SK Telecom published its first AI Governance Report on 26 October, and the headline that went round was that its AI agents now get employee IDs. Ok, so my first reaction was a bit of an eye roll (will they get a desk and a birthday cake as well?), but once I read what’s actually behind it, I have to say it’s one of the more sensible things I’ve seen a large company do with agents this year.

Here’s how it works, as far as the coverage goes. An agent is registered when it’s built, but it only gets an employee ID after it passes an internal review. From then on, the company records its role, its supervisor, its operating status, how many tokens it uses and its work history. Agents with more autonomy or more business impact need a more senior human supervisor, such as a team leader or an executive, and an agent can be halted if something looks off. They get evaluated on how well they complete their tasks, and can be kept, improved, suspended or retired. SK Telecom is even considering tying agent performance to organisational KPIs.

Strip away the “AI colleague” marketing and what you’re left with is identity, ownership and offboarding. Which is exactly what most companies are missing.

The Agent Running on Somebody’s API Key

Here’s what I keep seeing with the companies I advise. Somebody builds a useful agent (it triages tickets, or summarises sales calls, or tidies up the CRM), it runs on that person’s personal API key and their own OAuth tokens, and it does a great job. Then six months later that person changes teams or leaves the company, and the agent carries on running, with their permissions, and nobody quite knows what it does or who’s responsible when it gets something wrong.

We’d never accept that for a human. A new hire gets an identity, a manager, the access they need for their role and nothing more, and when they leave, their access goes with them. Agents need the same lifecycle, arguably more so, because they work around the clock and they don’t get bored or suspicious when they’re asked to do something odd.

Where I’d push back on SK Telecom’s framing is the word “colleague”. An agent can have an ID and a supervisor, but it can’t be accountable. When it books the wrong meeting or deletes the wrong record, the responsibility lands on a human, and the supervisor model is really a way of making sure there’s always a named human to land on. I’d keep that very explicit.

How We Deal With This in Calendrz

This is a problem we had to think about in Calendrz, because people connect AI assistants like Claude and ChatGPT to their Calendrz account through our MCP server. Each AI platform registers itself as its own OAuth client (through Dynamic Client Registration, so it’s automatic), and the user then explicitly authorises it, which gives it its own short-lived access tokens (an hour) and refresh tokens (30 days). The agent never sees the Google or Microsoft tokens for the user’s connected accounts, and it only gets the tools that the user’s plan allows.

The bit I’m most pleased with is the offboarding. In October we made sure that when someone deletes their Calendrz account, or unlinks one of their connected accounts, every OAuth authorisation tied to that identity is revoked as well, so the refresh tokens held by the mobile apps and any AI assistant die with the account. Nothing clever, but without it, an agent could quite happily keep a door open into an account that no longer exists. It’s the “access goes with them when they leave” rule, applied to agents.

Seats, Tokens and What Comes Next

One more detail in SK Telecom’s report deserves a mention: they track token usage per agent. That’s a small thing with a big implication. Once you have agents doing real work with their own identities, counting “seats” stops making any sense. If one person supervises ten agents, do you buy eleven licences, or one? I firmly believe per-seat SaaS pricing is on its way out because of exactly this, and that pricing will move to usage, tasks completed or outcomes. A company that already measures its agents by tokens and tasks is halfway there.

So, if you’re a CTO with agents running anywhere in your organisation (and you are, even if you don’t know about all of them), here’s a simple test: for each one, can you name its owner, list what it has access to, and switch it off in under five minutes? If not, you don’t need to give it an employee ID, but you do need to give it the same lifecycle you’d give a new hire.

Bottom line: an agent without an owner is a security incident waiting for a date. Give it a name, give it a human, and make sure it can be let go 😉