SK Telecom gave its AI agents employee numbers. Keep the approver off the scoreboard.

If you want a template for running AI agents inside a large company, SK Telecom just published one. Copy the dull parts: every agent gets an ID, the risky ones get a named boss, and somebody can press stop. Be careful with the part that made the headlines: agent results feeding KPIs and, per the report's own table, executive pay.

The Korean carrier published the report on Monday 26 October¹. It is the company's first, and most of its chapter on internal AI is about agents². The report is in Korean, so the wording below is our translation.

/01What SK Telecom actually built

Two tiers of identity. Every agent that finishes development gets a basic ID, so the company knows what exists. Agents that pass a separate review against company standards get an employee number. From then on, the agent's role, its owner, its operating status and its performance are monitored².

An owner sized to the risk. An agent that helps one person can be managed by that person. An agent with bigger impact or higher autonomy must have a team leader or executive designated to supervise it². For important decisions and for cost or permission escalation, the report says human involvement is designed to fit the work. It defines how far the agent may decide and act alone, and when a person steps in².

Three tests before more scope. Before an agent gets an employee number and a scope of work, SK Telecom checks three things. Is there an established standard operating procedure? Can performance be measured objectively? Can the agent complete the work without a human stepping in²? The more autonomous the agent, the stronger the human review and approval steps².

One record and a stop button. A central Agent OS shows which model and tools each agent uses, including MCP and API tools. It keeps step-by-step execution records, token usage and call volume, and a history of every settings or function change. When something looks wrong, an agent's execution can be halted, and that action lands in the history too².

An end date. Agents with employee numbers are reviewed regularly against their cost and results. Each is kept as is, rescoped, or retired once it is no longer needed².

SK Telecom also says its internal agent management system won ISO/IEC 42001 certification this month, on top of the 2024 certification for its customer-facing AI services¹.

/02The part to argue with

The report recognises numbered agents as a new workforce and evaluates them through a committee of AI experts². Its own table lists the short-term use of those evaluations: links to executive evaluation and compensation, plus rewards for staff. The longer-term plan is agent results inside organisational KPIs². The Seoul Economic Daily led with the workforce and KPI angle³.

Now put two rules side by side. High-impact agents must be supervised by a team leader or executive. Executive pay may be linked to how those agents perform.

That can leave one person both approving the agent's actions and getting credit for its output. An executive who scores better when the agent clears more tickets has a reason to widen its scope and wave through its holds. Once agent results sit in team KPIs, the same goes for every team leader. It is the core rule of segregation of duties: the person paid on a number should not be the only one who can move it. Putting an agent in the middle does not change that.

Reward the outcome if you like. But the owner should not be the only one who can loosen the agent's limits.

/03What to copy this week

1. /01

   Give every agent an ID on day one. Widen its access only after a review. Start with a list. If you cannot count your agents, you cannot govern them.

2. /02

   Assign the owner by blast radius, not by who built it. A personal helper belongs to its user. Anything that touches customers, money or permissions gets a named team lead.

3. /03

   Run SK Telecom's three tests before widening scope. No written procedure, no objective measure, or no clean finish without a person? Then the agent proposes and a person approves, every time.

4. /04

   Turn the two triggers into rules the agent cannot edit. SK Telecom names important decisions and cost or permission escalation. In your stack that reads like "any refund over $500, or any request for a new permission, waits for the owner."

5. /05

   Make loosening a second person's call. Tightening an agent's limits can be one click. Widening them needs someone outside the owner's reporting line. Our two-person rule piece shows how.

Gatehouse is built for exactly this split. The agent works under its own identity and limits. Anything outside them waits for a named approver, and the rules sit outside the agent, so the person scored on its output cannot widen them off the record. Every action and every rule change lands on that record. For the stop button itself, see kill switch, pause and rollback.

Pick your busiest agent. Find its owner, its limits and its last config change. If any of the three takes longer than five minutes, start there.

Want the next one? News with a take, three times a week. Follow by RSS →