QA That Used to Sample 5% of Calls Can Now Review 100%

What AWS Showed at re:Invent

At AWS re:Invent 2025, Amazon Connect unveiled 30 major launches around agentic AI, framed around what VP Pasquale DeMaio called the core distinction: “agentic is really about action,” an AI agent’s ability to actually do something on a customer’s or company’s behalf, not just summarize or suggest. Two customer case studies grounded the announcement in real numbers. Centrica, a 200-year-old energy firm running nearly 11,000 agents across its UK contact centers, cut handle times from 140 to 87 seconds and lifted NPS by 89 points on some customer journeys after deploying generative AI. Priceline reported saving 50 seconds per interaction through automated call summarization alone.

Why This Matters for BFSI and Collections

The most transferable result here is not the handle-time reduction; it is what Priceline is doing with quality assurance. Traditional QA processes sample only 3 to 5% of an agent’s calls weekly, with coaching delivered days or weeks after the interaction happened. Priceline is moving toward 100% of interactions reviewed by AI, with feedback delivered seconds after the call ends. For collections operations, this changes what oversight can actually mean. A compliance review that currently catches a problematic payment conversation days after it happened (once the damage to a customer relationship or a regulatory record is already done) could instead catch it while the pattern is still forming, whether that is an agent unable to apply for hardship pathways correctly or a script that consistently triggers complaints.

What the Numbers Do Not Say Out Loud

Neither Centrica nor Priceline are financial services companies (one is travel booking, the other is energy retail), so these results describe an operational mechanism, not a proven BFSI outcome. Worth noting too: Centrica’s own agentic voice-to-voice pilot, using Amazon’s Nova Sonic model, is still handling only 8% of queries and is explicitly “not live in production yet,” according to cloud director James Boswell, despite testing well. Even a company running an agentic, action-taking pilot is choosing to move slowly on the piece where AI actually acts rather than assists. That distinction between automation that assists a human (summarization, live transcription, QA scoring) and automation that acts autonomously is exactly where collections needs to be most careful, since acting on a payment plan carries different stakes than transcribing a hotel name correctly.

The Practical Read

For collections and BFSI operations, the actionable idea here is not “deploy voice-to-voice AI”; it is “close the QA and coaching gap first”. A collections agent who gets a debrief immediately after a difficult hardship call (what worked, what failed, what created escalation risk) improves faster and more consistently than one reviewed on a random 5% sample weeks later. That is a lower-risk, higher-leverage place to start than full autonomous action, and it builds the real-time visibility into agent and AI performance that regulators expect anyway. Automate the observation first. Let the action-taking follow once that foundation is proven.

[Read the full report]

Related Post

Only 11% Contact Centers Trust AI to Work Alone

The Adoption Picture Five separate 2026 research reports now paint a consistent picture of where contact center AI actually stands. Valoir found more than 95% of organizations run at least one AI or automation capability, with AI agents leading adoption at 61%, followed by copilots at 55% and knowledge search at 54%. ContactBabel found 21% […]

Gartner Predicts Half of AI-Driven Layoffs Will Be Reversed by 2027

A major Gartner forecast has delivered a striking reality check for corporate leadership: by 2027, 50% of companies that cut customer service staff and attributed the reduction to AI will rehire people to do similar work under different job titles. While AI-driven layoffs have captured significant media attention, the actual operational experience on the ground […]

Two-Thirds of Enterprises Are Scaling AI Without Any Way to Measure If It Is Working

The Finding New research from TELUS Digital’s Enterprise CX AI: 2026 Global Survey, polling 815 enterprise decision-makers across 12 countries, finds that 61% of organizations now spend over $10 million annually on CX delivery. Yet only 32% have the automated QA infrastructure needed to actually measure how that AI is performing. As Peter Ryan, President […]