How to decommission an AI agent

It happens to Bournes, Bonds, and bots alike: Eventually, some agents need to retire.

Professional services firm KPMG’s Q2 2026 numbers reveal that, among more than 200 senior business leaders at US companies with annual revenue of $1 billion or more, AI agents have a 53% adoption rate. That’s a solid rate for deployment—but what happens if a company wants to un-deploy one, two, or even all of them?

There are many reasons to cut an AI agent from the tech stack: New and better ones might come along, for example, or maybe that latest research bot just isn’t providing the sales team with the right answers. But doing so without careful consideration could lead to problems related to security, service outages, and angry users.

We spoke with agentic pros about how to give an AI agent its burn notice.

Same old, same old. A typical IT pro is no stranger to decommissioning software. In the course of their career, a company’s resident techie has likely had to disable everything from an unused email server to an old fax machine.

Shutting down an agent is fundamentally another change-management exercise, according to Guy Bourgault, head of agentic services at tech-services provider Concentrix, and requires familiar steps: preset rollback plans, audit trails, and dependency monitoring.

Agents, however, also introduce unique-to-agent considerations:

• Permissions (who has access)
• System access (what systems the agent can interact with)
• Tokens (the bits of input and output text that can lead to high costs)

“Those represent latent vulnerabilities,” Bourgault said.

Step 1: Audit. Before decommissioning, Bourgault recommends an audit of the agent’s last 60 days of activity “to really understand if some of the assumptions that you’ve made about decommissioning are backed up by the day-to-day performance and data that’s available.” That assessment should include conversational flow, use cases, knowledge bases, how agents are activated, and interactions with additional agents—and if all of those activities match expectations.

“It’s not a ‘pull the plug and see what happens,’” Bourgault said.

Bourgault frequently works with clients on mapping how AI fits into a business process; he sees that dependency graph also helping an IT pro understand who’s using the agent and what’s impacted if an agent disappears. Today’s asset-tracking ITSM tools, which can spot configuration items in configuration management databases (CMDBs), can be repurposed to discover an agent’s connections and impacted systems, he added.

Bourgault also recommends that agent decommissioners consider the following:

• Identify stakeholders and who’s impacted—and anticipate customer impacts and increased workloads as teams troubleshoot and work through the new workflow
• Plan the replacement process.
• Deactivate but don’t fully remove the agent. Archive prompts, knowledge base articles, context and definitions, guardrails and intents, and be ready to roll the agent back quickly if needed.

Keys and tokens, too. Agents are given access to services to do their jobs. Proper decommissioning means disabling the credentials to those third-party platforms so they can’t be exploited by threat actors.

An agent may use an API key to access an external party’s data source. Shanti Greene, head of data science and AI innovation for enterprise AI solutions company AnswerRocket, recommends one unique key per agent; that way, if a runaway API key is running on a loop and triggering charges, you know which agent is causing the problem.

When an agent needs to be retired, Greene advises IT pros to go to the provider’s platform and revoke (or disable) a given key. He has used an open-source tool called Infisical as a kind of access broker that keeps agents from possessing credentials at all; the manager grants agents access on demand without making the credentials directly visible to those agents.

And there are tokens, too—meaning credentials, not the AI processing units. An agent might have access-granting OAuth tokens or credentials, depending on the systems it touches, which is no different than a human user’s typical access. For example, an agent running in Entra might have permissions to interact with Google Cloud. A configuration setting in Google’s platform enables trust with an Entra token; you want to make sure you dispose of that, according to Will Pocknell, director of cybersecurity at solutions integrator Insight, to avoid inadvertent privilege escalation. (There has been extensive reporting on how hackers steal OAuth tokens to attempt access to platforms like Salesforce, Vercel, and LastPass.)

While it’s important for IT pros to log traces of an agent to avoid potential disruptions, it can add to the cleanup if an agent needs to be disabled, especially if the team must eliminate virtually every trace of the agent’s workflow.

“If an agent is offboarded, you are cleaning up all the artifacts associated with that agent, almost to the point where it almost as if the agent never existed,” Pocknell said.