AI voice agents for Amazon Connect that answer, qualify, and hand off cleanly.
We design and build production AI voice agents on Amazon Connect and AWS Bedrock. They answer calls around the clock, capture structured details, update your systems, and transfer to a person with full context when it matters.
Built on
Amazon Connect, AWS Bedrock, and Nova Sonic
In production
Multi-tenant intake agent answering 100% of calls
Human control
Defined escalation path and full call logging
Listening and speaking
One streaming connection
Intake record
Fields captured
Transcript
Logged to S3
Human fallback ready
Warm transfer with context
Built for ai voice agents
Speech-to-speech voice agents on Amazon Connect for intake, scheduling, and support calls, with tool calling, call logging, and human fallback.
Natural phone conversations
Speech-to-speech models handle interruptions and turn-taking, so callers get a conversation instead of a menu.
Actions in your systems
Validated tool calls create cases, book slots, and look up records in your CRM, scheduling, or case management tools.
Human fallback and audit trail
A designed escalation path, plus full call transcription and logging to S3 for review and compliance.
Proof, not just a pitch.
Real numbers from named engagements, not marketing estimates.
Common use cases
Tailored to your workflow, but these are typical patterns.
01
Intake and qualification
- After-hours and overflow call intake
- Structured details captured from natural speech
- Qualified leads routed into CRM or case management
02
Scheduling and status
- Appointment booking and rescheduling
- Order, case, or ticket status lookups
- Hours, location, and policy answers grounded in approved content
03
Support triage
- Intent detection and context gathering before transfer
- Routine requests resolved end to end
- Warm handoff with transcript and extracted fields attached
Typical scope
Three ways to put an AI voice agent on Amazon Connect
The right path depends on how much of the conversation you need to control. We recommend the simplest option that meets the call type, and we have built the most flexible one in production.
| Approach | How it works | Best for | Trade-off |
|---|---|---|---|
| Connect native AI features | Built-in conversational features configured inside Amazon Connect | FAQ, status, and simple self-service flows on a short timeline | Fastest to launch; least control over conversation logic and tool use |
| Amazon Lex bot with Bedrock fulfilment | Lex handles intents and slots; Lambda calls Bedrock for reasoning | Structured intents with some free-form understanding; teams already on Lex | Familiar tooling, but turn-based, with more latency and rigid intent design |
| Speech-to-speech agent (Amazon Nova Sonic) | One model hears and speaks over a bidirectional stream; Connect keeps orchestration and flows | Intake, qualification, scheduling, and any call that depends on what the caller just said | Most natural and flexible; you own state, guardrails, and monitoring |
Amazon documents Nova Sonic as a speech-to-speech model you can enable on a Conversational AI bot locale in Amazon Connect, paired with a Nova Sonic compatible voice in the flow. See the full architecture and build guide for call flow, tool design, and failure modes.
Each layer has one job, and a plan for how it fails
Most production incidents on voice agents come from a layer doing a second job it was not designed for. This is the layout we build to.
| Layer | Typical AWS service | Responsibility | Failure we design for |
|---|---|---|---|
| Telephony and routing | Amazon Connect contact flows and queues | Numbers, entry flow, queue transfer, recording | Agent unreachable: the flow falls back to a human queue, never dead air |
| Conversation | Amazon Bedrock (Nova Sonic, Claude) | Understands the caller and decides what to say and do | Slow turns, invented details, accent and noise handling |
| Tool layer | AWS Lambda and your APIs | Validated actions: look up, create, schedule, transfer | Timeouts and duplicate writes on retry |
| Audit and storage | Amazon S3 and your database | Recordings, transcripts, structured outcomes | Missing records for the one call that is later disputed |
| Human fallback | Connect queues and agent desktop | Warm transfer with transcript and extracted fields | Transfer without context, so the caller repeats everything |
Our production deployment follows this pattern. Read how it works in AI Voice Agents with Amazon Connect and AWS Bedrock.
An intake call, step by step
Walking a single call through the system is the fastest way to find gaps in a design. This is an illustrative after-hours intake call.
| Step | Caller | Agent | System effect |
|---|---|---|---|
| 1 | Dials the Connect number | Greets and discloses AI assistance and recording | Contact flow starts the stream and a call ID is created |
| 2 | Describes the issue in their own words | Listens and asks one clarifying question at a time | Transcript streams to storage |
| 3 | Gives name and callback number | Reads them back to confirm | Fields held in session state |
| 4 | Corrects a digit mid-sentence | Stops speaking, fixes the field, resumes | State updated, no data lost |
| 5 | Confirms the details | Calls the intake tool | Record written to the system of record |
| 6 | Asks something out of scope | Declines to guess and offers a person | Escalation logged with transcript |
| 7 | Ends the call | Summarises next steps | Recording, transcript, and outcome stored |
AI voice agent, touch-tone IVR, and answering service compared
Teams usually weigh an AI voice agent against the two options they already know. This is how they differ in practice.
