Amazon Lex is the AWS service for building conversational interfaces using voice and text. It uses deep learning for automatic speech recognition (ASR) and natural language understanding (NLU) to recognize intent and extract slot values from user input. For enterprise AI projects, Lex serves as the front-end interaction layer for chatbots, IVR systems, and automated customer service workflows.

Official documentation: https://docs.aws.amazon.com/lex/
Pricing: https://aws.amazon.com/lex/pricing/
Service quotas: https://docs.aws.amazon.com/lex/latest/dg/gl-limits.html

Core Concepts

Bot - The top-level resource. A bot contains one or more locales (language configurations), each with its own set of intents and slot types. Bots can be versioned and aliased, allowing you to develop new versions while production traffic uses a stable alias.

Intent - A goal the user wants to accomplish, such as “BookFlight” or “CheckOrderStatus.” Each intent has sample utterances (example phrases that trigger it), slots (parameters to collect), and a fulfillment action (typically a Lambda function).

Slot - A parameter within an intent. For a BookFlight intent, slots might include departure city, destination city, and travel date. Lex prompts the user for missing slots automatically. Each slot has a type (built-in types like AMAZON.Date or custom types you define).

Fulfillment - The action taken once all required slots are filled. Lambda fulfillment is the most common pattern: Lex invokes a Lambda function with the collected slot values, and the function executes business logic and returns a response.

Lex V2 vs V1

Lex V2 (current) introduced significant improvements over V1: multi-language support within a single bot, streaming conversation APIs, improved intent classification, and a redesigned console. All new projects should use V2. V1 bots can be migrated using the built-in migration tool, though custom slot types and Lambda fulfillment code may need adjustment.

Integration with Amazon Connect

The most common enterprise deployment pattern pairs Lex with Amazon Connect for contact center automation. Connect handles telephony (receiving calls, managing queues, routing to agents), while Lex handles the conversational logic. A typical flow: the customer calls, Connect plays a greeting, Lex identifies the intent (“I want to check my order status”), collects the order number, calls a Lambda function to look up the order, and returns the status. If Lex cannot resolve the issue, it hands off to a human agent with full conversation context.

This pattern reduces average handle time by resolving simple queries without agent involvement. Organizations typically see 20-40% deflection rates on well-scoped use cases like order status, account balance, and appointment scheduling.

Adding Generative AI with Bedrock

Lex V2 integrates with Amazon Bedrock to handle queries that fall outside defined intents. Instead of returning a fallback “I didn’t understand that” message, Lex can route unmatched utterances to a Bedrock foundation model that generates responses from a knowledge base. This hybrid approach provides deterministic handling for structured workflows (intent-based) and flexible handling for open-ended questions (generative).

Configure this through the QNAINTENT built-in intent, which connects to a Bedrock knowledge base. The model generates answers grounded in your documents, with source citations.

Conversation Design Best Practices

Start with a narrow scope. A bot that handles three intents well is more valuable than one that handles twenty poorly. Map the most common customer queries from support tickets or call logs, and build intents for those first.

Design for failure. Users will say unexpected things. Configure the FallbackIntent with a helpful response that either clarifies what the bot can do or offers to connect to a human agent. Never let the user hit a dead end.

Use confirmation prompts for high-stakes actions. If the bot is going to cancel an order or transfer funds, require explicit confirmation before fulfillment.

Test with real utterance data. Export utterances from production traffic and use them to identify missed intents and improve slot recognition accuracy.

Pricing

Lex charges per request: speech requests (voice) cost more than text requests. There is a free tier of 10,000 text and 5,000 speech requests per month for the first year. For high-volume contact center deployments, costs scale linearly with call volume, so model the expected request count carefully during planning.