Governing Non-Human Sprawl: AI, Bots & Access
By Dr. Elena Voss on 2026-09-08 · 1772 words · 5 min read
Why service accounts, bots, and AI agents are outpacing employee identities, and how modern identity and access management tools are adapting to govern them.
Governing the Non-Human Sprawl: How AI and Automation Are Reinventing Identity and Access Management Tools
Ask a security lead how many employees can log into the customer database, and you'll get a number fast, usually with a little confidence behind it. Ask the same person how many service accounts, API keys, bots, and AI agents have that same access, and watch what happens to their face. There's a pause. Sometimes a wince. Then, almost always, some version of "let me get back to you." That pause tells you more than any audit report would. Somewhere in the past few years, the machines quietly outnumbered the people, and almost nobody sent out a memo about it.
Nobody Was Really Watching These Accounts to Begin With
When a person joins a company, there's a whole ritual around it. Background checks, an account created with a name attached, permissions assigned by someone who's supposed to think it through, and — in a company that's actually on top of things — all of it revoked the day that person walks out for the last time. A service account gets none of that ceremony. A developer needs two systems to talk to each other on a Thursday afternoon, spins one up, gives it more access than it strictly needs because trimming it down properly would eat twenty minutes he doesn't have, and moves on. Nobody ever comes back to check. That account just keeps doing its job, quietly, for years, long after everyone who understood why it existed has left the company or simply forgotten.
Now multiply that by every integration a team has ever connected, every script someone wrote to automate a report, every AI agent now sitting inside a CRM or a support inbox or a code repo, and you end up with a population of digital identities that dwarfs the human headcount — most of it created faster than anyone can realistically track, almost none of it built with an expiration date in mind. People call this "non-human sprawl," which is a polite way of describing something that's less like sprawl and more like accumulation nobody's actually watching happen.
Why Nobody Notices Until It's Too Late
Here's the thing about a person who still has access to something they shouldn't — eventually, somebody notices. People gossip. People change jobs. Someone in a hallway conversation says "wait, why do you still have that?" A forgotten service account never gets that kind of social scrutiny. It doesn't show up on a headcount spreadsheet. It's not in anyone's one-on-one. It just sits there, quietly authenticated, running on permissions that made sense back in 2022 and haven't been looked at since, waiting for either absolutely nothing to happen, or for exactly the wrong thing to.
The Old Rulebook Was Never Built for This
Most identity and access management tools were designed around a fairly simple mental model: a human logs in, that human has a role, the role determines what they can touch, and every so often somebody reviews whether that mapping still makes sense. It's a decent model. It's held up fine for years, for the population it was actually built to handle. What it was never built for is an account with no manager to ask about it, no HR file to trigger an offboarding checklist, no calendar invite that might jog someone's memory about why it's still running. A system built entirely around how people behave runs into real trouble the second most of what it's governing stops behaving like a person at all.
And that's an uncomfortable spot for a lot of security teams to be in, because it's not that their tools are broken. They're just answering a question that's gotten a lot smaller than it used to be. Knowing who has access used to be most of the job. Now "who" increasingly means "what," and the "what" doesn't file a ticket when it changes roles, doesn't get flagged in an exit interview, and definitely doesn't raise its hand to say it's outgrown its original purpose.
Then AI Showed Up and Made It Weirder
You'd think AI would be the thing that cleans this mess up, and in some ways it genuinely is — but it's also making the underlying problem stranger before it makes it simpler. An AI agent isn't just a static service account sitting quietly doing one repetitive thing. It can request new permissions on its own, chain actions across several systems in a single task, decide what it needs to touch next based on the goal it's working toward rather than a fixed script a developer wrote once and forgot about. That's exactly what makes it genuinely useful. It's also exactly what makes it so hard to govern with tools designed for permissions that never change unless a human deliberately changes them. An account that can reason its way into needing more access tomorrow than it had this morning is a fundamentally different animal than one that's been sitting with the same fixed permissions since the day it was created.
