The Best AI for IAM Depends on Which AI You Mean

When I started looking for "the best AI for IAM", I found two different conversations happening under the same name.

The first is AI that helps identity teams do their job: assistants and agents that tidy up access policies, review permissions and spot risky sign-ins. The second is IAM for AI: giving AI agents their own identities and controlling what they're allowed to touch.

Most of the news in 2026 has been about the second one. Which tool is "best" depends on which problem you have and what you already run. So this post covers both, then gives my verdict.

Why this matters now

Identities that aren't people (service accounts, API keys, bots and now AI agents) outnumber the people in most organisations:

45 to 1 on average, and up to 144 to 1 in cloud-native environments, according to a Cloud Security Alliance whitepaper from May 2026.
109 to 1 in Palo Alto Networks' 2026 Identity Security Landscape, a survey of 2,930 security leaders. The same survey found that 9 in 10 organisations had a successful identity-related breach in the past 12 months.

Most organisations aren't managing these identities yet. The CSA paper cites research showing 51% have no clear owner for their AI identities, and 78% have no documented policy for creating or removing them.

The numbers vary because each vendor counts differently, and some of these vendors sell the fix. The direction is the same in every report, though: more machine and agent identities than ever, and far less oversight of them than of human accounts.

Meaning 1: AI that helps IAM teams

This is AI as a co-worker for the identity team. It reads your policies and logs, explains what's wrong, and in some cases fixes it. Every major vendor now has some of this.

Microsoft Entra with Security Copilot. The Conditional Access Optimization Agent runs every 24 hours. It finds new users, apps and agents not covered by a policy, and suggests fixes. It needs Entra ID P1. An Identity Risk Management Agent (in preview) investigates risky users in ID Protection. Suits anyone already on Microsoft 365.
SailPoint (Harbor Pilot, Atlas). AI agents you talk to in plain English. They guide access requests, right-size privileges and enforce policy. A Harbor Pilot Policy Agent is due around the end of 2026. Suits large organisations with heavy access reviews and audits.
Okta (Identity Threat Protection with Okta AI). Scores risk on every session and reacts when something looks off, such as a token being reused from a new location. Suits cloud-first, multi-vendor shops.
Palo Alto Idira (formerly CyberArk). Privileged access with zero standing privilege: nobody holds admin rights permanently, and access is granted just in time. Suits organisations where privileged access is the biggest risk.

Microsoft has the strongest cost argument. Security Copilot is now included in Microsoft 365 E5 and E7: 400 Security Compute Units a month for every 1,000 user licences, up to 10,000 a month. If you're already on E5, you can try the Entra agents without buying anything new. Microsoft says it plans to throttle usage over that allowance, with an option to pay US$6 per extra unit.

One thing I liked: when you turn on an Entra agent, it gets its own agent identity with fixed, listed permissions. So the AI that manages identity is itself managed as an identity.

Meaning 2: IAM for AI agents

AI agents read email, call APIs and change records, often with a shared API key that never expires and no named human responsible for them. This half of the market treats each agent as an identity in its own right: registered, owned by a person, given only the access it needs, and easy to switch off.

The vendors split into two camps. Issuers give identities to agents. Governors oversee agents that something else has already created.

The issuers

Microsoft Entra Agent ID. Agent identities with human sponsors, Conditional Access templates for agents (including blocking high-risk agents), access packages, and automatic clean-up when an agent is deleted. It works with non-Microsoft agents such as Amazon Bedrock and n8n. Generally available, and the agent registry is moving into Microsoft Agent 365.
Okta for AI Agents. Agent SSO swaps never-expiring keys for short-lived tokens, using Cross App Access. That's an open specification Okta started, now an official MCP authorization extension and an IETF draft. It's an emerging standard, not a ratified one yet. Okta for AI Agents also has rules for which agents can call each other, and access certifications for agents. Those three are generally available. An Agent Gateway, shadow-agent discovery and a gateway kill switch were announced in September 2026 as coming later.
Ping Identity for AI. Agent IAM Core, an Agent Gateway that enforces access at runtime (including for MCP), and agent detection through PingOne Protect. Generally available since March 2026.

The governors

SailPoint Agentic Fabric. Discovers agents, MCP servers and credentials. Adds prompt monitoring, a kill switch, just-in-time access and audit evidence reports. Announced at Navigate, October 2026.
Saviynt Identity Security for AI. Visibility, lifecycle management and runtime control for agents, alongside human and machine identities. Launched March 2026.

The big vendors have also been buying their way in, and all three deals have closed:

SailPoint completed its purchase of Entro Security in June 2026.
Okta completed its Permiso acquisition on 26 August 2026, after announcing it in July.
Palo Alto Networks completed its CyberArk acquisition in February 2026.

Even the analysts haven't picked a winner. Futurum's take on SailPoint's launch is that the real question is which approach buyers put in the path of an agent's actions first. That could be an issuer's gateway, like Okta's or Ping's, or a governor's identity graph, like SailPoint's. They also note a limit of kill switches: stopping an agent's session doesn't stop its goal, which can come back through a different credential.

My verdict

There is no single best AI for IAM. The best one is whichever fits where you are.

Already on Microsoft 365 E5: look at Microsoft Entra (Security Copilot agents and Agent ID). Security Copilot capacity is included in E5, the agents work right inside Entra, and Agent ID covers your own agents with the Conditional Access you already know. It's the easiest place to start.
Running a mix of clouds and SaaS apps: look at Okta. Agent SSO uses Cross App Access, an open specification rather than an Okta-only feature. Okta lists Claude, Cursor, VS Code, Zoom and Docker among the apps adopting it, though not all of them say their support is live yet. Anthropic's is in beta with Okta. Wider adoption is what would make it truly vendor-neutral in practice.
A large organisation with audits and access reviews: look at SailPoint or Saviynt. They aim to govern every identity they can discover or connect to, whether that's people, service accounts or agents, in one place. Agent discovery is still new, so check which platforms their connectors cover.
Most worried about admin and privileged access: look at Palo Alto Idira. Zero standing privilege matters most once agents start getting admin-level rights.

Personally, I'd start with Microsoft. It ties straight back to what I studied for SC-900: Conditional Access, ID Protection, least privilege. If you're on E5, you can try the agents without paying extra.

Whichever you choose, the basics don't change. Every agent gets its own identity, a named human owner, the least access it needs, and a way to switch it off. Most organisations aren't doing that yet.

The products above are changing month to month, so check each vendor's current documentation before you decide.