Sign in

Test the ai's configured behavior

An Admin exchanges test messages with the AI in a sandbox that runs the full configured pipeline — instructions, persona, sources, groundedness, forbidden topics, handover rules — without ever reaching a real customer or any channel.

When this happens: Admin opens Admin — AI Assistant — Test.

Step by step

  1. 1

    Open Admin — AI Assistant — Test, and optionally picks a channel to simulate (affects which capability restrictions the sandbox shows, UC-28.31).

  2. 2

    Types Do you ship to Moldova? and sends it in the sandbox.

    EasyCRM runs the same pipeline as Auto Reply — assembles context from the sandbox conversation only, calls IAiCompletionClient, applies the groundedness, forbidden-topic and handover checks — but writes nothing to any real Conversation or Message, and sends nothing to any provider. EasyCRM shows the resulting reply (or the fallback/handover outcome, annotated) in the sandbox transcript, together with which sources were used and whether it was grounded. EasyCRM records an AiInteractionLog with Test, latency, groundedness and any handover reason, for the Admin's own review.

  3. 3

    Continues the sandbox conversation to test multi-turn behavior, or clears it to start over.

Other paths

Test an unsaved draft

Open the sandbox from an unsaved Instructions/Persona edit (UC-28.3 , similar for other tabs). EasyCRM runs the pipeline against the in-memory draft instead of the saved settings, and labels the transcript Testing unsaved changes

If something goes wrong

Provider unavailable during a test

The provider call times out or errors. EasyCRM shows the same fallback outcome the real pipeline would produce, clearly labeled as a test result, and logs EasyCrm:Ai:ProviderUnavailable.

Good to know

  • The sandbox never creates a real Conversation, never sends through any ChannelSendJob, and is never visible to a customer.
  • The sandbox is scoped to the testing Admin only; other users' test sessions are independent.