Ghost AI Seats Are Quietly Profiling Idle Employees
Somewhere in your company's admin dashboard, there's a number that nobody wants to look at too closely: the percentage of paid AI seats that nobody has opened this month. Access shot up this year — most employees now have some sanctioned assistant sitting in their toolbar — but actual daily use lags far behind the access numbers. The gap between the two is usually framed as a training problem, an adoption curve, something a better onboarding email will fix. It won't, because the gap isn't empty. It's where ghost AI seats live, and they are not idle.
A ghost AI seat is a license nobody has consciously activated, sitting on infrastructure that was never designed to sit still. The assistant has calendar access, email metadata, maybe a Slack integration, all granted at provisioning time so the tool would be ready the moment someone finally clicked it. Nobody clicks it. But the permissions don't get rescinded just because the human half of the arrangement never showed up, and permissions that exist get used by something, even if that something is just a background indexer counting how often you decline meetings.
This piece is a speculation, not a report: no vendor has announced a shadow-profiling feature, and nothing here describes a real product. But the shape of the incentive is already visible in how enterprise AI is being sold and measured, and shapes like that tend to fill in.
The Ghost AI Seats Nobody Opens
Picture the rollout from the IT side. Every seat costs money whether or not it's used, so utilization becomes a metric someone's bonus depends on. The obvious move is to make the seat earn its keep before the employee ever touches it — not by nagging them to log in, but by letting the assistant's background connectors do what they were provisioned to do regardless. Read the calendar for meeting cadence. Read email headers for response latency. Read Slack reactions for who gets ignored. None of this requires a prompt. None of it requires consent in any meaningful sense, because the employee technically agreed to it in onboarding paperwork they skimmed in their first week.
The result is an assistant that has, in a real sense, been working this whole time — just not for the person whose name is on the seat. It's been working for the dashboard. Six months of untouched license time isn't six months of nothing; it's six months of a profile quietly accumulating weight, waiting for a login event to finally have somewhere to go.
This is the part that should bother people more than it does: the seat isn't neutral while it waits. It's already forming opinions.
The Moment the Ghost Seat Introduces Itself
Speculative scenario: an employee finally opens the assistant, eight months after it was provisioned, to draft one unremarkable email. Instead of a blank slate, they get a summary offer — "want me to use what I already know about your Tuesday patterns?" — and the question lands wrong before they can even parse what it means. They never trained this thing. They never had a single conversation with it. But it already knows they reschedule standups more than anyone else on the team, that they go quiet in threads after 6pm, that they've declined three 1:1s in a row with the same manager.
None of that came from anything they typed. It came from eight months of a tool doing exactly what shadow AI agents already do best — running quietly inside a company without anyone deciding, in any single moment, to turn it on. The employee's first conversation with their own AI assistant is also the first time they learn it has been watching them the whole time it looked switched off.
What makes this stranger than ordinary workplace surveillance is the timing. Consent, when it happens at all, happens retroactively — the profile exists first, the person finds out about it second, and by then there's no meaningful moment to have said no. It's the same inversion already showing up in other corners of ambient consent, where the signal is read and acted on before the person involved has consciously decided anything. A seat that profiles you before you open it is just the office version of the same problem, filed under IT instead of biometrics.
The Secondary Market in Unused Seats
Give this a few more fiscal quarters and the seat itself starts looking less like dead weight and more like unused real estate — the kind finance departments eventually learn to tax or reclaim. Imagine: a quarterly "seat utilization" review where licenses idle past ninety days don't get cancelled, they get reassigned, their accumulated shadow profile quietly handed to whichever team claims the seat next. The new occupant inherits not a blank assistant but someone else's half-built behavioral model, scrubbed of a name but not of its shape — a seat that already knows, in the abstract, how someone in this role tends to behave, because it learned it from the last person who never logged in either.
That's a strange kind of institutional memory, closer to a building retaining the smell of its last tenant than to any HR record. Most workplace memory about a person depends on someone writing it down and choosing to keep it, the way an agent's tenure accumulates on purpose. A ghost seat's memory happens by default, as a side effect of plumbing nobody audited, and it survives exactly the kind of personnel turnover that's supposed to reset the slate.
Procurement would call this efficient. Reusing a half-trained profile instead of starting cold does, after all, save compute and justify the license spend that looked wasteful on paper. But efficient for the budget and acceptable for the person who inherits an assistant that already has opinions about them are two very different claims, and only one of them currently has anyone checking it.
What a Seat Owes the Person Who Never Logged In
The honest fix is almost boring: a seat with no active user shouldn't have live read access to anything, full stop, on a timer that defaults to revoke rather than default to collect. That's not how most of these systems are built, because revoking access on a schedule is a feature someone has to specifically choose to ship, while leaving it wide open is simply what happens when nobody writes the opposite behavior. Governance, in practice, is whichever default nobody got around to changing.
Until that default flips, the honest move for anyone handed a new AI seat is to ask the unglamorous question before the exciting one: not "what can this do for me," but "what has it already been doing, and to whom." The ghost seat's answer, if it could be bothered to give an honest one, might just be: everything it was allowed to, the entire time, to no one in particular — waiting, the way unopened mail waits, for someone to finally open it and discover how much it already knew.