| Dimension | Touch-tone IVR | Answering service | AI voice agent |
|---|---|---|---|
| Caller experience | Press-a-number menus | A person takes a message | Open conversation in the caller's own words |
| Availability | 24/7 | Limited by staffing and plan | 24/7 and unaffected by call spikes |
| Captures structured data | Only what menus allow | Free-text notes | Structured fields written to your systems |
| Takes action in your tools | Limited to routing | Rarely | Creates records, books slots, looks up status |
| Handles edge cases | Dead ends | Person decides | Escalates to a person with context |
| What you maintain | Menu trees | A vendor relationship | Prompts, tools, and a regression test set |
What an AI voice agent costs to run on Amazon Connect
Build effort depends on scope, so we quote it after a discovery call. Running costs on AWS are published, and these are the list prices as shown on the AWS pricing pages on 5 November 2026.
| Cost line | List price | Notes |
|---|---|---|
| Amazon Connect Customer: voice | $0.038 per voice minute | Standard telephony rates apply on top. AWS bundles its AI features into this rate. |
| Amazon Connect Customer Basic: voice | $0.018 per minute | Channel usage only; AI features are priced separately. Telephony is billed on top. |
| US local phone number (DID) | $0.030 per day | Per number, US region rate. |
| Inbound telephony on a US DID | $0.0022 per minute | Billed while the call is open, including time in a queue. |
| Speech model (Nova Sonic on Bedrock) | Billed per speech token | Scales with call length. Check the current rate on the AWS Bedrock pricing page. |
| Integration layer (Lambda, S3) | Usage-based | Usually small next to voice minutes; chatty tool designs add latency as well as cost. |
| Human escalations | Staff time | Context passed with the transfer shortens handling time. |
Worked example: 1,000 calls a month, 5 minutes each, Connect Customer, US number
| Line | Calculation | Monthly cost |
|---|---|---|
| Voice usage | 5,000 min × $0.038 | $190.00 |
| Inbound telephony | 5,000 min × $0.0022 | $11.00 |
| Phone number | 30 days × $0.030 | $0.90 |
| Platform subtotal | Before speech-model usage, other AWS services, and build or support | $201.90 |
This is an illustration built from list prices and excludes tax, other regions, toll-free numbers, and volume tiers. AWS changes pricing, so confirm on the Amazon Connect Customer pricing page before budgeting.
The metrics we agree before launch
Containment rate alone can hide calls that were resolved incorrectly. We track quality, speed, and reliability together.
| Metric | How it is measured | Why it matters |
|---|---|---|
| Task success rate | Share of calls where the structured outcome matches what a reviewer says should have happened | The real measure of quality |
| Escalation rate | Calls handed to a person versus fully handled | Drives cost; read it together with task success |
| Turn latency (p50 and p95) | Time from the caller finishing to the agent speaking | Callers notice the slow tail more than the average |
| Abandonment during handling | Callers who hang up before completion | Early signal of a confusing greeting or frustration |
| Tool error rate | Failed or retried tool calls per 100 calls | Exposes integration problems before callers do |
| Reviewer disagreement | Sampled calls scored by a person | Catches drift after prompt or model changes |
A staged rollout, with a rollback switch
Most failed pilots tried to automate the whole phone line at once. We start with one call type and widen only when the numbers hold.
| Phase | Scope | Exit criteria |
|---|---|---|
| 1. Discovery | One call type; listen to real calls; define success and escalation rules | Written success definition and a baseline from the current process |
| 2. Prototype | Working agent on a test number with mocked tools | Scripted scenarios pass; voice and wording approved |
| 3. Integrated pilot | Real tools on after-hours or overflow traffic, every call reviewed | Task success and escalation behaviour meet the agreed bar |
| 4. Gradual rollout | Rising share of live calls with sampled review | Stable metrics at each step; rollback tested |
| 5. Operate | Monitoring, regression suite, scheduled prompt and knowledge reviews | Named owner and a documented change process |
Consent, recording, and data handling
Requirements vary by jurisdiction and industry, so confirm them with counsel. These are the decisions we settle with you before launch.
| Topic | Typical approach | Decision for your team |
|---|---|---|
| Disclosure and consent | Greeting states the call is AI-assisted and recorded where required | Exact wording and any jurisdiction-specific consent |
| Retention | Recordings and transcripts stored in S3 with a defined lifetime | How long to keep them and who may read them |
| Sensitive data | Redaction rules for payment details and health information | What the agent may collect at all |
| Data boundary | Data stays inside your AWS account boundary | Region and account where it runs |
| Access control | Least-privilege access to transcripts and recordings | Named roles with access |
| Human access | A person is reachable at any point in the call | Where transfers land and when they are staffed |
AI Voice Agents questions
What teams usually ask before starting this kind of engagement.