To be fair to the technology, automation is also doing real, useful work on the other side of this — it's worth not painting the whole picture as doom. Modern identity and access management tools are getting better at noticing things a person combing through a spreadsheet never would: an account gone quiet for ninety days, a permission set that's unusually broad compared to anything similar, a new integration that inherited access nobody actually signed off on. None of it requires a person to remember to go check. The system just notices, the same way a decent smoke detector notices instead of requiring someone to periodically walk around sniffing the air.
Building Governance That Fits How This Access Actually Gets Created
The organizations handling this well haven't found some clever hack. They've just stopped treating non-human identities like a side issue and started treating them like the main event, because honestly, in a lot of environments now, they are. That starts with discovery that goes past whatever IT officially provisioned, because a real chunk of this access never went through that process at all — a developer testing something, a team plugging a new AI tool straight into their workspace without ever looping security in, each one adding quietly to a population that the official records never fully capture.
It also means giving these identities something closer to an actual lifecycle, the way a human employee has one. Something gets created for a reason, it should have that reason documented somewhere, it should have an owner who's actually on the hook for it, and there should be a point down the line where somebody checks whether it's still earning its keep. Right now, most non-human identities have none of that. They were born out of convenience, and left unattended, they tend to die out of neglect — sitting there until an audit finally catches them, or worse, until an incident does.
How Zluri Fits Into This Picture
Zluri approaches the problem by extending identity governance across the entire population of accounts an organization actually has, not just the neat, HR-tracked human ones. That means surfacing service accounts, integrations, and AI agents right alongside employee access, so a security team isn't stuck manually stitching together answers from five different systems just to answer something as basic as "what can this thing actually reach." Pair that visibility with automated lifecycle management — flagging accounts gone dormant, catching permissions that have quietly outgrown what the role should need, tying every non-human identity to a real, accountable owner — and you start closing the gap between how big this problem has actually gotten and how small most teams' current toolset still is.
What This Actually Costs When Nobody's Looking
It's worth being blunt about what's really on the line here, because "governance gap" is the kind of phrase that can sound abstract in a way that lets people tune it out. An old service account still sitting on top of a customer database is a live risk right now, whether or not anyone ever touches it. It's exactly the kind of thing that shows up months later in a breach report, described in a sentence that always reads roughly the same way — access that should've been pulled years ago, that nobody had any reason to remember, until someone else found it before you did. The account was never malicious. It was just forgotten. And forgotten access doesn't need to be exploited on purpose to become a real problem. It just needs to sit there long enough for the wrong person to find it first.
Where This Actually Goes From Here
The honest prediction for the next few years isn't that this population of non-human identities somehow shrinks back to something manageable. It won't. Every new integration, every AI agent a team brings on to move faster, adds to a headcount that was already outpacing the human one before AI adoption even really took off. The teams that come out ahead of this aren't going to be the ones who tried to slow that growth down — that's not a fight anyone wins, and trying usually just pushes the behavior further underground, the same way banning software purchases outright just pushes shadow IT into the shadows it's named after. The ones who come out ahead are going to be the ones who built visibility and lifecycle discipline in early enough that the growth never outran their ability to actually see it happening. It's a far less dramatic story than a breach headline. It's also the only version of this worth actually building toward.
Frequently Asked Questions
What exactly counts as a "non-human identity"? Service accounts, API keys, bots, integrations, and increasingly AI agents — basically any account or credential taking action inside a system that isn't a person sitting down and logging in the way an employee does.
Why do traditional identity and access management tools struggle here? Most were built around human patterns — logins, roles, org charts, an HR event that triggers offboarding. Non-human identities don't follow any of that, so the signals these tools rely on to catch risk often never fire at all.
How is governing an AI agent different from governing a regular service account? A regular service account has fixed permissions that only change when a person changes them. An AI agent can request new access and chain actions across systems on its own, based on the task at hand, which makes its footprint a lot harder to predict or contain with rules built for something static.