Rehearse multi-turn conversations against the real assistant — same prompts, same tools, zero side effects.
Why this matters
Grade outcomes automatically: bookings proposed with the right details, questions answered inside the trust rules.
Real conversations that go wrong — a thumbs-down, a human takeover, an abandoned deposit — become draft rehearsals automatically.
Compare a candidate model or setting against the current one on identical conversations before anything changes in production.
Implementation workflow
A clear path from setup to production-grade performance.
Review the rehearsals drafted from your real conversations, or script your own.
Activate the ones that matter — each carries the customer turns and a checklist.
Run all active rehearsals whenever the assistant, knowledge, or model changes.
Read the graded results per conversation before any change meets a customer.
Expected outcomes
Side effects
Zero
Rehearsals run in a sandbox — no bookings created, no emails sent, no charges made.
Test source
Real conversations
Failed real interactions are drafted into rehearsals automatically, ready for review.
Comparison
Side by side
Candidate models and settings run the same conversations as the current setup.
Frequently asked questions
Does a rehearsal create real bookings or send real messages?
No. Rehearsals run in a sandbox: the assistant behaves exactly as it does in production, but booking, lead, and messaging actions are captured for grading instead of executed.
Where do rehearsals come from?
Two places. You can script them by hand, and Amfion drafts them automatically from real conversations that went wrong — a thumbs-down on an answer, a conversation handed to a human, or a booking abandoned at the deposit step.
Does Amfion use Rehearsal Studio itself?
Yes. Changes to how the assistant retrieves knowledge, remembers long conversations, and selects tools are tested through Rehearsal Studio before they ship — the same runs every account can trigger, with a monthly allowance of included runs.