05 / Governance

Zombie AI: the AI nobody is using, and why it is still a risk

AI can remain installed, connected, credentialed or paid for after its work ends. How Zombie AI connects evidence of disuse to an owner, a decision and a supported action.

An unused or unowned AI tool can remain connected to credentials.

Key takeaways

  • Unused does not mean harmless: An installation, connection, credential or paid allocation can remain after the work that justified it ends.
  • Approval and continuing need are different questions: An approved tool can become Zombie AI, while an unapproved tool can be heavily used.
  • Inactivity is a reason to review: Different resources need different observation windows, and missing activity does not prove that something has never been used.
  • A person makes the decision: Zombie AI is off by default. An administrator enables it, and an authorised person chooses each action and gives a reason.

AI can outlive the work that justified it. At Holistic AI, we call this Zombie AI: AI that remains installed, provisioned, credentialed, scheduled or paid for after useful activity fades or accountable ownership breaks down.

It can start with one person trying a tool and never opening it again. It can also be a once-useful agent that falls out of use, or a scheduled workflow that outlasts its owner. A completed pilot is one example of the problem.

We built a way to find it, review it and decide what happens next in Holistic AI EndLayer, the Guardian Agent Platform. Where an administrator enables Zombie AI, a daily sweep brings supported activity and ownership records into a shared review list. Nothing is removed automatically.

FilmZombie AI in a minute: how AI gets left behind, how the sweep finds it, and why nothing is turned off without a person. Example data, not a live fleet.

How AI gets left behind#

Someone installs a coding agent to try it. They connect a service, add a key and take a paid seat. Later, they switch tools or stop opening it.

The installation stays. So can the connection, the key and the subscription. The same thing can happen when a team finishes a project, abandons a pilot or changes who owns a workflow.

A resource can also need review while it is still running. A scheduled agent may keep doing yesterday's job after its owner leaves, with nobody accountable for whether it should continue.

No new application has to appear. The organisation can simply accumulate resources whose continued existence nobody has recently considered. The question is whether each one still has useful activity, an accountable owner and a reason to remain.

Each resource can leave something behind that still matters: access, authority, operational responsibility or cost. A valid credential still grants access. An installed tool or connected server still needs an accountable owner. A paid seat still has a renewal date. The end of useful work does not retire any of those things on its own.

Zombie AI, explained

AI can outlive its purpose.

Unused or unowned AI can still hold access, keep running or keep billing.

  • Installed and forgotten. Someone tries a tool and forgets it.
  • Use fades. An agent, key or seat falls out of use.
  • Owner leaves. The owner leaves. The resource stays.
  • Work moves on. The work ends. Jobs and connections stay.

What can remain: tools and agents, keys and connections, seats and services, jobs and projects.

Access

Keys and connections can remain valid.

Ownership

Responsibility can lapse while an agent keeps running.

Cost

Seats and services can keep billing.

Evidence and an owner inform a decision. A person chooses whether to keep, reclaim or retire the resource.

Inactivity is a reason to review, not proof to remove.

Holistic AI EndLayer, the Guardian Agent Platform. Zombie AI is off by default. Nothing is removed automatically.

The pilot that never ended#

In large organisations, AI often arrives through a proof of concept. A team trials several coding agents side by side, connects a few MCP servers to test data, issues a key for each model and buys seats to compare.

The evaluation ends with a decision, and one tool goes forward. The others rarely get a decommissioning step of their own. The agents that were not chosen stay installed on the testers' machines, the test connections stay declared, the keys stay valid and the seats renew.

Every team that runs an evaluation can leave a set like this behind. None of it is unapproved, because it was approved for the pilot, so a Shadow AI review has no reason to flag it. What changed is that the work ended.

Figure 1One proof of concept, six months on. Four coding agents are evaluated, one is chosen, and the other three keep their installation, key, connection or seat after the decision. Example only.

Approval and continued need#

A quiet agent may be needed next month. A rarely used deployment may support an important customer. A project may be deliberately paused. Evidence of disuse gives a team a reason to review the resource and establish whether it still has a purpose.

That question is different from the one raised by Shadow AI. An approved tool can become a zombie. An unapproved tool can be heavily used. Approval alone does not establish continuing need.

ReviewThe question it answers
Shadow AIHas this use been evaluated and sanctioned? What has arrived outside the organisation's intended governance?
Zombie AIShould this resource still be here? Does it still have useful activity, an accountable owner and a reason to continue?

The distinction matters across teams. Security cares about lingering access. FinOps cares about idle allocations. IT and platform owners care about resources that keep running, reporting or requiring maintenance after the work has moved on.

Discovery tells you what arrived. Lifecycle review asks what should still be here.

Eight forms of Zombie AI#

We built the EndLayer review queue around eight kinds of resource. They bring installed tools, connections and paid allocations into the same conversation, while keeping the evidence and available response specific to each kind.

