The call dropped. You should still get the lead.

An eighteen-minute call once produced nothing, because the agent tried to hand off the whole job at once. Fixing that changed how we build every agent that has to write something down.

A contractor spent about eighteen minutes on the phone dictating a large remodel estimate — demolition, three flooring types across fifteen rooms, a master bath. The agent handled the conversation well. It captured the scope and read it back to him correctly.

WORKFLOW AUDITSAFTER THE CALL, NOT DURINGhand off mid-callnothing savedCALL STARTCALLER HANGS UPfull transcriptrecord · CRM · emailSTILL COMPLETESThe transcript always survives.Do the work after the call, not during it.

Then it tried to pass the entire job to the backend in a single handoff, and produced nothing at all. It couldn't pack fourteen minutes of detail into one structured payload. So it started re-narrating the job from the top, rambled, apologized, went quiet — and the call timed out on silence.

From his side, eighteen minutes of careful dictation had simply vanished.

That's not a software bug. That's a trust problem, and it's much more expensive.

Build it after the call, not during.

— The transcript always survives.

The fix: stop doing the work during the call

Every call produces a complete transcript the moment it ends. That transcript is the most reliable artifact in the whole system — it survives rambling, interruptions, dead air, and hangups.

So we stopped asking the agent to repackage a long conversation in real time. Now, when a call ends, the platform hands the full transcript to a single backend step that does the actual work: builds the record, writes to the CRM, sends the email.

Four things get better at once. No mid-call latency: nothing happens while the customer is sitting there listening to silence. Early hangups still count: if the caller drops at minute nine, the transcript still exists, so the work still completes. No duplicates: the whole thing is keyed on the call ID, so a retry can't create two records, and where it has to match an existing customer it matches on a strong key like phone or email, never on a name. No model constraints: the agent doesn't need to be good at structured handoffs anymore, it just needs to be good at talking.

We paired it with one prompt rule: submit once and stop. Give the caller a single line — "you'll get it by email, feel free to hang up" — and never re-narrate the job after the read-back. That re-narration is what stalled the original call.

What this means for you

Here's a question worth asking any vendor: what happens if my customer hangs up halfway through?

If the answer is "then we don't capture it," the work is being done in the wrong place. It's being done during the conversation, where it's fragile, instead of after it, where it isn't.

The moment a call ends shouldn't be the moment the information is at risk. It should be the moment the work starts.

Want us to look at where a phone agent could take work off your plate? Grab a free workflow audit.

← Back to all posts Free Workflow Audit