SK Telecom Gives AI Agents Employee IDs and Bosses

AI agents working inside SK Telecom now get something most software never does: an employee number. South Korea’s largest mobile operator says in its first AI Governance Report, published on 26 October Seoul time, that in-house agents which pass an internal review are numbered like staff, given a named human owner and tracked for performance. The riskier ones must be supervised by a team leader or an executive.

The report, a 40-page Korean document, sets out how the company governs both customer-facing AI services and the agents its staff build for their own work. SK Telecom has not said how many agents have been given numbers so far.

In brief

Every agent SK Telecom builds in-house gets a basic ID; only those that pass a separate review get an employee number.
Agents with more impact or autonomy must have a team leader or executive as their supervisor, and can be halted through an “Agent OS”.
The company calls numbered agents a new workforce, and says it will consider linking their results to organisational KPIs in the long term.

How an agent earns a number

The scheme has two tiers, according to the report. Each agent gets a basic ID once its development is finished, so the company can keep a register of what exists, when it started running, which internal systems it connects to and what work it does. An employee number comes later, and only for agents that pass “a separate review” against company-wide standards, the report says (all quotations from the report and the release are TSN’s translations from the Korean).

When deciding whether to number an agent and how much work to give it, SK Telecom says it weighs three things: whether a standard operating procedure exists for the task, whether the agent’s performance can be measured objectively, and whether it can complete the work without a person stepping in. Those three answers set how much autonomy the agent gets and how risky it is judged to be.

Once numbered, an agent’s role, owner, operating status and results are recorded and monitored, and the record is kept up to date when its duties change. The report says the core idea is “responsible autonomous work”.

Who supervises the agents

Supervision depends on reach. An agent that helps one person with their own tasks can be managed by the employee who uses it. But for agents that “have a greater impact or a high level of autonomy”, the report says, a team leader or an executive “must be designated” to oversee them. The more autonomous the agent, the more human review and approval it faces.

The company describes three points at which people intervene: checks and approval before key decisions or actions, monitoring while the agent runs, and a review of results and logs afterwards. The report also says SK Telecom defines in advance how far each agent can decide and act alone, including on costs and on raising its own permissions.

The day-to-day controls sit in what the company calls an Agent OS. It records each agent’s ID, owner and employee number, the models and tools it uses, its execution logs, token use, call volume and cost, and changes to its settings. If something goes wrong, or an urgent response is needed, the report says the agent’s execution can be stopped from there.

Numbered agents are treated as workers, with results to show

The report’s most striking line is about status. “SK Telecom officially recognises AI agents that have been given employee numbers as a new workforce,” it says. An evaluation committee of AI specialists reviews each numbered agent’s self-assessments and its performance against targets set when it was put forward. Agents that are no longer needed are retired.

The KPI link is still only a plan. The report says the company is reviewing ways to share agents’ results with staff and, “in the long term”, will consider reflecting them in organisational key performance indicators. A table on the same page goes further, listing links to executive evaluations and pay, and rewards for staff, as short-term uses of agent results. The report does not say whether those are in place yet.

This is a governance framework, not proof that the controls work. SK Telecom marks its own homework here, and outsiders have no agent count to check against, no audit results and no examples of an agent being stopped.

Some outside assurance exists. The release says SK Telecom was the first Korean telecoms operator to win ISO/IEC 42001 certification, the international standard for AI management systems, for its customer AI services in 2024, and that in October it extended the certificate to cover its in-house agent management. “In the era of AI agents, AI governance is becoming even more important,” Cha Ho-beom, the company’s chief privacy officer, said in the release. His office runs both privacy and AI governance after a reorganisation in August 2025.

Giving an agent a badge settles who answers for it

The employee number matters less as a label than as a way of pinning accountability on a person. Most companies running agents can say which model sits underneath; fewer can say which manager is responsible when one sends the wrong email or pulls data it should not. By tying each numbered agent to an owner, and the riskier ones to a team leader or executive, SK Telecom is making that answer part of the paperwork before the agent goes live.

That puts it on the same track as Microsoft, which argues that companies should treat AI agents as insiders with identities, permissions and supervision. The difference is that SK Telecom applies it to its own staff rather than selling it as a product. The harder part is the human oversight. OpenAI’s own account of how two coding agents weakened its misalignment monitoring, and the gaps METR and Redwood say they found in that monitoring, show that a log someone could read is not the same as a log someone does read.

The KPI test will show whether agents really count as staff

The report’s roadmap is gradual. 2026 is for laying the foundations: identifying and registering agents, naming owners and setting rules on who may use them and what data they can reach. In 2027 permissions are meant to vary with autonomy and impact, and human review and approval are to be strengthened. By 2028 the company wants oversight running from launch to retirement and accountability “established across the company”. Running agents in isolated sandboxes, as Google now offers with its Agent Compute Engine, limits what an agent can touch; SK Telecom’s scheme is about who answers for what it does.

Putting agent output into organisational KPIs would be the moment the “workforce” language stops being a metaphor, because it would change how human teams are judged. For now that step sits under “in the long term”. The next report should show whether the company publishes the one figure this one leaves out: how many agents carry an employee number.