For most of security’s history, an “identity” meant a person with a username. Today the people are outnumbered. Service accounts, API keys, app connections and now AI agents all hold keys to company systems, and most of them are watched far less closely than the humans they work alongside.
What is a non-human identity?
A non-human identity (NHI) is any account or credential that lets software, not a person, access company data. That includes:
- Service accounts that run scripts and integrations.
- API keys and tokens stored in apps and automation tools.
- App connections (OAuth grants), such as an AI note-taker someone allowed to read their calendar and mail.
- AI agents that sign in, read data and take actions on their own.
By one widely cited count, machine identities now outnumber people by 82 to 1 in the average organization. (CyberArk 2025 Identity Security Landscape)
Why AI agents are the riskiest kind
Older non-human identities did one narrow job. AI agents are different:
- They act, not just read. An agent can send email, change files, raise tickets and call other apps.
- They get broad access by default. It’s often easier to give an agent an admin account than to work out the narrow access it needs.
- They’re often nobody’s job. An agent built by one team keeps running after its creator moves on.
- They chain together. One agent calls another app, which calls another, so access spreads beyond what anyone approved.
The cost of getting this wrong is already measurable: 97% of organizations that had an AI-related breach lacked proper AI access controls. (IBM Cost of a Data Breach Report 2025)
The five questions to answer for every agent
AI agent governance comes down to being able to answer these, for every agent, at every client:
- Who is it? Does the agent have its own identity, or is it borrowing a person’s or a shared admin account?
- Who owns it? Is there a named person responsible for it, and do they know?
- What can it reach? Which mailboxes, files, apps and data can it open, and is that more than its job needs?
- Is it allowed? Was it approved, and does it fit the company’s AI rules?
- Is it behaving? Is it doing what it normally does, or something new, at a new time, on new data?
If you can’t answer all five, you don’t control the agent. You’re trusting it.
A practical starting plan
For a fuller, step-by-step version, read how to secure AI agents in Microsoft 365 and Google Workspace.
Week 1: find them. List every app connection, service account and AI agent in Microsoft 365 and Google Workspace. Most companies find connections nobody remembers approving.
Week 2: give each one an owner. Tie every agent and app connection to a person who can say what it’s for. Anything without an owner is a candidate for removal.
Week 3: cut access to what the job needs. Replace shared admin accounts with scoped access. Remove permissions the agent hasn’t used.
Ongoing: review and watch. Set a review date for each agent, and watch for behaviour outside its normal pattern. Make sure you can switch any agent off in one step.
Where AI agent governance meets AI governance
Agents are one half of the picture. The other half is the AI tools people use directly: chat assistants, browser extensions, note-takers. Governing both from one place matters because they connect. A person approves an AI app, and that app starts running agents with the person’s access. Read more in our guide to shadow AI and non-human identities.
Who supervises the agents once they’re live? Read what guardian agents are and how the Praxis guard agent works.
How Centeye helps
Centeye Synthwatch, now in early access, gives every AI agent an identity, an owner and an off switch. It finds the agents working in Microsoft 365, Google Workspace and connected apps, ties each one to a human sponsor, shows what it can reach and what it actually did, and lets you revoke its access in one step. Centeye Lumen keeps the register of the AI tools people use, so people and agents are governed together.
Want early access for your clients? Book an intro call.
Figures cited from the CyberArk 2025 Identity Security Landscape and the IBM Cost of a Data Breach Report 2025.