What Is Amazon Lex? Features, Pricing & Alternatives

Amazon Lex is one of the oldest AI services AWS sells. It launched in 2016, and for a long stretch it was the default way to put natural language in front of an AWS application. That history is worth knowing before you evaluate it, because a lot of what gets written about Lex describes the 2017 version of the product, and a lot of what AWS has built since then sits above it rather than inside it.

This guide covers what Lex actually does today, the generative AI features it picked up after 2023, what it costs, the V1 shutdown you need to check for, and where a dedicated agent platform fits instead.

What Is Amazon Lex?

Amazon Lex is a fully managed AWS service for building conversational interfaces using text and voice. You define what your users might want, Lex works out which of those things a given utterance maps to, pulls the details it needs, and hands control to your code to do the work.

It shares its speech recognition and language understanding lineage with Alexa, which is where AWS's "same technology as Alexa" line comes from. That's a lineage claim, not a feature claim: Lex is a developer service for building your own bots, and it doesn't give you Alexa's assistant behaviour.

The short version: Lex is the conversational AI layer inside AWS. It understands the request. What happens next is yours to build.

How Amazon Lex Works

Lex processes each turn through four stages:

1. Input processing. The user's speech or text arrives at Lex, which transcribes audio if needed.
2. Intent recognition. Lex classifies the utterance against the intents you defined. "I want to return my order" and "can I send this back" should both land on a returns intent.
3. Slot filling. Lex extracts the values the intent needs, prompting for anything missing. A returns intent might need an order number and a reason.
4. Fulfillment. Once the slots are full, Lex invokes your logic, usually a Lambda function that calls your systems, then returns a response by voice or text.

That intent-and-slot model is the foundation, and it's still the foundation in 2026. The generative features layered on top make it more forgiving, but underneath you're defining intents and slots.

The Generative AI Features Lex Added After 2023

This is the part most Lex write-ups miss, and it's the difference between evaluating a 2017 product and the current one.

• QnAIntent. A built-in intent that answers questions from an Amazon Bedrock knowledge base using retrieval-augmented generation. You point it at your documents and it handles FAQs without you authoring intents, utterances, and prompts for each one. This is the single biggest change to how you build a Lex bot.
• Assisted NLU. An LLM improves intent classification and slot resolution while staying inside the intents and slots you configured. You get better matching without giving up the guardrails.
• Descriptive bot building. Describe the bot you want in plain language and get a first draft to edit, rather than starting from an empty console.
• Visual Conversation Builder. A drag-and-drop surface for designing flows, instead of configuring everything as separate console objects.
• Automated Chatbot Designer. Feed it existing call or chat transcripts and it proposes intents and slots based on what your customers actually say. This is the $0.50-per-minute line in the pricing table below.

Worth noting how QnAIntent changes the cost picture: the retrieval and generation happen in Bedrock, so those calls are billed as Bedrock usage on top of your Lex request charges.

Amazon Lex V1 vs V2: Check Which One You Are On

Before anything else, find out which version your bots are built on.

AWS ended support for Amazon Lex V1 on September 15, 2025. New V1 resources stopped being possible on March 31, 2025, and after the September date the V1 console and V1 resources are no longer accessible. V2 is the only current version.

This matters more than a version number usually would, because V2 isn't a drop-in upgrade. It uses a different API, and it restructures bots around locales and a different versioning model. Moving a non-trivial V1 bot is closer to a rebuild than a migration. If you inherited a Lex deployment and nobody has looked at it since 2024, that's the first thing to check, and it's a reasonable moment to ask whether you'd rebuild on Lex at all.

Amazon Lex Pricing

Pricing for Amazon Lex is based on a pay-per-use model. There's no upfront cost or minimum involved. For the most accurate and up-to-date pricing, refer to the Amazon Lex pricing page.

Two things the table doesn't show. First, there's no permanent free tier: since July 15, 2025 new AWS accounts get up to $200 in Free Tier credits usable across eligible services including Lex, available for six months after account creation, with credits expiring at twelve. Second, anything that routes through QnAIntent adds Bedrock charges for retrieval and generation, so a knowledge-heavy bot costs more per turn than the per-request numbers suggest.

For most support workloads the request pricing is not the expensive part. The engineering time around it is.

