Digital Intelligence Actor Classification

Actor Classification

When the Digital Intelligence enrichment runs in a RiskOS™ workflow, the Evaluation response contains an actorClassification object. This object tells you whether automation was detected in the session, whether that automation looks like a script or an AI agent, and, when we can tell, who or what is behind it.

The object is returned at data_enrichments[].response.digitalIntelligence.actorClassification, alongside the device, behavioral, and velocityMetrics objects.

{
  "data_enrichments": [
    {
      "enrichment_name": "Socure Digital Intelligence",
      // ...
      "response": {
        "digitalIntelligence": {
          // ...
          // highlight-start
          "actorClassification": {
            "summary": {
              "isAutomated": true,
              "isAgentic": true
            },
            "agentAutomationDetails": {
              "type": "personal_agent",
              "name": "muse",
              "indicator": "device"
            }
          }
          // highlight-end
        }
      }
    }
  ]
}

How actor classification works

Actor classification looks at the device data and the on-page behavior (typing, clicks, field fills, and timing) collected by the Digital Intelligence SDK for the di_session_token. Each evaluation is classified on its own, so a session can receive different results across evaluations. For example, a person may start a session and an AI agent may take over later.

Classification answers three questions:

• Is the session automated? (isAutomated)

• If it is automated, is it driven by an AI agent or by a script? (isAgentic)

• When it can be shown, which agent or which script operator is behind it? (agentAutomationDetails, scriptAutomationDetails)

Fields

Empty fields are omitted from the response.

• summary:

  • isAutomated: true if software, not a person, drove the session. false means no automation was detected. It does not prove a person was present.
  • isAgentic: Only set when isAutomated is true.
    • true: an AI agent drove the session, meaning software that reads the page and decides what to do next, usually on a person's behalf.
    • false: a script drove the session, meaning software that follows fixed steps.
    • Empty: automation was detected, but the evidence fits a script and an agent equally.
  • scriptAutomationRisk: Only set when isAgentic is false.
    • low: we know who runs the script and confirmed it from the network it came from, such as an account aggregator or a known search crawler. This is usually expected traffic.
    • medium: a script we can't name or confirm. Whether it is wanted is your decision.
    • high: the script claims to be a known crawler but comes from a network that crawler doesn't use. Someone is pretending.

• scriptAutomationDetails: Present when isAgentic is false or empty.

  • type: account_aggregator, accessibility_reader, ads_crawler, search_crawler, or unknown.
  • name: the operator, when identified. For example: plaid, mx, yodlee, googlebot, bingbot, google_read_aloud.
  • indicator: how the operator was identified.
    • isp, isp+device, ua+asn: confirmed by the operator's own network.
    • asn+device: inferred from a known network and device setup.
    • device: the device setup matches, but the network doesn't confirm it.
    • ua_only: the network contradicts the claimed identity.

• agentAutomationDetails: Present when isAgentic is true or empty.

  • type:
    • personal_agent: a known consumer AI agent acting for its user.
    • vision_agent: an agent that works from screenshots and types and clicks like a person at a computer.
    • dom_agent: an agent that reads the page's code and fills fields directly.
    • unknown: the kind of agent couldn't be determined.
  • name: the personal agent, when identified: muse, instinct, grok, or grok_bot.
  • indicator: device (the device matches a known agent), device_partial (part of a known agent's device setup matched), or behavior (classified by how it types, clicks, and pauses).

• inconclusivenessReason: Present only when summary is empty.

Examples

Not automated

No automation was detected in the session.

"actorClassification": {
  "summary": {
    "isAutomated": false
  }
}

Known account aggregator

A script ran the session, and we confirmed from its network that it is MX, an account aggregator. The risk is low because this is usually expected traffic.

"actorClassification": {
  "summary": {
    "isAutomated": true,
    "isAgentic": false,
    "scriptAutomationRisk": "low"
  },
  "scriptAutomationDetails": {
    "type": "account_aggregator",
    "name": "mx",
    "indicator": "isp+device"
  }
}

Unknown script

A script ran the session, but we couldn't identify who runs it. The risk is medium.

"actorClassification": {
  "summary": {
    "isAutomated": true,
    "isAgentic": false,
    "scriptAutomationRisk": "medium"
  },
  "scriptAutomationDetails": { "type": "unknown" }
}

AI agent identified by behavior

An AI agent ran the session. We couldn't name it, but its typing and timing match an agent that fills fields directly through the page.

"actorClassification": {
  "summary": {
    "isAutomated": true,
    "isAgentic": true
  },
  "agentAutomationDetails": { "type": "dom_agent", "indicator": "behavior" }
}

Inconclusive

The session used an SDK older than 2.9.0, so no classification could be made.

"actorClassification": {
  "inconclusivenessReason": { "code": "IC_002", "message": "Old SDK Version" }
}