AWS Contact Center

Customize AI in Amazon Connect Customer: Agents, Prompts, and Guardrails

Introduction

Artificial intelligence (AI) Agents, Prompts, and Guardrails give contact centers extensive control over how generative AI engages customers and supports human agents. Built into the next generation of Amazon Connect Customer, these three features let you select the right large language model (LLM) for each task and configure how it behaves within your contact center. In this post, this post walks through each component and shows you how to customize them.

The power of AI in Amazon Connect Customer

Amazon Connect Customer now delivers first-party AI across supported channels, using foundation models from Amazon Bedrock. This unified approach brings self-service, agent assistance, and analytics together under a single AI framework.

At the heart of this AI-powered experience are three interconnected components:

  • AI Agents: Autonomous assistants that handle customer interactions and support human agents
  • AI Prompts: Customizable instructions that guide how AI responds to specific situations
  • AI Guardrails: Safety controls that ensure AI behaves appropriately and within defined boundaries

Understanding AI agents

AI Agents in Amazon Connect Customer are generative AI-powered assistants designed to handle customer interactions across voice and digital channels. These agents can:

  • Orchestrate complex workflows: Coordinate multiple tools and data sources to complete tasks such as checking order status, processing refunds, or updating records
  • Provide customer self-service: Answer frequently asked questions, process routine requests, and guide customers through common workflows without human intervention
  • Assist human agents: Deliver real-time recommendations, suggested responses, and relevant knowledge base articles to help agents resolve issues faster

AI Agents use LLMs in Amazon Bedrock to understand customer intents. These models draw on relevant sources of information to deliver accurate responses and actions in real time.

Customizing AI behavior with AI prompts

While AI Agents provide the framework, AI Prompts give you granular control over how the AI behaves. An AI prompt is essentially a task or instruction for the large language model, defining how it should respond in specific situations.

Types of AI prompts

Amazon Connect Customer provides templates for various prompt types:

Prompt Type Purpose
Orchestration Coordinates different use cases based on customer needs
Answer Generation Creates solutions using knowledge base content
Intent Labeling Identifies customer intents for agent display
Query Reformulation Constructs relevant searches for knowledge bases
Self-Service Pre-Processing Evaluates conversations and selects appropriate tools
Self-Service Answer Generation Generates self-service responses from knowledge bases
Email Response/Overview Handles email-specific interactions
Note Taking Creates structured notes from live conversations
Case Summarization Summarizes customer cases

Selecting your AI model for each prompt type

A key capability in AI Prompts is the ability to select which large language model powers each prompt. This capability, available directly in the Amazon Connect Customer web UI, allows contact center administrators to optimize AI performance for their specific use cases.

Prerequisites

Before configuring custom AI prompts or agents, ensure you have the following:

  • An active Amazon Connect Customer instance with AI features enabled
  • Administrator-level access, or a security profile with AI agent designer permissions (AI agents, AI prompts, and AI guardrails Create/Edit)
  • Basic familiarity with Amazon Connect Customer flows to integrate AI agents into your contact routing

How to select your AI model

Follow these steps to choose and configure your preferred LLM in the Amazon Connect Customer console:

  • Navigate to AI Prompts: Log in to your Amazon Connect Customer admin website and go to AI agent designer → AI prompts

Screenshot of the Amazon Connect Customer admin console showing the AI agent designer navigation menu with AI prompts highlighted

  • Create or Edit a Prompt: Choose Create AI Prompt and select your prompt type, or edit an existing prompt
  • Access the AI Prompt Builder: After creating the prompt, you’ll be taken to the AI Prompt builder page
  • Select Your Model: In the Models section, you’ll see a dropdown menu displaying the system default model for your AWS Region. Choose from the dropdown to view all available models and select your preferred LLM

Screenshot of the AI Prompt builder showing the Models section dropdown with available LLM options

The available models vary by region and include options such as:

  • Amazon Nova Pro, Nova Lite, and Nova Micro models
  • Anthropic Claude Haiku and Claude Sonnet models

Model selection recommendations

Choosing the right model for each prompt type can significantly impact performance, cost, and response quality. While each prompt type suggests a default model for the reasons outlined below, Amazon Connect Customer gives you the flexibility to choose a different model at any time.

