What can it still reach?
A key that still works and a server that is still connected keep their access long after anyone needs it.
Holistic AI EndLayer / Zombie AI
Zombie AI.
Zombie AI.
The pilot ended. The access stayed.
Find the coding agents, MCP servers, keys, seats and cloud agents that are still installed, connected or paid for after the work moved on. See who owns each one, then keep it or turn it off through the control that owns it.
From what is still standing to a decision
01 / Still standing
Coding agentInstalled, no runs in 30 days
MCP serverDeclared, never called
API keyValid, quiet for 58 days
Paid seatNo metered use this month
03 / A row you can decide on
● Still standing
MCP server: github-legacy
- Last call
- Never recorded on this machine
- Owner
- Account disabled in Entra ID
- Holds
- A credential: the variable name, never the value
- Can do
- Remove it from this machine on its next check-in
Next decision
Keep it for 90 days with a reason, or remove it. Either way, the decision is recorded.
Illustrative recordStanding → Reviewed → Decided
What gets left behindA pilot finishes.
A pilot finishes.
Its footprint stays.
What is Zombie AI?
Zombie AI is AI that stays installed, connected, credentialed, scheduled or paid for after the useful work has faded, or after the person accountable for it has gone.
A team tests a coding agent, connects it to an internal service, issues a key and buys a few seats. The pilot ends and the project moves on. The agent, the connection, the key and the subscription stay. Nothing announces it, because nothing new arrived.
What are we still paying for?
Seats with no use this month, and agents that keep running a job nobody reads, still cost money.
Who answers for it now?
When the owner leaves, a scheduled agent can keep doing yesterday's job with nobody accountable for it.
The lifecycle gapThe work can end
The work can end
before the access does.
Approval answers whether a tool should arrive. Nobody is asked whether it should still be here. Zombie AI brings that second question into review.
Has this been approved?
The arrival question: what came in outside the organisation's governance. Shadow AI detection ↗
Should this still be here?
The lifecycle question: does what remains still have use, an owner and a reason. An approved tool can become a zombie too.
What EndLayer findsEight kinds of AI
Eight kinds of AI
left standing.
Each kind gets its own test, because a seat has a billing month, a schedule has a rhythm and a new key needs time to be used. One label for all of them would hide the difference.
Coding agents
Installed, with no recorded runs in the window.
MCP servers
Still declared, with old or no recorded calls.
Scheduled agents
Missing their rhythm, running for an owner who left, or running with nobody using them.
Devices
Gone quiet, never reporting, or tied to a disabled directory account.
Cloud agents
Agent Engine deployments and Agent 365 agents with old or no use.
Keys and connections
Valid, with no observed use after a seven-day grace period.
Paid seats
No metered use this month: a concrete question before renewal.
AI projects
No dated AI work on any of their machines inside the window.
Findings reflect the devices that are enrolled and the accounts that are connected. "Never" means no event in the evidence a source keeps, not in the resource's whole life.
Explore the 21 signals ↗How Zombie AI worksFrom scattered evidence
From scattered evidence
to one decision.
EndLayer already holds the evidence. The sweep turns it into one review list, and a person decides what happens to each row.
- 01 / Observe
Gather what is known.
Installed tools, agent runs, MCP calls and device reports from the Layer 1 Agent, with directory, cloud and usage records from connected accounts.
Surface capabilities ↗ - 02 / Identify
Apply the right window.
The sweep runs daily and on demand. Most kinds use a window you choose: 14, 30, 60 or 90 days. Seats and schedules follow their own rules.
Monitor capabilities ↗ - 03 / Decide
One list, with context.
Each row shows the resource, its last recorded activity, its owner where known, its source and the next step it supports.
Govern capabilities ↗ - 04 / Act
Turn it off, or keep it.
One action reaches the control that owns the resource, with its own permissions. Requested, pending and done are tracked as separate states.
The Layer 1 Agent ↗
A next step with a clear scope"Turn it off" means
"Turn it off" means
something specific.
Each kind of resource is switched off by the mechanism that owns it. Some decisions act straight away; others record work for a person to finish.
ResourceWhat turning it off doesWhat stays in your hands
Agents and MCP servers
Removed from that one machine on its next check-in.
The device reports back, so you see when it is gone.
Keys
An EndLayer key is paused or revoked. An authorised owner can deactivate a supported Anthropic key.
Each action keeps its own permission, and is recorded.
Devices
The device is revoked in EndLayer, or the agent is uninstalled through Intune.
Intune and the directory stay the systems of record.
Cloud agents
A permitted Google Cloud connection deletes the Agent Engine deployment.
Only where the connection allows it.
Seats and projects
A seat is marked to reclaim; a project is marked retired.
The vendor subscription and the repository are untouched until you make the change.
Nothing is removed on its own.
The sweep finds and explains. An authorised person chooses the action and gives a reason, and both are recorded.
Keeping it is a decision too.
Keep an item out of the sweep for 90 days with a reason. Retaining something belongs in the record as clearly as retiring it.
Revenue-linked work is protected.
Rows marked revenue-linked stay out of bulk actions, and their name must be typed before anything is switched off.
Savings are counted when they are real.
A seat's monthly price shows what could be recovered. It becomes a saving only once the subscription is actually changed.
Start with one pilot
What did your last AI pilot leave behind?
Pick one team, one kind of resource or one finished pilot, and give everything still standing a next decision.
Five practical answers
Before you start.
What it is, how it differs from Shadow AI, what it will and will not do, and how to turn it on.
What is Zombie AI?
AI that stays installed, connected, credentialed, scheduled or paid for after its useful activity has faded, or after the person accountable for it has gone: a coding agent nobody runs, an MCP server nobody calls, a key nobody uses, a seat with no use this month.
How is Zombie AI different from Shadow AI?
Shadow AI asks whether something was approved. Zombie AI asks whether it should still be here. An approved tool can become a zombie, and an unapproved tool can be in heavy use, so the two need different reviews.
Does EndLayer delete anything automatically?
No. The sweep finds resources and explains why each one is listed. An authorised person decides what happens and gives a reason, and the action then runs through the control that owns the resource.
How does EndLayer decide something is unused?
Each kind has its own test. Most use a window you set to 14, 30, 60 or 90 days; keys get a seven-day grace period; seats are judged on the current billing month; scheduled agents on the rhythm they usually keep. "Never" means no event in the evidence the source keeps.
How do I turn Zombie AI on?
Zombie AI is off by default, and an administrator turns it on in the console. What it finds depends on the devices that are enrolled and the accounts that are connected. Book a demo to see it on a fleet like yours.
EndLayer
Give AI room to work.And clear boundaries.
See the Guardian Agent Platform on your own fleet. Find every AI tool, understand what it reaches and costs, and apply the controls your team needs.
One connected layer of control
Devices · Cloud · ApplicationsSurface / Monitor / Govern