---
title: What happens to AI agents when their owner leaves the company?
url: https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What happens to AI agents when their owner leaves the company?

**Unless you plan for it, nothing happens: most agents keep running on the leaver's keys and permissions, with nobody accountable for them. A sound leaver process uses your identity provider's deprovisioning event to find every agent, API key, skill and MCP server the person owned, hands each one to a named owner within a fixed window, and revokes or re-issues the credentials that were tied to the person.**

Offboarding for people is mature. Offboarding for the agents those people built is not. The OWASP Non-Human Identities Top 10 lists *improper offboarding* as its first risk (NHI1:2025), and AI agents make the problem worse because they are created quickly, by many people, in many tools.

## Why do AI agents become orphaned when people leave?

Agents are usually created by an individual, not by a central team. That individual's identity ends up woven into the agent in several ways:

- **Personal API keys.** A developer creates a model provider key under their own account and puts it in the agent's config.
- **Maker connections.** Low-code agents often use the creator's connections to email, CRM or file stores.
- **Delegated tokens.** An agent acting on behalf of a user holds tokens issued to that user.
- **Service accounts.** The person created a service account and was its only owner.
- **Local assets.** Coding agents, skills and local MCP server configs live on the person's laptop and in shared repos.

When the HR record closes, the identity provider disables the person. Some of these links break. Many do not.

## What breaks and what keeps running?

| Asset tied to the leaver | Typical behaviour after the account is disabled | Risk |
|---|---|---|
| Agent using delegated on-behalf-of tokens | Stops working once tokens are revoked or expire | Silent business-process failure |
| Agent using a personal model provider key | Keeps working if the key is not bound to SSO | Spend and data access with no owner |
| Agent running as a service account the leaver created | Keeps working | No accountable owner, no reviewer |
| Low-code agent using the maker's connections | Varies by platform: check yours | Either breaks or keeps the leaver's reach |
| Skills and MCP configs copied to shared repos | Keep working for everyone who copied them | Unreviewed tools with no maintainer |
| Scheduled cloud agents | Keep running on schedule | Actions nobody is watching |

The pattern is clear. Agents that use delegated identity fail safe. Agents that use static secrets fail open.

## What should a leaver workflow for AI agents look like?

Treat it as an extension of your existing joiner-mover-leaver process, triggered by the same event.

1. **Trigger on deprovisioning.** Use the SCIM deactivate or delete event from your identity provider (SCIM is defined in RFC 7644), or the HR termination feed, as the start signal.
2. **Query the inventory.** List every asset where the leaver is owner or creator: agents and their versions, API keys, MCP servers, skills, scheduled jobs and gateway routes.
3. **Split active from dormant.** Use last-invoked or last-used dates. A 30-day activity window is a reasonable first cut.
4. **Assign an interim owner.** Default to the leaver's manager from the identity provider, so ownership is never blank.
5. **Set a reassignment deadline.** Give the interim owner a fixed window, for example 14 days, to confirm a permanent owner or retire the asset.
6. **Revoke personal credentials.** Disable the leaver's personal provider keys. Rotate any shared secret they could have read.
7. **Re-issue credentials properly.** Give the agent a workload identity or a key owned by a team, not a person.
8. **Quarantine what nobody claims.** After the deadline, block or quarantine unclaimed agents, then retire them after a second window.
9. **Keep the evidence.** Record who reassigned each asset, when, and which credentials were revoked. Auditors will ask.

Steps 6 and 7 must happen in that order for active agents, or you will break a production process. Re-issue first for anything business-critical, then revoke.

## How do you stop agents becoming orphans in the first place?

Prevention is cheaper than clean-up. Four rules cover most cases:

- **Two owners per agent.** One business owner and one technical owner, both resolved against the identity provider. One leaver then never leaves an agent ownerless.
- **No personal keys in shared agents.** Production agents use workload identity federation or team-owned credentials.
- **Delegate rather than copy.** Where an agent acts for a user, use on-behalf-of delegation so access ends with the user's session. Our [guide to agent identity and on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation) covers the patterns.
- **Periodic attestation.** Ask owners to confirm their agents each quarter. This mirrors the account review control in NIST SP 800-53 AC-2, applied to agents.

If you have not yet linked agents to people at all, start with [assigning owners to AI agents using your identity provider](/blog/assign-owners-to-ai-agents-with-your-identity-provider).

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, each with an owner. Auto-discovery finds agents, models and API keys across AWS, Azure and Google Cloud, and connectors bring in agents from platforms such as Salesforce, SAP, Databricks and HubSpot, so the leaver query covers the whole estate rather than one tool. OIDC SSO, delegated on-behalf-of identity and agent workload identity federation help you move agents off personal credentials. Posture flags risky permissions and configurations, and per-agent Block and Quarantine controls act through the connected platform when an agent goes unclaimed. Opencontroller Edge extends the same view to coding agents, skills and local MCP servers on developer devices. To see it on your estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do AI agents stop working when their creator's account is disabled?**
Only some do. Agents using delegated tokens usually stop. Agents using personal API keys, service accounts or shared configs usually keep running.

**Who should own an AI agent after the creator leaves?**
Make the leaver's manager the interim owner automatically, then require a permanent business and technical owner within a fixed window.

**Should we delete an orphaned agent straight away?**
No. Quarantine it first so you can confirm nothing depends on it, then retire it after a second review window.

## Key facts

OWASP's Non-Human Identities Top 10 lists improper offboarding as NHI1:2025. Agents that use delegated on-behalf-of tokens tend to stop when the owner is deprovisioned; agents using personal API keys or service accounts tend to keep running. A leaver workflow should trigger on the SCIM (RFC 7644) deprovisioning event, assign an interim owner, set a reassignment deadline, re-issue credentials before revoking, and quarantine unclaimed agents. Opencontroller by Lyzr provides a governed registry with owners, auto-discovery of agents, models and API keys across AWS, Azure and Google Cloud, delegated identity, workload identity federation and per-agent Block and Quarantine controls. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you assign an owner to every AI agent using your identity provider?](https://opencontroller.lyzr.com/blog/assign-owners-to-ai-agents-with-your-identity-provider.md)
- [What should you do with a shadow AI agent once you find it?](https://opencontroller.lyzr.com/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it.md)
- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
