Amazon Connect

What it is

Amazon Connect is AWS’s cloud contact center: voice, chat, SMS, WhatsApp, email, routing, workforce management, conversation analytics, and AI agents, all billed by the minute or message with no seats. It began as the system behind Amazon’s own retail customer service and launched as an AWS service in 2017. AWS disclosed a $1B annualized revenue run rate on its Q3 2025 earnings call, and named customers include Capital One, State Farm, Hilton, Air Canada, and United Airlines.

On 28 April 2026 AWS split Connect into a family of four products. The contact center is now Amazon Connect Customer; the other three are Connect Talent (high-volume hiring, covered in our Amazon Connect Talent entry), Connect Decisions (supply-chain planning), and Connect Health (clinical documentation and patient coordination). This page covers Connect Customer — the part a support or CX-ops team buys.

What changed in 2026

• NLX acquisition. AWS bought NLX, a no-code conversational-AI builder, on 23 April 2026. The pitch is that business teams, not only developers, can build AI self-service journeys; AWS cites United Airlines going live in three months instead of twelve. That is a vendor claim. In June AWS previewed Agentic CX Designer, the visual builder that follows from it, alongside Live Sync, which ties a voice call to on-screen steps.
• Agentic voice got broader. A 20 July 2026 update took agentic voice to 50+ languages, including Spanish, French, Portuguese, and Japanese, with more than 100 new voices and controls for speed, volume, emotion, and turn-taking.
• MCP as a tool layer, not a server. Since November 2025, Connect’s AI agents can call tools over the Model Context Protocol: prebuilt Connect actions, your own flow modules exposed as tools, and outside APIs through an Amazon Bedrock AgentCore Gateway. Connect consumes MCP; it does not ship an MCP server you point Claude at, which is why we mark mcp_available false.
• Voice ID is gone. Connect’s caller voice-biometrics feature reached end of support on 20 May 2026. AWS points customers to Pindrop on AWS Marketplace or to SMS one-time passcodes.

Pricing reality

AWS publishes the default Connect Customer rates: $0.038 per voice minute, $0.010 per chat message, $0.014 per SMS or third-party message, $0.080 per email, with no seats, minimums, or contracts. That rate includes AI agents and self-service, conversation analytics and summaries, real-time agent assist, forecasting and scheduling, and evaluations. Telephony is billed on top. A September 2026 third-party rate card (Macha) puts US numbers at $0.03/day for a DID and $0.06/day toll-free, inbound at $0.0022/min on a DID and $0.012/min toll-free, and outbound at $0.0048/min, and lists Cases ($0.12 each), Tasks ($0.07), and Bedrock model costs as separate meters.

The same rate card lists a cheaper Customer Basic mode — $0.018 per voice minute with each AI feature priced separately (analytics $0.015/min, agent assist $0.008/min, forecasting and scheduling $27/agent/month, evaluations $12/agent/month) and no agentic AI agents. AWS’s own page links to Customer Basic without listing its rates, so confirm them in the console before modelling.

Here is the arithmetic the “pay-per-minute is cheaper” pitch skips. A seat-priced platform charges the same whether an agent handles 20 hours of calls a month or 150. Connect charges for every minute. Five9’s Core seat lists at $159/month, which equals about 4,200 Connect minutes — roughly 70 handled hours. A 50-agent voice team handling 6 hours of calls a day over 22 days runs about 396,000 minutes a month: roughly $15,000 in Connect charges plus $870 of DID inbound telephony, or about $190K a year. That is double Five9’s $95K Core floor on seats alone, though Connect’s figure already includes WFM, analytics, and AI that Five9 sells in higher tiers. Connect wins clearly on seasonal, low-utilization, or AI-first traffic. An AI-contained call costs about $0.04/min including DID telephony, against $0.09–$0.15/min all-in on a typical Vapi or Retell build.

The bill that does not appear on the pricing page is engineering. Connect is configured in the AWS console and extended with Lambda, Lex, and Bedrock. Budget either an AWS-certified partner or a named in-house engineer who owns contact flows, IAM, and the CRM integration after go-live.

Best for

CX-ops and support-engineering teams that already run on AWS and have at least one engineer who can own contact flows — especially centers with seasonal peaks, many part-time or low-utilization agents, or a plan to move a large share of volume to AI self-service, where per-minute billing beats paying for idle seats.

Skip it if you have no AWS skills in-house and want a contact center a supervisor can configure on day one; a packaged CCaaS will get you live faster.

Versus the alternatives

Genesys Cloud and Five9 are the market-share leaders. Pick them when your agents are busy most of the shift, so a flat seat price beats per-minute billing, or when supervisors need to change routing and WFM without engineering tickets. Pick Connect when utilization is uneven or AI will carry most volume, because idle capacity costs nothing.

NiCE CXone and Talkdesk are the pick when you want vertical packages (banking, healthcare, retail) ready-made instead of built on AWS primitives.

Parloa and PolyAI are the fastest-growing AI-native voice-agent vendors. Pick one of them when the voice agent’s quality is the whole project and you will run it in front of an existing contact center; pick Connect’s own AI agents when you want AI and human queues, analytics, and billing in one AWS account.

For under 20 agents with no engineers, Dialpad gets you live faster with seat pricing and built-in AI.

If none fit, put one queue on Connect for a month alongside your current platform. Export the minutes, telephony, and Bedrock charges from AWS Cost Explorer and compare cost per resolved contact, not per seat.

Watch-outs

• Per-minute billing punishes busy agents. A center at high occupancy can pay more on Connect than on seats. Guard: before signing, pull 90 days of handled minutes per agent from your current platform and multiply by $0.038 plus telephony. If the result beats your seat bill, keep going; if not, the case has to rest on AI deflection, so model that separately.
• It is an AWS project, not a SaaS login. Changing a contact flow, a Lambda function, or an IAM role is developer work, and the console assumes you know AWS. Guard: name the engineer or partner who owns flows, Lambda functions, and IAM roles after go-live, and build flows in version-controlled templates (CloudFormation or the Connect APIs), not only by hand in the console.
• Meters multiply. Connect usage, telephony, Cases, Tasks, Customer Profiles, Lambda, and Bedrock each bill separately. Guard: put Connect in its own AWS account or cost-allocation tags, and set an AWS Budgets alert at 80% of the expected monthly spend.
• Features retire on AWS’s timetable. Voice ID ended in May 2026 with a one-year notice, and Zendesk is ending support for the TTEC-built Amazon Connect integration from April 2027. Guard: keep caller authentication and CRM integration behind interfaces you own, and track the AWS “Services in Sunset” page and the Connect release notes quarterly.