How Treater agents improve

Offline: new versions are proposed, screened and approved
Online: rewards shift traffic toward what's winning

Start with a goal

A software company wants demos with staffing agency owners. The catalog starts with one version of the agent.

GoalA 20-minute demo booked with the owner
Who2,000 staffing agency owners
AgentVersion A: prompt, voice, pace and backchannels

Reward every call

Each call is scored on the customer's goal and on how the conversation went.

Owner“We tried a scheduling tool last year, and honestly my recruiters—”
Agent“Mm-hmm.”
Owner“—they never really got used to it, and then the—”
Agent“Right, right.”
Owner“…yeah. Anyway, just send me an email.”
Demo bookedNo
DeferenceLow
SentimentDisengaged
Rules followedYes

The recording disclosure isn't scored. Decisions enforces it on every call.

Propose changes

An offline review reads the low-reward calls and proposes new configurations.

Agent talked over the owner with backchannels1 in 5 calls
Fewer backchannels while the other person is talking
Slower pace
Mirror the caller's energy
Calmer voice

Screen, then approve

Each proposal is simulated against past calls. A model predicts the reward it would earn, and the weak ones are dropped before the rest are manually reviewed or automatically introduced.

Fewer backchannels
Keep
Slower pace
Drop
Mirror energy
Keep
Calmer voice
Keep
The customer reviews the three survivors.Waiting for review

Real conversations shift traffic

Approved configurations share traffic in real time based on expected reward.

Share of callsDay 1
Demos per 100 calls3.0
Simulated example.