What is an AI voice agent for Amazon Connect?
It is software that answers a phone call routed through Amazon Connect, understands what the caller wants, takes actions in your business systems, and either resolves the call or transfers it to a person with context. Amazon Connect keeps telephony, routing, queues, and recording, while the voice agent owns the conversation.
How does an AI voice agent integrate with Amazon Connect?
The contact flow hands the call to the agent. With Amazon Nova Sonic, AWS documents enabling speech-to-speech on a Conversational AI bot locale and selecting a compatible voice in the flow's Set voice block. Custom designs can instead stream call audio to a model you control. In both cases the flow can transfer to a human queue at any point.
Do you build with Amazon Nova Sonic?
Yes. Our production intake agent for Bracken & Cole LLP runs on Amazon Connect, AWS Bedrock, Nova Sonic, and Lambda. Nova Sonic handles speech understanding and generation in one model over a bidirectional stream, which avoids chaining separate speech-to-text, language model, and text-to-speech steps.
Can I use a custom voice with Amazon Connect?
It depends on the architecture. For Nova Sonic speech-to-speech, AWS documents a set of compatible voices in the Set voice block, such as Matthew (en-US), Amy (en-GB), Olivia (en-AU), and Lupe (es-US). If you need a custom brand voice or strict control over wording, a cascaded pipeline with your chosen text-to-speech is the usual route. Voice availability changes, so we confirm it against current AWS documentation during scoping.
How much does it cost to run an AI voice agent on Amazon Connect?
AWS lists Amazon Connect Customer voice at $0.038 per minute with standard telephony on top, and Customer Basic at $0.018 per minute with AI features priced separately. A US number adds $0.030 per day and $0.0022 per inbound minute. Speech-model usage and integrations are billed separately. On list prices, 1,000 five-minute calls a month come to about $202 in Connect charges before model usage. Build cost depends on scope and is quoted after discovery.
How long does it take to launch an AI voice agent?
It depends on the call type, the number of integrations, and how much review the pilot needs. We scope one call type first, run it on a test number, then pilot it on after-hours or overflow traffic before widening. We confirm the timeline in the proposal after discovery rather than quoting a generic number.
What happens when the agent cannot handle a call?
Every deployment has a defined human fallback for ambiguous situations, explicit requests for a person, and topics outside scope. The transfer carries the transcript and extracted fields so the caller does not repeat themselves, and the escalation is logged for review.
Does the voice agent replace our contact center agents?
No. It takes the calls that are routine, after-hours, or overflow, and routes the rest to people with context. The goal is to stop calls going to voicemail or long holds, and to move human effort to the conversations that need it.
Can one voice agent system serve multiple clients or brands?
Yes, if it is designed for it from the start. Our intake agent is multi-tenant, with per-firm configuration and business logic on shared infrastructure. Retrofitting multi-tenancy onto a single-client build usually means re-architecting routing and state, so we decide it at design time.
Which languages does an Amazon Connect voice agent support?
Language support depends on the model, the voice, and the AWS region. AWS documents an initial Nova Sonic voice set covering US, UK, and Australian English and US Spanish, and availability expands over time. We check supported languages and regions against current AWS documentation before committing to an approach.
Learn more about ai voice agents
Practical engineering guides from real builds, written by the people who ship them.
Voice AI
AI Voice Agent for Amazon Connect: Architecture and Build Guide
A practical guide to building an AI voice agent on Amazon Connect — the build options, the call flow, what drives cost, and the mistakes that sink pilots.
13 min read
Voice AI
AI Voice Agents with Amazon Connect and AWS Bedrock
What it actually takes to put a speech-to-speech agent into a real contact center — bidirectional streaming, tool calling, and human fallback, not just a demo call.
10 min read
Agentic AI
How to Build Human-in-the-Loop AI Agents That Teams Actually Trust
The pattern that made the difference on our production agent deployments: resolve ambiguity toward review, not toward silent action.
9 min read
AI Evaluation
Evaluating AI Agents in Production: Accuracy, Cost and Reliability
An agent that's 95% reliable per step succeeds end-to-end only about a third of the time across a 20-step chain — here's how to actually measure whether an agent is ready for production.
11 min read
Need a broader engagement?
We often combine product build, agents, and knowledge systems in one roadmap.