Most organisations already believe they govern identity. They have a directory and run access reviews. They use role-based controls and have an approval process for new starters. For a workforce made up mostly of people, that approach has worked reasonably well for years. The problem is that the workforce is no longer only human. And governance designed for people starts to creak when the things being governed can act on their own.
The shift that changes the picture
Identity used to mean people. An organisation’s identity estate was, broadly, employees and contractors, with a quieter group of service accounts in the background. Most teams didn’t think about those too often, which was understandable, if not ideal.
But that picture has changed, organisations now manage two kinds of actors:
- Human users.
- Non-human identities, such as service accounts, workloads, API keys and AI agents.
The non-human side has been expanding for years through service accounts, automation, integrations and APIs. AI has accelerated that growth. In many enterprises, non-human identities already outnumber human users by a wide margin, with commonly cited ratios from 40:1 to more than 100:1. One survey found the ratio to be as high as 144:1.
This doesn’t just add volume. It adds complexity.
More identities means more owners, or no owners. More access paths. More permissions. More places where something can go wrong without anyone spotting it quickly.
Microsoft’s identity engineering leadership has described the direction clearly: by 2026, enterprises may have more autonomous agents than human users. The question is whether organisations are ready to secure and govern them.
Identity is no longer just a user problem, it’s a control problem across the whole environment.
AI agents change the risk
It would be convenient if AI agents were simply another type of account to add to the directory. But they’re not. A human user or traditional application mostly accesses systems but an AI agent acts inside them.
To do useful work, agents often need access across multiple systems. They may need permissions that last beyond a single session. They may also complete a task end-to-end without stopping for a person to check each step.
In practice, an agent can behave like a highly privileged service account, with one important difference: it can reason, make decisions and call other agents or services to complete multi-step work.
Microsoft’s security team makes the same point. Autonomous agents are not a small extension of existing identity or application governance, they’re a new kind of workload and should be treated as one. They’re also easy to create, sometimes by non-technical staff. That can lead to sprawl, shadow agents and a set of dependencies that are hard to see and harder to monitor.
So, the question is no longer just who has access? But what can act with that access?
Where governance starts to break
When a governance model built for people is applied to identities that act independently, the weak points are fairly predictable.
No clear ownership
Non-human identities are often created to get something working. When the project ends, the identity may live on, still active, still credentialed and owned by nobody in particular.
Limited visibility
Many organisations cannot produce a reliable inventory of which agents and service accounts exist, what they can reach, or what they are doing. Microsoft has described agents as scaling faster than some companies can see them. That visibility gap is not just an IT problem, it’s a business risk.
Weak lifecycle management
Human users usually have a joiner, mover and leaver process. Most non-human identities do not. There may be no review when their purpose changes and no guaranteed decommissioning when that purpose ends.
Too much static access
Role-based and standing permissions assume a stable identity doing predictable work. Agents are more dynamic, their tasks and risk profiles can change quickly, while their permissions often stay the same.
The design flaw underneath all of this is simple enough: most governance models were built for people who log in, complete defined tasks and can be reviewed at a human pace. AI agents do not work that way.
What happens when governance does not keep up
The Salesloft breach shows what this can look like in practice.Attackers exploited OAuth tokens linked to a third-party AI chat agent and accessed data across more than 700 Salesforce environments. The identity existed and the access was real. What was missing was the governance around it: clear ownership, tight scope and visibility of what that identity could reach.
That is the pattern organisations need to pay attention to. AI doesn’t just add more identities, it increases the number of identities, the speed at which they act and the impact when one is misused or makes a mistake. All at the same time.
In a human-paced environment, weak controls may be spotted before too much damage is done. With autonomous agents operating in seconds across several systems, the same gap can be used at scale before anyone reviews a log.
That’s why traditional periodic controls, such as quarterly access reviews or annual recertification, are no longer enough on their own. They still have value, they’re just not fast enough to govern actors that do not work on a human clock.
Why identity governance becomes critical
AI does not necessarily create brand new problems, it speeds up the problems already there. Every over-permissioned account becomes more exposed, every orphaned credential becomes more risky and every gap in visibility becomes harder to tolerate. The moment an autonomous agent can use that access, the stakes change.
This is why identity governance becomes central to AI security, not as a compliance exercise or a tidy-up project but as the control layer that determines who, or what, can act, under what conditions, and with what limits. Good governance gives organisations the confidence to use AI without letting access sprawl quietly in the background.
What needs to change
The fix is not exotic, it’s the disciplined extension of the governance principles already applied to people, expanded to cover every identity and made continuous rather than occasional.
Every identity needs an owner, a purpose and a lifecycle. Human or non-human, each identity should be created through a defined process. It should be clear who owns it, what it is for, what access it needs and when that access should end.
But you can’t govern identities you can’t see. A reliable inventory of agents, service accounts and integrations, including shadow AI, is the starting point for everything else. Governance needs to move from static to continuous. Access should be assessed in context and at the point of action. The aim is to prevent a policy breach as it happens, rather than find it later and hope the damage is modest.
Steven Parker, practice lead for secure identity at CWSI, offers a useful test: hold agents to the same lifecycle standard you apply to staff.
When a new employee joins, HR approves the role and IT provisions access. When they move role, access is reviewed. When they leave, that access is removed.
Now ask the same questions of an AI agent.
- Who approves its scope and permissions when it is created?
- Who reviews access when its task changes?
- Who switches it off when it is no longer used?
In many organisations, those questions do not yet have clear owners. That’s the governance gap.
Strengthen identity governance before AI scales
AI adoption moves faster when identity governance is clear. Through our AI readiness assessment, CWSI helps organisations understand whether their identity foundations are ready for human users and non-human identities.
That means looking at what exists, what it can access, who owns it, and where governance may need tightening before AI scales further.
