Amazon Connect Customer
Cloud contact center for voice, chat, and SMS, renamed from Amazon Connect, with AI agents that resolve routine contacts
A call center used to mean a room full of phones and a rack of hardware you bought, housed, and repaired. Connect Customer is that call center as a cloud service. You log into a browser, claim a phone number, and draw your call menu as a flowchart (press 1 for billing, press 2 for support). Agents wear headsets and sign in from anywhere. You pay for what you use instead of buying agent seats, so adding 50 people for a busy month needs no purchase order. The newest part is who answers first: an AI agent picks up, works out why the caller rang, and often finishes the job outright, speaking more than 50 languages. When it cannot, it hands the human agent the call, the transcript, and its notes. The flowchart still runs the show; the AI works inside it. You build that AI conversation on a visual canvas called the Agentic CX designer, the same draw-it-out way you built the call menu.
Connect Customer keeps the Amazon Connect architecture: contact flows drive every interaction from IVR prompt to queue to agent, routing profiles set agent queue membership, priority, and concurrency, and Lambda functions pull live account data into the flow. The agentic layer sits on top. An orchestrator AI agent handles self-service, and applications built on the Agentic CX designer canvas are invoked from a flow through the Agentic CX block, so deterministic steps like authentication stay enforced. Eight AI agent metrics, including goal success rate and faithfulness score, measure whether the AI actually resolved the contact.
Key Capabilities
- Omnichannel contacts: voice, chat, email, SMS, and tasks land in one agent workspace with queues and routing profiles
- Contact flows: visual editor for IVR menus and routing, with Lambda lookups for live customer data; reusable flow modules now work in every flow type, even nested inside other modules
- Agentic self-service: an orchestrator AI agent recognizes the caller, inherits context from web or app sessions, and resolves issues before a human joins
- Agentic voice: expressive AI speech in over 50 languages with more than 100 voice options, plus support for third-party speech models
- Agentic CX designer: no-code canvas that mixes open-ended AI conversation with deterministic steps for authentication, verification, and required disclosures; apps plug into flows through the Agentic CX block
- Global Resiliency: two-Region active routing that offers contacts to available agents across paired Regions, with unified analytics and continuous live-traffic validation
Gotchas & Constraints
Gotcha #1: instances are single-Region by default. Two-Region active routing needs Global Resiliency, which requires AWS account team approval and paired Regions. Gotcha #2: Agentic CX designer runs in nine Regions, including N. Virginia, Oregon, Frankfurt, London, Tokyo, and Sydney; elsewhere, complex conversations fall back to standard flows and bots. Constraints: pay-as-you-go billing for telephony and AI features, and Tasks and Email concurrency limits are now set per workload type, not per channel.
A regional insurer runs 120 agents on an end-of-life on-premises phone system. Open enrollment triples call volume to 24,000 calls a day each November, and most questions are routine: deductibles, plan comparisons, replacement ID cards. They migrate to Connect Customer over eight weeks. An AI agent answers first in English or Spanish, pulling the caller's plan record through Lambda before speaking. An Agentic CX designer app runs identity verification as a deterministic step, then handles routine plan questions conversationally, passing anything unusual to a person with the transcript and AI notes attached. During November the AI resolves 41% of calls outright. Human agents handle the rest with AI-generated response drafts from plan documents, cutting average handle time from 9 to 6.5 minutes. Sixty temporary agents go live in one afternoon: accounts created, routing profiles assigned, two hours of browser training.
The Result
the six-week seat provisioning cycle is gone, AI deflection saves about $180,000 a year in agent hours, and peak season absorbed a 25% volume spike with zero hardware.