TL;DR: Your identity program was designed for a knowable world: employees in the IdP, applications in the catalog, service accounts on a list. On The Non-Human & AI Identity Podcast, C1.ai CEO Alex Bovee argued that this world is gone, and that the instinct to file agentic AI under the existing NHI program misreads the problem.
- The old assumption: identity security meant humans using SaaS applications, a bounded population touching a bounded catalog.
- What broke it: in Bovee's words, "the world is very, very, very messy right now," with credentials on endpoints, vibe coded MCPs, and agents landing in all sorts of places.
- The tempting shortcut: agents authenticate like service accounts, so extend the NHI program and move on.
- Why it fails: agents bring their own identity problems, from delegated authorization to standalone agent identities that need real-time authorization. It's "a blurring of the boundaries."
- The shift that follows: "from orchestration to runtime," security that's enforced as actions happen rather than cleaned up afterward.
For twenty years, identity security had a stable shape. Employees lived in the directory, applications lived in the catalog, and the job was connecting the two safely: sign-on, provisioning, deprovisioning, the occasional service account tracked in a spreadsheet nobody loved.
Asked on The Non-Human & AI Identity Podcast whether we've built identity security for a world that no longer exists, Bovee agreed: "I think generally that is the case, yes. The surface area now in businesses is no longer about humans using SaaS applications and some on-premise technology."
It's now also about humans leveraging agentic identities and harnesses to build software, use credentials, deploy workloads, and reach data, all at machine speed. His conclusion points at architecture, not tooling: "There's just fundamental shifts we have to move from, from orchestration to runtime. We have to be a lot less reactive in our security."
What actually changed about the enterprise surface area?#
Two things, compounding each other.
The first is volume. Non-human identities were already outrunning human ones, but agentic AI changed the slope of the curve. Every agent carries an identity, most agents carry several credentials, and agents are cheap to create in a way employees and even applications never were.
The second matters more: who creates them. In the old world, identity creation had choke points. IT provisioned accounts, platform teams minted service accounts, and the process, however imperfect, meant someone wrote things down.
Bovee's description of the current state is blunter: "The world is very, very, very messy right now. There's credentials on endpoint. There's vibe coded MCPs. People are building vibe coded applications and agents and putting those in all sorts of places."
Elsewhere in the conversation he described AI as "turning every single person into a builder," a Cambrian explosion of new workloads operating the business. And the pressure cuts against caution, since adoption is a board-level mandate: "it does feel a little bit like we're racing towards a cliff."
This is the environment we described in You Can't Govern the AI Agents You Can't Find: identities entering sideways, through teams shipping fast. The shadow AI problem and the identity problem are converging into the same problem.
Why isn't agentic AI security just NHI security with more rows?#
Asked whether agentic AI security equals NHI security, Bovee's answer was "not necessarily," and his reasoning starts with the term itself: NHI means different things to different people. Ask five practitioners and you'll get five definitions, some centered on service accounts, some on keys, tokens, and API credentials.
Agents, he argued, bring their own set of identity problems that don't map cleanly onto either definition. Some agents act on behalf of a person, the personal productivity kind, where the problem is delegated authorization: the agent's effective permissions have to trace back to the human it's working for. Others are enterprise, SaaS, or infrastructure agents with standalone identities of their own, which look more like service accounts, except that you need to authenticate them and authorize their activity in real time rather than trust the credential and walk away.
"Fundamentally, agents are non-humans," he said, "but they do act and behave sometimes like humans. I think it's a blurring of the boundaries for sure."
That blurring is why the NHI playbook is necessary but not sufficient. Securing a service account means securing a credential attached to predictable code, which is inventory-and-hygiene work that mature NHI programs do well. An agent is a credential attached to behavior that's composed at runtime, so the controls have to move from the identity's lifecycle to the identity's actions: authorization decided per action, access scoped per task, every call attributed and logged.
Where should an identity program start over?#
Not from scratch, but from a different first question. The old program started from the directory and asked what each known identity should access. The new one starts from the environment and asks what's actually operating in it, because the defining feature of the new world is that creation outruns registration. Discovery first, then ownership, then runtime governance, the arc we've been building across this series.
How does C1.ai fit?#
C1.ai maps every identity in the environment, human, service account, and agent, including the ones nobody registered, then governs them at the speed they're created rather than the speed of a quarterly campaign. AI-native identity governance for the builders' era: find everything, own everything, and decide access where the actions happen.
The world that identity security was built for isn't coming back. The programs that thrive will be the ones designed for the one that replaced it.
Listen to the full hot-seat episode, or see how C1.ai governs the new surface area.
