AI Playground
The AI Playground runs a prompt against a real conversation's transcript, live, without persisting anything to the conversation. It is the prompt-design surface for the whole analytics stack: use it to validate a new global task, debug a tenant's override, or compare engines before a switchover.
The playground opens from the Playground tab of Administration › Conversation Analytics › AI Tasks, and from the Save and Test / Test buttons on AI task pages.
Experiment types
Five playground experiments exist, one per prompt style:
| Experiment | What it tests |
|---|---|
| Generic AI task | A free-form prompt with a text/JSON response — the general-purpose playground. |
| Auto QA form | An Auto QA prompt scored against a selected scorecard's questions. |
| Custom field | A prompt that extracts one or more custom-field values. |
| Sentiment score | The sentiment-scoring prompt. |
| Topic form | LLM topic extraction against your topic definitions. |
Workflow
- Pick a conversation that has a transcript — the playground's first step is a filtered conversation search.
- Edit the prompt — the task instructions and inputs, response settings, and the engine to run on.
- Run — the prompt executes against the conversation live and the result is displayed.
- Iterate, then optionally Save AI Prompt to write the tuned prompt back to the AI task (or the tenant's override) it belongs to.
Nothing is written to the conversation itself: no summary, score, topic, or evaluation is stored by a playground run. The only persisted change is the prompt, and only when you explicitly save it.
Permissions
Playground access requires permission to view transcripts, plus the permission for the experiment's resource (for example, AI task or evaluation-form access for those experiment types). Tenant administrators have the same playground within their own scope — see AI Playground in the Administration Guide.
Operator use cases
- Validating a global prompt change — before editing a global task that tenants rely on, reproduce the current behavior in the playground, apply the change, and compare outputs across a few representative conversations from different tenants.
- Debugging a tenant complaint ("the summary misses X") — open the playground, load one of the tenant's conversations, and iterate on the override prompt until the output is right, then save it to the override.
- Engine comparison — run the same prompt and conversation against two engines before repointing the AI Assistant job.