AI assistant detection

Last reviewed: 2026-10-17

AI assistant detection estimates whether the customer is an automated personal assistant. It can recognize an assistant that introduces itself as AI and look for behavioral signals when it does not. Treat the result as a fallible estimate. A human may write formally or use translation, dictation or AI writing help.

Voice and Chat agents

  1. Select the model node and open Capabilities.
  2. Choose Add capability → AI assistant detection.
  3. Configure the Logic SLM auxiliary model and its connection.
  4. Keep the default thresholds initially, or adjust them using your own labeled tests. Publish the agent and start a new test conversation.

The agent can call get_counterparty_assessment to read the score, supporting customer quotations and when the assessment was updated. Its optional refresh input waits for pending evidence. Normal conversation continues while background assessments run.

Voice detection needs finalized customer transcripts. For native audio models, enable input transcription if the model does not provide it by default.

Copilots

Select the AI assistant detection node on the Copilot canvas and enable Monitor the customer continuously. Choose the model connection and the Persistent assessment card. If no card is listed, publish the updated Copilot cards catalog containing CounterpartyAssessmentCard.

This experimental monitor supports OpenAI and AWS Bedrock model connections. The initial live benchmark used OpenAI GPT-5.5; validate other models separately.

The node reads the original customer transcript before enhancement and the actionability gate. When it detects a possible AI assistant, the card appears and updates in place. The card stays visible with an updated assessment if evidence later suggests a human has taken over. Other Copilot prompts can use the same live variables on their next invocation.

Scores and live variables

Variable Use
counterparty_ai_likelihood A score from 0 to 1; initially empty. It is not a verified probability.
counterparty_ai_detected True when the detection threshold is reached. False can mean insufficient evidence.
counterparty_ai_status unknown, likely_ai or likely_human.
counterparty_ai_available False when an assessment could not be completed.
counterparty_ai_pending True while a newer customer turn awaits assessment.
counterparty_ai_assessed_through The customer turn included in the latest assessment.
counterparty_ai_updated_at When that assessment completed.
counterparty_ai_disclosure Whether the customer explicitly claimed to be AI.

These variables are maintained by the capability and cannot be set by the agent. Templated Voice/Chat prompts update as their values change. Use the pending and availability fields when freshness matters.

By default, the flag switches on at 0.8 after at least three customer turns. Explicit AI self-disclosure can trigger sooner. Two checks at 0.35 or below clear it, reducing flicker around a threshold. A timeout preserves the previous score and marks it unavailable.

The Detection prompt section allows a custom rubric. Test it against both human and AI conversations, including polite humans, professional representatives and assistants instructed to hide their identity. Different models can produce substantially different results. The initial benchmark is experimental; no accuracy level is guaranteed for your callers.

Detection does not authorize access to an account and does not indicate fraud. Continue the normal identity, consent and authorization steps. Configure any change in service behavior explicitly in your agent's instructions.