Skip to main content
Audact
See the offer

Method, not results

Retell to Audact, measured rather than asserted.

Not running · the control arm no longer exists

This page publishes a measurement design, and the measurement is not running. It was written for a side-by-side comparison in which Tirza IP Limited — a related party that shares a founding shareholder with Audact — would split production traffic between its existing Retell setup and Audact. On 11 August 2026 Tirza stopped as an independent commercial product and moved off Retell entirely, which removes the control arm the design depends on. No traffic was ever split, no data was collected, and there are no results to expect. What remains is a full migration to Audact on a related party: internal validation, not customer proof. The design stays published because a method written down before the outcome is worth more than one written after, and because deleting it would be the more convenient of the two options.

Disclosure

A related party, stated before anything else.

Tirza IP Limited and Audact share a founding shareholder. A measurement run on a related party is internal validation and not customer proof. That was true when the design was written, and it stays true now that the comparison is not running.

Method

How the comparison was set up.

  • Element
    Traffic split
    How it was fixed
    50/50, randomised by a hash of the caller identifier and locked for the length of the study window.
  • Element
    Control and treatment
    How it was fixed
    The existing Retell setup was to be the control and Audact the treatment. Same conversation design, same script.
  • Element
    Jurisdictions
    How it was fixed
    The Netherlands, Belgium, Germany, Italy and France: the full footprint, not a favourable subset.
  • Element
    Primary metrics
    How it was fixed
    Completion rate, average handle time, post-call satisfaction, escalation rate to a human.
  • Element
    Secondary metrics
    How it was fixed
    Disclosure delivery rate, evidence-chain coverage, exceptions per thousand calls.
  • Element
    Window and publication
    How it was fixed
    As designed: collection to the end of Q4 2026, publication with confidence intervals in Q1 2027. The window never opened — see the note above.

What was to be compared

Two things that would have differed, and one that would not.

The conversation was to be held constant on purpose. What differs between the two setups is what happens around it: the approved wording is checked before the model speaks rather than sampled afterwards, the AI disclosure is versioned per language and jurisdiction and its position in the call is recorded, and every governed call leaves a receipt.

That means the interesting question is not whether governance improves completion rate (there is no reason it should) but whether it costs anything. A conversation-quality measurement is the honest way to find out, and publishing the method first is the only way to make the answer worth reading.

Questions

Why publish a case study with no results in it?
Because the method is the part that can be checked, and because publishing it first removes the option of choosing which outcome to tell afterwards. Here it left nothing convenient either: the window never opened, and this page says so at the top rather than quietly disappearing.
Will you report the numbers if they are bad?
There are no numbers: the comparison never ran. The commitment stands for any comparison that does run — published with its confidence intervals and its full method, whichever way it falls. A comparison you can only publish when it flatters you is a comparison nobody should have believed in the first place.
Does this count as customer evidence?
No. Tirza is a related party with a shared founding shareholder. It is internal validation, and it is labelled that way here, on the Tirza page, and on the customers page.

See what the receipts look like

The artefact this design was built to compare is the one a governed call produces. Recompute three of its fields yourself, without a key and without an account. And the fourth, inclusion in the Merkle root, is checked against the published root with the inclusion proof from the call's Article 50 report.

Open the receipt explorer

Related