prepaid.ai docs
StatusOpen the console

Guides

A/B testing your agent

Switchback testing, and how to read its results honestly.

Testing here is a switchback test, and the distinction matters for reading your results.

One variant is live at a time#

The exposure window alternates on a timer, and every call is attributed to whichever arm was live when it started. Two callers ringing at the same moment hear the same thing.

Reading the verdict#

  • No verdict is offered before 30 calls per arm.
  • Confidence intervals are Wilson intervals, which stay inside 0–100% at the small numbers a single business produces.
  • Average handle time is shown but deliberately not tested — a p-value we did not compute would be worse than none.
Worth knowingStopping a test publishes the untouched configuration; it is not an undo that has to remember what it changed. Only one test can run per agent at a time.