---
title: "AI engine"
description: "Answer-rate prediction, real-time agent assist and automated QA scoring. The same models run before, during and after every call Firetone handles."
url: https://firetone.com.au/ai
site: Firetone
provider: RemotIQ Pty Ltd
keywords: "AI contact centre, conversation intelligence, real-time agent assist, call centre QA automation, answer rate prediction, speech analytics, compliance monitoring, explainable AI"
---

# The same models run before, during and after the call.

Firetone treats a call as three moments where a decision improves: who to dial and when, what to say while the line is open, and what the business learns after.

## In the call path, not alongside it

Most contact centre AI is a reporting layer. Calls happen, recordings pile up, a model reads them later. Useful for trends, incapable of changing any call it analyses.

Firetone's models sit in the path. Prediction runs before dialing and changes who is dialed. Assist runs mid-call and changes what the agent says. Compliance flags a miss the day it happens.

## What the models actually do

### Answer-rate prediction

A per-record probability of a live answer, used to decide dialing order.

### Best time to call

Contact-level reachability windows that re-order the list.

### Lead and risk scoring

Routes the records worth a senior agent to one.

### Real-time agent assist

Live transcription with next-best action, objection responses and required wording on screen mid-conversation.

### Conversational intent capture

Natural-speech understanding that completes routine requests and escalates the rest with context attached.

### Live sentiment

Escalation alerts to a supervisor while the call is still open and still recoverable.

## Why explainability is a requirement, not a feature

A decision you cannot explain is a decision you cannot defend. Someone may have to account for why one account holder was called at seven and another was not called at all.

So every score lists the factors behind it, in the console and the export. A 0.74 breaks into its contributions. Supervisors can disagree, override, and pin thresholds by hand.

## What scoring every call changes

Traditional quality assurance is a sampling exercise, and everything about it is shaped by that constraint. A reviewer listens to a handful of calls per agent per month because listening is slow. Coaching is therefore retrospective and thin. Systemic problems stay invisible until they are large. An agent's score depends substantially on which calls happened to be pulled.

Scoring all of them changes the questions that can be asked. Not 'how did this agent do on four calls' but 'which disclosure is missed most often across the floor, is it missed more on long calls, and did it start after the script change three weeks ago'. Those are process questions with process answers, and they are invisible at a two percent sample.

## AI engine questions

### Is our call data used to train your models?

No. Recordings and transcripts are tenant-isolated and are not used to train shared models, in any deployment mode.

### Can we see why the AI scored something the way it did?

Yes, and this is deliberate rather than incidental. Every prediction lists the factors that produced it, in both the console and the export.

### What does agent assist actually show during a call?

Live transcription of the conversation, the next-best action or offer given what has been said, responses to the objection just raised, and any wording your scripts require at that point — for example a mandated affordability line before an arrangement is confirmed. It also shows live scores such as promise-to-pay likelihood and script compliance as they update.