  1. Coding agents: A supported agent remains installed, but recorded runs show old or no use in the available collection window.
  2. MCP servers: A declared connection remains available while a supported call ledger shows old or no recorded calls.
  3. Scheduled agents: A headless workflow misses its cadence, loses its owner or keeps running without recent interactive use in the supported evidence.
  4. Endpoints: A machine goes quiet, fails to establish reporting or remains associated with a disabled directory account.
  5. Cloud agents: Connected Agent Engine or Microsoft Agent 365 records identify a deployment that deserves review.
  6. Keys and model connections: Supported credential and inference connection records show old or no observed use after the applicable grace period.
  7. Paid seats: A paid seat has no priced, metered use recorded in the current month, giving its owner a concrete renewal question.
  8. AI projects: The latest dated AI work across a project's observed machines is older than the review window.

These resources need different tests. A scheduled job has a cadence. A seat has a billing period. A newly created credential needs time to establish use. A single inactivity label would hide those differences.

A recurring agent can also need review while it is still running. If its owner's directory account is disabled, continued execution can itself be the reason to ask who is now accountable.

From evidence to a decision#

The architecture follows the investigation a team needs to carry out.

First, EndLayer brings together evidence it already collects or receives from supported connections: installations, agent runs, MCP calls, device reports, directory ownership, cloud records and usage data. The sweep reads those records to identify resources that deserve review.

Second, rules appropriate to each resource identify candidates. The sweep runs daily where enabled, and an administrator can run it again. Most kinds use a configurable window of 14, 30, 60 or 90 days. Seats and scheduled work follow their own billing or cadence rules, and keys have a seven-day grace period.

Third, the findings become one review list. Each row joins the resource with its last recorded activity, available owner, source and supported next step. A known seat price can also show a monthly recovery opportunity.

Figure 2Where Zombie AI is enabled, a daily or on-demand sweep reads existing device and connected-account evidence. Resource-specific tests create review candidates with activity, ownership and source context. A person decides whether to keep an item or approve a supported action. The owning control applies that action, with permissions, a reason and an action record. Pending requests remain distinct from completed work and available outcome evidence.

This gives security, IT and FinOps a shared starting point. Teams can inspect the same evidence and distinguish an access decision from a subscription decision.

Coverage remains part of that conversation. Findings reflect supported records, source freshness and observation windows. "Never" means no event in the evidence the source keeps. It does not describe the resource's entire life.

A review queue with context#

Zombie AI puts never-observed and older candidates first. Teams can browse the review queue or focus on one kind: agents, servers, schedules, endpoints, cloud agents, keys, seats or projects.

Opening a row explains why it is there, where the evidence came from, who owns it where known and what has already been done. A disabled directory account is visible. Missing ownership stays visible too.

ItemLast seenNext step
github-legacy: MCP server on Mina's MacNever recordedReview removal
Inventory key: EndLayer integration key58 days agoPause key
Cursor seat: no metered use this monthNo use recordedMark to reclaim
Figure 3A simplified editorial reconstruction. These records are fictional and the labels are condensed.

The activity signal and its source explain why the item needs attention. An available owner gives the team a place to resolve the business question. Requested, pending and completed describe different stages of the work.

The sweep posture separates completed actions, pending requests and items still awaiting a decision. Sending a removal request and observing its outcome are different steps.

Teams can also keep an item out of the sweep for 90 days. A reason to retain something belongs in the lifecycle as clearly as a reason to retire it.

An action with clear scope#

"Turn off" means something specific for each resource.

For a supported installation or MCP declaration, it can mean removal from one machine on its next check-in. For an EndLayer key, it can mean pause or revoke. An authorised owner can deactivate a supported Anthropic key. A permitted Google Cloud connection can delete an Agent Engine deployment.

Other decisions record work to complete. Marking a seat to reclaim leaves its vendor subscription in place until the commercial change is made. Marking a project retired records its lifecycle decision while leaving the repository intact. Some findings support review without offering a direct action.

Zombie AI reaches the mechanism that owns the resource, with that mechanism's permissions and consequences.

Figure 4Retention and retirement are both accountable outcomes. A person reviews activity, ownership, business importance and the available response. Keeping an item can defer its next review for 90 days. A supported action proceeds through its owning control with an explicit scope and reason, then a pending or completed status and available outcome evidence. A seat decision still needs the commercial change. Revenue-linked items require individual review and stay out of bulk actions.

An authorised person chooses the action and supplies a reason. In the review interface, resources marked revenue-linked stay out of bulk selection and require their name to be typed before an individual action proceeds. Bulk operations are limited to supported actions with an appropriate scope.

The review stays deliberate. A quiet month does not trigger automatic destructive cleanup. A finding brings the decision into view; an authorised person decides what happens next.

Potential seat recovery becomes savings only after the organisation completes the relevant commercial change. A pending request remains a request until the outcome is established.

The value is a repeatable way to close the lifecycle: find the residual resource, explain the evidence, identify responsibility and follow through on a decision.

Give it a next decision#

AI adoption creates a lifecycle obligation. Organisations need to know who introduced a capability, what it can reach, whether it still serves a purpose and what happens when that purpose ends.

Zombie AI makes the last part visible. It connects evidence of inactivity or lost ownership to someone who can decide, then connects that decision to a supported action.

Start with one team or one kind of resource. Look at what remains standing, and give it an explicit next decision.

Zombie AI is off by default, and an administrator turns it on in the console. Coverage and available actions depend on the deployment, connected sources and the user's permissions. Explore Zombie AI or book a demo to review what your organisation still has installed, connected or paid for.