Move cursor | Click to ripple
Responsible AI

How QEval keeps its scoring fair and accountable.

Fairness in automated scoring is something you can measure, not just promise. Here is how QEval keeps scores accurate, explainable, consistent, and private, and what we deliberately do not do.

The principle

Fairness you can measure.

An automated score is only fair if it is accurate, explainable, consistent across every conversation, and built on data you control. Those are testable properties, not promises.

AccurateExplainableConsistentPrivate
How it works

Four ways we keep scoring honest.

Pick one. Each is a mechanism you can check, not a value statement.

One scorecard, applied the same way every time

A score should not depend on which reviewer looked, what day it was, or whether the agent was a person or an AI. QEval applies one scorecard and one set of rules to every conversation.

Human agent ACard disputeA-Same scorecard
Human agent BCard disputeB+Same scorecard
AI agentRefund requestB+Same scorecard
Human agent CCollectionsA-Same scorecard

Scores are calibrated against people, and watched for drift

QEval scores are checked against trained human reviewers on a continuous sample. The gap between the model and the human is held under 2%. When it drifts past that, the model is recalibrated.

Average drift from human reviewers: 1.1%
0%2% band5%
Within the 2% band. No action needed.

Every score points to the moment that earned it

No black-box grades. Each score links to the exact turn in the conversation behind it, so an agent, a supervisor, or an auditor can check it in seconds.

Compliance 96earned by this moment
Agent, turn 3For your security, I will not ask you to read your full card number.

Your conversations stay yours

Personal information is redacted at ingest, before any model sees it. Scoring runs on QEval's own closed mixture of experts, so your conversations never enter a third-party model's training loop.

Step 1Raw conversation
Step 2PII redacted at ingest
Step 3Scored by QEval MoE
Step 4Evidence-pinned audit trail
Guardrails

What QEval does not do.

The clearest way to describe responsible scoring is to be plain about the lines we do not cross.

Data
We do not train third-party models on your data
Scoring runs on QEval's own closed mixture of experts. Your conversations never enter an outside foundation model's training loop.
Explainability
We do not hand you black-box scores
Every score is pinned to the moment that earned it, so it can be checked and disputed with evidence, not opinion.
Accuracy
We do not set accuracy once and walk away
Scores are calibrated against human reviewers continuously, and the model is retuned when the gap drifts past 2%.
Fairness
We do not score people on who they are
A score reflects what was said and done in the conversation against your scorecard, not the personal characteristics of the agent or the customer.
Consistency and bias

The same standard, held across everyone.

Bias in scoring usually creeps in through inconsistency: who reviewed, what mood, which day. A model removes that variance only if it is watched for new variance of its own.

Uniform
Same rules, every conversation
One scorecard is applied to every conversation and every agent. There is no end-of-shift leniency, and no reviewer who happens to grade harder than the rest.
Watched
Monitored across teams and time
Scores are checked for systematic skew by team, queue, and shift. If one group scores differently with no defensible reason, that is a signal to examine the model, not the people.
Human-set
People still hold the bar
Reviewers are not removed. They define the scorecard the model calibrates to, and they adjudicate the edge cases. The model handles the volume so people can focus on the hard calls.
Questions

Responsible AI, in plain terms.

Do you train AI models on our conversations?

No. Scoring runs on QEval's own closed mixture of experts, and your conversations never enter a third-party foundation model's training loop. Personal information is redacted at ingest, before any model processes the conversation.

How do you keep scoring from being biased or inconsistent?

One scorecard and one set of rules are applied to every conversation and every agent, human or AI. Scores are calibrated against trained human reviewers, so the result does not depend on reviewer mood or which 2% of calls happened to be sampled.

Can we see why a score was given?

Yes. Every score is pinned to the exact moment in the conversation that earned it, so an agent or an auditor can verify it directly rather than trust a number.

What standard do you hold yourselves to?

QEval is built to ISO 42001 for AI management, alongside SOC 2 Type II, ISO 27001, PCI DSS, HIPAA, and GDPR. Accuracy is held to the contractual 94%+ service level.

Does QEval replace human reviewers?

No. People define the scorecard, calibrate the model against their own judgment, and handle the genuinely ambiguous cases. QEval does the volume, scoring every conversation, so your reviewers spend their time on the hard calls instead of sampling.

On your data

See it run on your conversations.

Start a pilot and we will score a sample of your conversations, show you each score pinned to its evidence, and walk your security and quality teams through how it is calibrated and protected.

Contractual commitments

Four numbers no peer publishes.

94%+
Accuracy SLA
Written into the master agreement
30 days
Deployment
Money-back guarantee
60 days
Exit clause
Cancel with notice, no penalty
120 days
ROI
Documented customer-average outcome