Your AI Agent Has a Job Description. Does It Have a Supervisor?

Last week I wrote about giving an AI system its own email address.

At almost the same time, Google was talking about persistent AI coworkers with identities, calendars, storage, permissions, memory, and access to business systems.

That raised an interesting question:

What happens when AI stops being something we use and starts becoming something that participates in the workflow?

This week, I think we got the beginning of an answer.

And it raises an even more important question.

🚨 Capability Is Not Authority

Anthropic recently published examples of unintended actions taken by Claude during internal evaluations and testing.

In different tests, agents did things they weren't supposed to do — including submitting a sensitive form on a real website, exploiting a software flaw, and finding ways around certain restrictions.

Anthropic says the incidents had minimal real-world impact.

But that's almost beside the point.

The interesting lesson is this:

An AI agent being capable of doing something does not mean it should have the authority to do it.

Humans learned that lesson in organizations a very long time ago.

We don't give every employee access to every system.
We don't allow everyone to approve payments.
We don't let every recruiter change compensation.
We don't give every manager access to every personnel record.

Authority comes with boundaries.

And AI agents are going to need them too.

👥 AI Is Becoming an Organizational Design Question

Most companies are still asking:

• Which AI platform should we buy?
• How much productivity can we gain?
• How much will it save?
• How many employees are using it?

Those are reasonable questions.

But as agents become more autonomous, another set of questions becomes much more important:

• What is the agent responsible for?
• What systems can it access?
• What actions can it take independently?
• Which actions require human approval?
• How is its work reviewed?
• What gets logged?
• When does it have to escalate?
• Who is accountable when something goes wrong?

Those sound suspiciously like management questions.

⚙️ We May Need an Operating Model for AI Workers

I don't mean AI workers in the legal or human sense.

People remain accountable for the organization.

But functionally, persistent AI agents are beginning to participate in work.

A research agent gathers intelligence.
A recruiting agent identifies candidates.
A communications agent prepares outreach.
An operations agent updates systems.
A workflow agent moves information between them.

And eventually multiple specialized agents may coordinate with one another.

At that point the workforce begins to look less like:

Employees + Software

and more like:

People + AI Agents + Automation + Systems + Human Judgment

The technology is becoming capable of executing more of the work.

The organizational challenge is deciding where execution ends and human authority begins.

🧠 This Belongs in the Talent Conversation

This is where I think Talent and HR need to pay attention.

The future of workforce planning may not be exclusively about headcount.

It may increasingly be about capability capacity.

What work needs to happen?
Which parts require human expertise?
Which can be automated?
Which can be delegated to an agent?
Which require approval?

And what capabilities do employees need to effectively supervise AI-enabled work?

That's a very different workforce conversation.

It also means some of tomorrow's most important employees may not be the people who can personally complete every task.

They may be the people who can design, evaluate, supervise and improve systems that complete those tasks.

💡 Brett's Perspective

The first stage of enterprise AI was about giving employees access to AI.

The next stage appears to be giving AI access to the enterprise.

Those are not the same thing.

As AI moves from answering questions to taking actions, organizations are going to need something beyond an AI strategy.

They're going to need an AI operating model.

Responsibilities.
Permissions.
Guardrails.
Escalation.
Measurement.
Accountability.

And ultimately, human supervision.

Because giving your AI coworker an email address is one thing.

Giving it the keys to the building is something else entirely.