For high-complexity tasks (Orchestration, Case Summarization)

Recommended: Anthropic Claude Sonnet

These tasks require sophisticated reasoning and the ability to synthesize information from multiple sources. Claude Sonnet models excel at:

  • Complex decision-making
  • Nuanced understanding of context
  • Generating comprehensive summaries

For speed-critical tasks (Query Reformulation, Intent Labeling)

Recommended: Amazon Nova Lite or Claude Haiku

When latency matters most, lighter models deliver faster responses with minimal accuracy tradeoff for straightforward tasks:

  • Lower latency for real-time interactions
  • Cost-effective for high-volume operations
  • Sufficient capability for focused, single-purpose tasks

For self-service interactions (Self-Service Pre-Processing, Answer Generation)

Recommended: Amazon Nova Pro

Nova Pro offers an excellent balance of capability and performance for customer-facing self-service:

  • Strong reasoning for tool selection
  • Reliable answer generation from knowledge bases
  • Optimized for conversational interactions

For agent assistance (Answer Generation, Note Taking)

Recommended: Claude Sonnet or Claude Haiku

Agent-facing features benefit from models that can quickly process conversation context:

  • Fast generation of suggested responses
  • Accurate real-time note taking
  • Contextual understanding of ongoing conversations

Cross-region inference

Many models support cross-region inference, which can improve performance and availability. When you see “(Cross-Region)” next to a model name, it means the model can route requests across multiple AWS regions to optimize response times and handle capacity constraints.

Implementing AI guardrails

AI Guardrails help verify your AI behaves responsibly and within defined boundaries. Amazon Connect Customer has six guardrail policy types including a customizable blocked message.

  • Content Filters: Blocks input prompts or model responses containing harmful content
  • Denied Topics: Define up to 30 custom topics you want blocked. You provide a name, definition, and examples. If detected in user queries or model responses, they’re blocked
  • Word Filters: Block specific words, phrases, or profanity via exact match
  • Sensitive Information Filters: Block or mask personally identifiable information (PII) and sensitive data in both user inputs and model responses
  • Contextual Grounding Check: Detects and filters responses not grounded in provided context (hallucinations)
  • Blocked Messaging: Customize the default message shown when a guardrail triggers

Guardrails work alongside your AI Prompts and Agents to create a comprehensive governance framework for AI interactions.

Putting it all together: A practical example

The following example walks through setting up a customized AI experience for a retail contact center:

Step 1: Customize AI prompts

Create custom prompts for each interaction type:

Answer Generation Prompt

  • Model: Anthropic Claude Sonnet
  • Purpose: Generate accurate responses from your product knowledge base

Screenshot of the AI Prompt builder configured for Answer Generation with Anthropic Claude Sonnet selected

Step 2: Configure guardrails

Set up guardrails to:

  • Prevent discussion of competitor products
  • Ensure pricing information comes only from approved sources

Screenshot of the AI Guardrails configuration page showing content filter and denied topics settings.

Step 3: Create your AI agent

Configure an AI Agent to answer customers’ frequently asked questions, select the prompt created in step 1, attach the guardrails created in step 2 and configure tools as needed.

Screenshot of the AI Agent configuration page with prompt and guardrail attachments.

Step 4: Deploy and monitor

Publish your AI Agent and monitor performance through Amazon Connect Customer analytics.

Best practices for AI prompt optimization

To get the most out of your AI Prompts:

  1. Position static content before variables: This enables prompt caching for improved latency
  2. Use prompt prefixes with at least 1,000 tokens (approximately 750 words): Longer static sections optimize caching performance
  3. Test different models: Experiment with various LLMs to find the best fit for each use case
  4. Monitor and iterate: Use conversational analytics to identify opportunities for prompt improvement
  5. Use cross-region inference: Turn on cross-region models for better availability and performance

Getting started

Follow these steps to start customizing your Amazon Connect Customer AI experience:

  1. Access the Console: Log in to your Amazon Connect Customer admin website
  2. Navigate to AI Agent Designer: Find AI Prompts, AI Agents, and AI Guardrails in the navigation menu
  3. Start with Templates: Use the provided prompt templates as a starting point
  4. Experiment with Models: Try different LLMs to optimize for your specific needs
  5. Deploy Incrementally: Test changes in a controlled environment before full deployment

Cleanup and cost considerations

The AI features configured in this post incur costs based on usage:

  • AI Prompt model changes take effect immediately and incur Amazon Bedrock inference charges for each invocation
  • To revert to default models, edit each AI Prompt and select the system default option from the Models dropdown
  • AI Guardrail evaluations may incur additional charges depending on your configuration
  • Review Amazon Bedrock pricing for current model invocation costs in your Region

Conclusion

The ability to customize AI models in Amazon Connect Customer provides additional flexibility for contact center personalization. Selecting the right LLM for each prompt type lets you optimize for speed, accuracy, cost, and customer experience. You can do this directly from the admin console.

With AI Agents, Prompts, and Guardrails, you can create differentiated customer experiences. Start with self-service scenarios like the retail FAQ example above, then explore agent-assistance use cases as your needs grow.

Learn More

Ready to transform your customer service experience with Amazon Connect? Contact us