A Major Telecom Just Gave Its AI Agents Employee IDs, Human Supervisors, and a Retirement Policy

SK Telecom published its first AI Governance Report today, and buried in it is the most concrete answer anyone has given to a question most companies are still avoiding: what is an AI agent, organisationally?

Their answer is that it is staff. Not a tool, not a licence, not a line item under software. Staff, with everything that word drags along behind it.

Four things that make it an HR system

The report is not a statement of principles. It describes machinery, and four pieces of it matter.

Every agent gets an employee ID number. The same kind humans get. This started with their AX Innovation 2.0 announcement in June, where agents were given employee IDs, departmental assignments, job functions and access permissions, managed from onboarding through to offboarding.

Every agent has a named human supervisor. And here is the sharpest line in the whole document: for agents with greater impact or autonomy, that overseer should be a team leader or an executive, not a rank-and-file employee. Somebody senior owns the agent, by name.

Every agent is evaluated. The company tracks each agent's role, supervisor, operating status and performance, and examines token usage and work records. They are reviewing whether agent performance should count toward organisational KPIs, though that part is still under review rather than settled policy.

Every agent can be retired. Agents judged less necessary based on evaluation results get suspended or retired. There is an exit.

Permissions follow the job description, so a marketing agent does not reach network operations data, and a Zero Trust architecture authenticates both the AI and human users. A Korean report today suggests the framework extends to operation, evaluation and disposal of agents company-wide by 2028, though I could only see that article's headline, so treat the timeline as unconfirmed.

The supervisor rule is the interesting part

Most agent governance writing to date has been about permissions. What can it touch, what can it not touch, where does the audit log go. Important, and also the easy half.

SK Telecom's harder move is insisting that a senior human's name sits next to each consequential agent. That is not a technical control. It is an accountability control, and it solves a failure mode that permissions cannot touch.

Here is that failure mode. An agent does something expensive or embarrassing. The engineer who built it says it was doing what the business asked. The business owner says they did not know it could do that. The security team says the permissions were approved. Everybody is telling the truth, and nobody is accountable, because accountability was never assigned to a person in the first place. Requiring a team leader or executive by name closes that gap before it opens.

It also quietly fixes the enthusiasm problem. When owning an agent means being answerable for it, people stop spinning them up for fun.

Retirement is the part nobody plans

The piece of this that most companies will skip is the last one. Agents accumulate. Somebody builds one for a quarterly process, the process changes, the agent keeps running, and two years later it is still firing on a schedule nobody remembers approving, holding credentials nobody has reviewed.

Humans leave and there is a process. Software gets decommissioned and there is a ticket. Agents just sit there, because nothing in a normal company's machinery is designed to notice a quiet, cheap thing that still works.

Building the exit at the same time as the entrance is the unglamorous decision in this report, and probably the one that ages best.

What this looks like without a telecom budget

You do not need an agent operating system or a 2028 roadmap to take the useful part. The four questions scale all the way down:

Can you list them? Write down every agent running in your business, including the ones in someone's personal account. If you cannot produce that list in five minutes, you do not have a governance problem yet, you have an inventory problem, and it comes first.

Does each one have a human name attached? Not a team, a person. The one who answers for it when it goes wrong.

What is it allowed to reach? Scope it to the job, not to whatever the integration happened to offer.

When do you next look at it? Put an actual date in the calendar. An agent with no review date is a permanent hire you never interviewed.

That is four lines in a spreadsheet, and it is most of what SK Telecom just formalised across a telecom.

The org chart question, answered by someone who runs a network

We have argued for a while that AI employees end up on the org chart, and the honest objection has always been that this is vendor framing. A company selling AI employees would say that.

This is different. SK Telecom is not selling the concept. They run a national network, they have a workforce, and they looked at what they had deployed and concluded that the existing categories did not fit. Software does not have a supervisor. A tool does not have a performance review. A licence does not get retired for being unnecessary. They reached for HR vocabulary because the IT vocabulary genuinely ran out.

That is the tell. When a serious operator independently arrives at employee IDs, named supervisors, evaluations and an exit process, the question has stopped being whether agents belong on the org chart and become what you write in the box.

Our own bet has always been that if a thing is going to be managed like a colleague, it should arrive like one: with its own identity, its own email address and phone number, memory that persists, and an owner who knows what it is for. SK Telecom just wrote the management layer for exactly that, at scale, and published it.

The companies that will struggle are not the ones without a governance framework. They are the ones who cannot yet answer the first question: how many are running right now, and who owns each one?

Want to test the most advanced AI employees? Try it here: https://Geta.Team