Anthropic’s usage data from earlier this year is useful here.[1] Among Claude Code users, full auto-approval rose from about 20% of sessions for newer users to over 40% by 750 sessions. Experienced users also interrupted the agent more: in around 9% of turns, against 5% for newer users. It’s observational data from one product, mostly used for software engineering.

Anthropic interprets this as active monitoring. That’s plausible. It still leaves the question an executive cares about: did people catch the mistakes that mattered?

I’d design the review around that question. Give the reviewer the proposed action, the evidence behind it, the uncertainty that matters and the consequences of going ahead, plus a way to stop the work at any point.

The approval click is the last inch of oversight. Most of the work has to happen before it.

Count finished work

When a team asks for more autonomy for an agent, I’d ask for operating results next to the demo. How much work reached an acceptable finish? How much checking and correction did it need? What did the agent miss? Were the right decisions escalated?

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support workers and found AI assistance raised issues resolved per hour by 15% on average, with the biggest gains among less experienced staff.[2] That’s assistance in one workflow, so it says little about autonomous agents company-wide. What I like is the unit: a resolved customer issue, with finished work behind the number.

In engineering, the equivalent is changes that get accepted and hold up after review. In revenue, it’s a blocked dependency that gets a decision before it puts a commitment at risk. More tickets, summaries and messages are activity.

An agent that saves ten minutes of drafting and creates fifteen minutes of checking has added five minutes of work. Celebrating its output won’t fix that.

Write the job description before the promotion

My practical recommendation: give every deployed agent a short written mandate, the way you’d define a role before promoting someone into it. It should answer five questions:

Then write down what would justify expanding it. Test the ordinary cases, the ambiguous ones and the ones where the right move is to stop. Count what the agent missed, too. A quiet agent isn’t necessarily a reliable one.

For the agent in that security-review example, the next promotion might be permission to send internal reminders. Before granting it, I’d want evidence that it reaches the right owner, avoids duplicate messages, recognizes when the issue is already resolved, and escalates conflicting information. Customer commitments would stay outside its mandate.

And make the authority reversible. You can revoke permission. You may not be able to undo what happened before you did. A model update, a new data source or a changed workflow can invalidate the evidence behind the last promotion, so someone has to own the call to pull it back. Agents are easier than people here: you can demote one without a difficult conversation.

Full autonomy is the wrong target

I’d resist making full autonomy the destination for every workflow. Start with the simplest setup that does the job, and add independence only where it improves the result.

Sometimes the right answer is an agent that finishes the work on its own. Sometimes it’s a system that does the preparation exceptionally well and leaves the consequential call with a person. Both count as success. The aim is to remove work while keeping the judgment the business depends on.

At your next AI review, pick one agent and ask the team to describe the boundary of its authority. Then ask what evidence would justify moving it.

If the answer is mostly about how good the demo looked, hold the promotion.

Which decision would you keep outside an agent’s authority even after it proved it could make it well, and why?

For more on AI and quantum as business decisions, subscribe to Uncharted.

Sources