October 28, 2026

Your AI Agents Should Not Borrow an Employee’s Login

A lot of business automation runs on someone’s personal account. A script uses a developer’s token. An integration was set up with a manager’s password. It works, until that person changes roles, leaves the company, or nobody can tell which actions were theirs and which were the automation’s.

AI agents make this question more urgent, because they do more and they cost money every time they run. This week, two large software companies shipped features that point to the same answer: an agent should have its own identity, its own limits and its own off switch.

What changed this week

On October 27, GitHub added server-to-server authentication for Copilot, in public preview. Services and agents built with the Copilot SDK can now make requests on behalf of an organization through a GitHub App, without a signed-in user’s credentials.

The details are practical. The app gets a read-only Copilot Requests permission and is installed on the organization. It uses a short-lived token that expires after one hour. Usage is billed to the organization, and the organization’s model policies, budgets and spending controls apply. Owners can see which apps have access, find app usage in their AI usage reports, and cut off an app by suspending or uninstalling it.

The same day, at SAP TechEd, SAP announced an Enterprise AI Gateway to manage AI traffic in one place, covering authentication, policy enforcement, routing, monitoring, usage tracking and auditing. SAP also described a governed agent gateway that lets third-party agents call its Joule agents with enforced access control, plus new controls in Joule Studio to see agent behavior, model usage and spend.

Why borrowed logins cause problems

When an agent runs on a person’s account, three things get blurry.

  • If that person leaves, the automation can stop working, or keep working on access nobody remembers granting.
  • The usage looks like that person’s usage, so it is hard to see what the agent itself is spending.
  • In the logs, the agent’s actions and the person’s actions look the same.

None of this is new. Service accounts have existed for years. What is new is that agents act more often, across more systems, and with a running bill attached.

Treat an agent like a new hire with a narrow job

A useful way to think about it: give an agent what you would give a new employee with one specific task. Its own account, only the access that task needs, a budget, and a manager who can see what it did.

GitHub’s setup follows that pattern. The permission is narrow, the token is short-lived, the bill goes to the organization, and the access can be removed in one step. SAP’s gateway does the same at a larger scale, by putting authentication, policies and usage tracking in one place instead of scattering them across tools.

You do not need either product to apply the idea. It works for any agent that touches your code, your CRM or your ERP.

Practical steps

  • List every automation and AI agent that touches a business system, and note whose account each one uses.
  • Move the important ones to their own app or service identity, with only the permissions they need.
  • Prefer short-lived tokens over long-lived keys stored in a config file.
  • Set a spending limit per agent, and check that you can see its usage separately.
  • Make sure someone owns each agent and knows how to switch it off.

Agents are starting to do real work inside business systems. Giving them their own identity is a small step that makes them easier to trust, cheaper to track and safer to remove.

How many of your automations would stop working if one specific person’s account were disabled tomorrow?

← Back to field notes