Amazon Connect at re:Invent: AI Agents and a Unified Platform Reshape Contact Centers
Most contact centers want AI to cut handle time, lift CSAT, and keep costs predictable. The issue: legacy stacks need custom work for every new feature, and point tools break down on complex questions. Amazon Connect is taking a different path with a single platform, embedded AI, unified data, and per-minute pricing across both basic and advanced features.
At re:Invent, AWS leaders reinforced a simple message: one engine for routing, context, analytics, and assistance across voice, chat, and automation. No pricing guesswork per feature. No stitching together separate voice and digital systems just to get a clear report.
Why this matters for support teams
Customers expect fast, relevant, and empathetic help-without repeating themselves. If your data is split by channel or vendor, your agents pay the price with longer calls and higher repeat contacts. Connect is built to keep context as the interaction moves from bot to chat to voice to a live agent.
Usage is billed by the minute whether you're doing basic routing or running real-time recommendations and summaries. That lowers the mental math of "can we afford this feature?" and puts the focus back on outcomes: time saved, issues resolved, and customers retained.
The architecture: one brain, consistent data
Amazon Connect processes voice, chat, and automated interactions through a single routing and context engine. Both human and automated interactions are recorded on the same schema, so you can apply one set of quality metrics everywhere instead of running parallel evaluation systems.
Traditional centers manually review 1%-5% of interactions. Connect runs AI across all of them for real-time signals and coaching. Plenty of vendors claim this, but the value is clear when QA moves from sampling to full coverage.
Learn more about Amazon Connect
Voice AI without duct tape: Nova Sonic
Connect now supports voice-based AI agents using Amazon Nova Sonic. Instead of wiring together separate speech-to-text, language, and text-to-speech components, Nova Sonic handles speech understanding and generation in one model. The result: more natural conversations and lower latency.
It also plugs into Amazon Bedrock knowledge bases to ground answers in your policies, pricing, and product data-reducing hallucinations and repeat contacts.
Proof at scale: Centrica's rollout
Centrica, a UK energy provider with roots back to 1812, moved from fragmented tools to Amazon Connect across all UK contact centers. About 11,000 agents now run on one platform with integrated AI. The shift also brought support in-house, speeding up delivery and ownership over outcomes.
- Average handle time dropped from 140s to 87s via AI auto-wrap and better routing
- Net Promoter Score rose 89% across specific journeys with escalation detection and vulnerability flagging
- Chat fulfillment rate doubled overnight after switching from ML to generative models
- After-call work shrank by ~30 seconds, freeing time for cross-sell between brands
- Management assessments launched in weeks, not months, using built-in evaluation frameworks
Two real-time interventions stood out. The system flags customer escalations agents might miss, preventing satisfaction drops. It also detects vulnerable customers struggling with payments and alerts agents to offer the right assistance, in the moment.
Beyond Connect, Centrica used Bedrock to analyze complaint transcripts, identify root causes, and reduce complaint volume by 28%. Nova Sonic voice agents are in pilot with ~8% containment-promising, but deployed carefully.
Scale is hard: what trips teams up
Many generative AI pilots never reach production. The usual culprits are fuzzy goals, missing metrics, weak data foundations, and poor governance. Fully autonomous agents are still rare because live failures are costly and risky.
Amazon Connect's stance: start with agent assistance, then expand automation as confidence grows. The platform centralizes data and workflows to simplify integration, but if you ever move off, recreating deep, native connections will take effort.
Practical playbook for contact center leaders
- Quantify the pain: handle time, repeat contacts, transfers, escalations, and containment. Tie each to cost and CSAT.
- Start small: pick 2-3 use cases (e.g., after-call summaries, knowledge retrieval, automated QA). Benchmark before-and-after.
- Fix data early: map sources, resolve duplicates, and set rules for data freshness and access.
- Governance beats intuition: define escalation paths, human-in-the-loop review, compliance checks, and rollback plans.
- Measure what matters: AHT, FCR, NPS/CSAT, containment, agent effort, and QA coverage. Review weekly, not quarterly.
- Iterate fast: ship improvements in days or weeks; avoid "big-bang" releases that stall.
What to watch next
Expect more unified analytics across human and AI performance, stronger grounding with enterprise data, and faster voice experiences. Per-minute pricing will pressure vendors that sell AI features a la carte. For most teams, the winning move is clear: augment agents first, then automate the repeatable pieces with tight monitoring.
If you're upskilling your team for AI-assisted support and QA at pace, explore curated programs built by job role: AI courses by job.
Your membership also unlocks:
