When COVID sent telesales agents home, the disclosure audit went to full volume.
Every telesales call audited automatically for mandatory disclosures and explicit consent, with human review for low-confidence cases, built for American Express via IBM in 2020.
In March 2020 telesales agents moved to work from home with minimal live supervision. Agents under pressure to close could skip or rush the mandatory disclosures (APR, fees, rewards conditions) and the explicit customer consent. Manual QA sampled only a small fraction of calls, so a regulatory finding could surface long after the fact.
I saw the risk at the onset of COVID, created the proposal and solution architecture, and after approval directed delivery to production. The pipeline: ingest call recordings; transcribe with speaker diarization, word-level timestamps and a custom language model trained on product names and the exact disclosure wording; detect each required disclosure against a canonical library; assess whether it was read legibly using speaking rate, completeness and transcription confidence; classify the customer's response as affirmative, negative or ambiguous and confirm it followed the disclosure; score each call with reasons; route failures and low-confidence calls to human review with jump-to-timestamp playback; report compliance trends with PII redaction, encryption, role-based access and a full audit trail.
Reached production. Compliance QA moved from a limited manual sample to automated audit across the full telesales call volume, reducing regulatory and penalty exposure and catching non-compliant behavior within days rather than at audit time. The knowledge base records no quantified metric for this platform and none is published.
Classical speech and language services, 2020. No generative AI.
The governance and auditability work that makes an AI system acceptable in a regulated environment was my job before the models were generative.