Skip to content

AI Agents & Non-Human Identities

Before this page

Assumes you have read How Scoring Works. Brand new? Start at The words you need first.

Not every account belongs to a person. Applications, automation and AI agents hold access too, and nobody reviews them the way they review a colleague. Identity Atlas stores them as accounts with their own type, so you can tell them apart and score them with their own classifiers.

Account types for non-human identities

Type (principalType) What it is
ServicePrincipal An application's identity in the directory
ManagedIdentity An identity attached to a cloud resource and managed by the platform
AIAgent A service principal recognised as an AI agent
WorkloadIdentity An identity for a workload that signs in with a federated credential

How the Entra ID crawler decides the type

For each service principal, the first rule that applies wins:

  1. Entra ID marks it as a managed identity → ManagedIdentity.
  2. It carries one of the tags CopilotStudio, PowerVirtualAgents, AzureOpenAI, CognitiveServices, AgenticInstance or AgenticApp, or a tag that starts with power-virtual-agents- → AIAgent.
  3. Its display name contains copilot, openai, azure-ai or cognitive-service, or the separate word gpt or bot → AIAgent.
  4. Otherwise → ServicePrincipal.

The Entra ID crawler does not assign WorkloadIdentity. The CSV and SQL crawlers accept all four types, so a source that knows the difference can supply it.

Rule 3 is a name match, so check the outcome: an application called Robot Arm Bot becomes an AIAgent, and an agent with a neutral name stays a ServicePrincipal.

How they are scored

Non-human accounts go through the same four layers as every other account. One thing differs:

  • Layer 1 uses the Agents classifiers. Accounts of the four types above are matched against the Agents tab of the active classifier set. If the set contains no agent classifiers, the Users classifiers are used instead.

Layers 2 to 4 are identical for human and non-human accounts: the same membership, hygiene and propagation signals, with the same points. There is no extra score for being non-human.

To see the agent classifiers, open Admin → Risk Scoring, expand Risk Classifiers and choose the Agents tab. See Classifiers.