A call coach can produce clever questions and still be useless.

The question has to arrive while the conversation can still benefit from it. It also has to know enough about the person and project to say something better than “ask about their pain points.” That was the reason I built Live Call Coach.

Observed test evidence from the current candidate. It proves what failed and what improved; it does not prove better calls or sales outcomes.

Prove the advice before building the plumbing

I started with a replay test instead of microphone capture. One real discovery-call transcript contained 96 turns. The replay found 27 moments where another question or follow-up might have helped, including four recurring questioning mistakes.

That test asked the important question cheaply: can the context and coaching rules produce suggestions worth seeing at all? A generic live-coaching product already exists in the market. My version only deserves to exist if knowledge from earlier work makes the next question materially better.

Then I built the smallest honest live candidate

The personal Windows app captures the two sides of a call, transcribes audio locally and shows suggested questions in a separate window. Audio stays on the machine. A person still decides what to ask, and real-client use remains blocked until the consent and legal boundary is cleared.

The installed candidate passed its offline checks. That proved packaging and local behavior, not whether it helped during an actual conversation.

The live test changed the verdict

During the latest test, the suggestions followed the conversation more closely than earlier versions. My verdict was still clear: “not bad,” but far too slow. One ended call produced only one generic cue, and seven of eleven model attempts failed the evidence check.

The spending limit created a second failure. A conservative per-call budget could stop the coaching before the conversation ended. A feature that disappears halfway through a call is not dependable enough to become part of the call.

What happens next

The next pass is not another model shopping exercise. It has to measure where the delay actually comes from: transcription, request spacing, generation, retries or the time taken to display the cue. Cost has to be measured beside usefulness so the system can last through the intended call.

The candidate is installed, but the product has not passed live acceptance. There is no claim of better calls, more sales or client results. The useful result so far is knowing exactly what has to improve before more building is justified.