Bias in Automated Quality Scoring

Automated evaluation applied to a workforce will work better for some employees than others unless someone checks. The mechanism is documented and the check is straightforward.

Where the bias enters

Transcription accuracy. Higher error rates for some accents and dialects mean required phrases are less often detected for those speakers, producing compliance failures that did not occur.

Language models applied to transcripts inherit the transcript's errors and may add their own, since training data over-represents some varieties of English.

Scorecard items themselves may encode norms of one variety of speech — pace, directness, formality — that penalise others.

Sentiment and emotion inference, which vary by cultural expression.

Training data for any scoring model, which if built from historical human evaluations inherits whatever bias those contained.

Why it matters more than in other applications

The output affects employment. Scores feed coaching, performance management and sometimes pay.

The disparity is systematic, not random, so it accumulates over time for the same individuals.

It is invisible without measurement, and it produces a plausible alternative explanation — that the affected agents perform worse.

It is a discrimination exposure in most employment law regimes.

Measuring it

Compute transcription accuracy by speaker group, as described in its own note.

Compute automated compliance failure rates by group, and compare with human-verified rates on a sample from each group. A gap between automated and human findings for one group is the disparity.

Compute dispute upheld rates by group. A higher upheld rate for some agents means the system is wrong about them more often, and it is the cheapest available signal because the data already exists.

Compare automated and human scores by group on the same calls.

Which groupings are lawful and appropriate varies by jurisdiction. Take advice, and consider using proxies that are operationally meaningful — queue, site, language — where demographic data cannot be used.

What to do with a finding

Suspend automated scoring for the affected items until the gap is addressed. This is the immediate step and it is the one that protects people.

Improve the inputs: audio quality, custom vocabulary, model adaptation on your own speakers.

Raise it with the vendor with your measurements, and ask what their per-group testing shows.

Keep the human in the loop for anything with a consequence.

Document the finding and the response. Evidence that you measured and acted is a different position from having never looked.

The governance arrangement

Someone outside the operational line should review the disparity analysis. The team running the programme is not well placed to audit itself here.

Annually is adequate, and after any model change.

Record the review, including a finding of no disparity, because the record of having checked is what matters when the question is asked.

The agent-facing part

Agents should be told that automated scoring is used, and that its accuracy is monitored.

The dispute route should be visible and used.

An agent who says the system mishears them should be taken seriously, and the dispute log should be checked for a pattern rather than the case being handled individually.

A programme where agents believe the system is unfair, and cannot demonstrate otherwise, has a problem regardless of whether the belief is correct. Measuring and publishing the accuracy by group is the only response that resolves it.

The annual fairness review

A defined review, conducted outside the operational line, producing a written finding.

Transcription accuracy by speaker group, measured on a fresh sample.

Automated determination rates by group, compared against human verification on a sample from each.

Dispute upheld rates by group.

Score distributions by group, for both automated and human evaluation.

Which groupings are used, and why they are lawful and appropriate in your jurisdiction.

The finding, including a finding of no disparity, which is worth recording.

The action, where a disparity is found, with dates.

Reviewed by someone who does not run the programme, because self-audit on this question is not credible and would not be treated as such if challenged.