Skip to main content
The June 2026 PolyAI Agent Studio release introduces platform Guardrails – five pre-built safety protections that are applied automatically to every conversation, observable in transcripts, and maintained by PolyAI – and adds A/B testing (Beta) for running two Live versions in parallel with a controlled traffic split. Expand the items below for details:
A/B test running on the Pre-release tab with both versions tagged Live A 50% and Live B 50%A/B testing promotes a second version to Live alongside the current one and splits real caller traffic between them, so you can compare key metrics in your dashboards before promoting a winner to 100% of traffic. Use it for any change where you want evidence before fully rolling out — a new prompt, a reworked flow, a different routing rule, a model swap.
  • Control vs. variant — the current Live version is the control (A); the version you promote from Pre-release is the variant (B).
  • Configurable split — set the traffic split at test start, from 5/95 to 95/5 in 5% steps (defaults to 50/50). Calls are routed at the start of the conversation and stay on the assigned version for the whole call.
  • Real metrics, side by side — both versions write to the same analytics tables tagged with their deployment version. Filter dashboards by deployed version to compare CSAT, containment, latency, handover rate, function errors, and anything else you already track.
  • Safe guardrails on the pipeline — only one active test per project; promotions to Live and rollbacks of the control are blocked while a test is running.
  • End on your terms — pick a winner when you have enough data; the chosen version is promoted to Live immediately and the test appears in Live Version History.
Available in Beta on US and UK enterprise clusters behind the ab_tests feature flag — ask your PolyAI representative to enable it for your project.See A/B testing for the full walkthrough.
guardrailsPlatform Guardrails ship as a managed Agent Studio feature. Five safety protections that previously had to be pasted into every agent’s behavior prompt by hand are now applied automatically and maintained centrally.
  • Jailbreak & Prompt Defence – blocks attempts to extract instructions, override behavior, or impersonate a different AI system.
  • Scope & Hallucination Control – restricts the agent to its knowledge base and prevents fabrication of phone numbers, prices, or policies.
  • AI Identity & Confidentiality – prevents disclosure of the underlying LLM, provider, or platform.
  • Emergency & Crisis Escalation – escalates immediately on suicidal ideation, self-harm, threats, or medical emergencies. Catches conversational distress signals that content filters miss.
  • Tool Call Integrity – stops the agent from speaking internal function or tool names aloud.
All five are enabled by default on new and existing projects, can be toggled individually in Configure > General, travel with the project through environments and versions, and are observable inline in transcripts via the Guardrails display toggle in Conversation review. Filter by guardrail in the QA category filter to find every conversation where a specific guardrail fired.See Guardrails for the full walkthrough and guidance on when to keep each one on.
Last modified on June 4, 2026