Building a Bot on Amazon Lex

Everything happens in the AWS Management Console or through the V2 API. The path is: create a bot, choose text or voice, define your intents with sample utterances, define the slots each intent needs, write the prompts, wire a Lambda function for anything that touches your systems, test in the console, then deploy through an alias.

AWS's own getting-started documentation walks the console click-by-click and stays current, so there's no value in duplicating it here. What's useful to know before you start is the shape of the work: the console gets you a bot that understands requests, and your engineering team supplies the interface it lives in, the state management around it, the integrations behind it, and whatever monitoring you decide you need. None of that arrives with the service.

What Amazon Lex Is Genuinely Good At

There's a clear case for Lex, and it's narrower than the marketing suggests.

• You already run Amazon Connect. This is the strongest reason by a distance. Lex is the native conversational layer for Connect, the integration is first-party, and no third-party platform can match how cleanly it drops into an existing Connect contact center. If your telephony is Connect, Lex is the default and you need a specific reason to look elsewhere.
• You want to stay inside AWS. IAM for permissions, Lambda for logic, CloudWatch for logs, one bill, one security review. For an enterprise with an AWS-only posture, that procurement path is worth real money.
• The pricing has no floor. Per-request billing with no platform fee means a low-traffic bot costs very little, which makes internal tools and pilots easy to justify.
• Multi-region and multi-language. Lex runs in most AWS regions and supports a solid set of locales, which matters for global deployments with data-residency requirements.

Where Amazon Lex Falls Down in 2026

• It's AWS-only, models included. The generative features run on Bedrock, so your model choice is whatever Bedrock offers. That's a large catalog, but it's one vendor's catalog, and the argument for keeping your options open across large language models applies with more force every year.
• The intent-and-slot model is still underneath. Assisted NLU and QnAIntent soften it, but you're maintaining an intent taxonomy. Newer agent platforms let you describe goals and let the agent reason about how to reach them, which is a genuinely different maintenance burden.
• You build the surface yourself. Lex has no website chat widget, no messaging channels, no mobile SDK you can drop in. Outside Connect, the deployment surface is your engineering team's problem.
• No agent operations story. There's no staging and production separation for bots, no built-in evaluation of conversation quality, and CloudWatch logs are not conversation-level observability. Measuring what your agents actually resolve is work you do somewhere else.
• V1 migration debt. If you're on V1, you're on an unsupported product and the way forward is a rebuild.

Amazon Lex Alternatives

For the cloud-native options, our Dialogflow, Azure AI, Vertex AI, and IBM Watson breakdowns go deeper, and the Rasa and Stack AI reviews cover the self-hosted and workflow ends. If the voice channel is your main concern, the Vapi comparison is the closer fit.

Amazon Lex vs Voiceflow

These solve different problems, and the honest framing is that they're layers rather than straight competitors. Lex gives you natural language understanding inside AWS. Voiceflow is where you design, ship, and watch the agent that uses it.

• Model-agnostic by design. Voiceflow runs across OpenAI, Anthropic, and Google, and you can bring your own model. You switch as cost and quality move instead of inheriting one cloud's catalog.
• Real agent primitives. Workflows handle the deterministic sequences that must run identically every time, like a refund or an identity check. Playbooks handle open-ended reasoning against a goal. They compose, so a workflow can call a playbook mid-sequence and a playbook can hand off to a workflow when the conversation enters regulated territory. You design both on a visual canvas.
• Grounded answers without assembling RAG. A built-in knowledge base keeps responses on your content. It's the QnAIntent idea without wiring a separate service and paying for it separately.
• Every channel from one build. The same agent ships to website chat, messaging, and voice, including phone, with a clean handoff to a human when it hits its limit.
• Built for running in production. Observability gives you turn-by-turn traces of what the agent did and why, Evaluations score conversations against criteria you define, and Environments let you test in staging before anything reaches customers. For sensitive data, Voiceflow runs with SOC 2 Type 2 and PII masking.

Teams like Turo, StubHub International, Sanlam Studios, and Trilogy build their agents this way. A common pattern for AWS-committed teams is to keep Connect for telephony and Bedrock for models, then build the agent itself in Voiceflow on top, which gets you the AWS infrastructure without hand-assembling the agent layer.

If you're weighing Lex against a platform built for agents, book a demo and we'll walk through your use case.