Skip to main content
XenonStack

Agentic AI Systems

XenonStack Blog Banking Chatbots on Amazon Lex with Agentic AI

and | Published | Updated

Banking Chatbots on Amazon Lex with Agentic AI
12:38

Key takeaway

  • Build banking chatbots on Amazon Lex with agentic AI: architecture, Bedrock Agents hand-off, use-case fit, CFPB risks, controls and an adoption sequence.

A banking chatbot becomes agentic when it stops only answering and starts completing governed tasks: it understands the request, plans the steps, calls approved banking APIs and hands anything risky to a person. Amazon Lex supplies the conversation layer (speech and text understanding, intents and slots), and an agent layer such as Amazon Bedrock Agents supplies the reasoning and action. The bank supplies the controls that decide what the agent may do.

This guide is for CTOs, CIOs, heads of digital banking, enterprise architects and risk leaders who are deciding how far to take a Lex-based assistant. It sets out the architecture, a decision table for use cases, the security and governance controls and an adoption sequence. For the wider picture, see our guides to agentic process automation in finance and banking and AI in banking.

What changes when a banking chatbot becomes agentic?

A rule-based bot matches a phrase to a scripted answer. An agentic assistant treats the phrase as a goal, breaks it into steps, gathers missing information through conversation and acts on bank systems through defined actions. According to AWS, Amazon Bedrock Agents extends foundation models to understand requests, break tasks into smaller steps, collect additional information from users and take actions by calling company APIs defined in action groups.

That shift moves the design question from "what can the bot say?" to "what is the bot allowed to do, and who can see why it did it?". The rest of this guide follows that question. Our overview of agentic workflows covers the underlying pattern.

What does Amazon Lex provide, and where does the agent layer sit?

According to the Amazon Lex V2 developer guide, Lex is an AWS service for building conversational interfaces for voice and text, with automatic speech recognition (ASR) and natural language understanding (NLU). It takes speech or text, identifies the intent and fulfils it by invoking a business function. Bots can be published to mobile apps, web apps and chat services, and Lex works natively with AWS Lambda and Amazon CloudWatch. Billing is pay-as-you-go by text or speech request, with a free tier.

LayerWhat AWS documentsRole in a banking assistant
Amazon Lex V2 botASR and NLU, intents, slots, channels, Lambda integrationUnderstands the customer and manages the conversation
AWS LambdaOne function per bot alias per language, invoked at the conversation stages you chooseValidates input and calls bank services
Generative AI features in LexBuilt on Amazon Bedrock: descriptive bot builder, utterance generation, assisted slot resolution, AMAZON.QnAIntent, AMAZON.BedrockAgentIntentImproves understanding and answers knowledge questions
Amazon Bedrock AgentsOrchestration, action groups, optional knowledge bases, tracingPlans multi-step tasks and calls approved actions

AWS states that generative features may give inaccurate or inappropriate responses and may incur additional charges. Treat both points as design inputs, not footnotes.

How does the architecture work end to end?

The documented hand-off is the clearest way to add an agent to Lex. The AMAZON.BedrockAgentIntent routes a request to an Amazon Bedrock Agent or knowledge base. Once active, the conversation stays in that intent and user messages go to the agent until the agent marks the conversation finished, at which point Lex resumes control. AWS lists three requirements: generative AI enabled for the bot locale, the agent ID and agent alias ID, and the agent's User Input setting set to enabled so it can ask follow-up questions.

Two-column diagram pairing conversation steps (request, agent hand-off, action call, response and record) with controls (identity, approved content, limits and approvals, encrypted logs).
  1. Channel and identity: the customer reaches the Lex bot through the app, web or voice channel and is authenticated by the bank's own identity layer.
  2. Intent and slots: Lex classifies the request and collects required details.
  3. Agent hand-off: open-ended or multi-step requests go to the Bedrock Agent through the agent intent. Simple, high-volume requests stay in standard Lex intents.
  4. Actions: the agent calls only the actions defined in its action groups, which front the bank's APIs.
  5. Policy checks: a control layer applies limits and approvals before any action that moves money or changes customer data.
  6. Logging: conversation logs and agent traces record what happened.

Two limits are worth planning for. AWS notes that AMAZON.BedrockAgentIntent, AMAZON.QnAIntent (without sample utterances) and AMAZON.KendraSearchIntent cannot be used together in the same bot locale. AWS also states that Amazon Bedrock Agents Classic is no longer available for new sign-ups and is in maintenance mode, and points to the Strands Agents SDK with Amazon Bedrock AgentCore for new code-defined agents. Confirm the current option with AWS before you commit a design. See also our customer engagement platform use case for how we approach omnichannel engagement.

Which banking use cases fit, and which need a person?

Use caseWhen it fitsWhere a person stays in the loop
Balance, transaction and status enquiriesRead-only, well-defined intentsFailed authentication or customer dispute
Loan and credit enquiriesGuiding eligibility questions and document checklistsAny credit decision
Card and fraud alertsGuided steps after a fraud system raises a flagConfirming or reversing transactions
Onboarding supportExplaining steps and collecting documentsIdentity verification outcomes
Product questionsKnowledge-base answers from approved contentAdvice that needs a qualified adviser
Fund transfers and profile changesOnly with strong authentication and limitsAbove-limit or unusual requests

Our banking industry page describes AI agents that operate within limits set by the risk team and send high-risk steps to human approvers, which is the pattern the right-hand column reflects. For analytics behind these journeys, see our banking and finance analytics platform.

What security and governance controls does a banking agent need?

The US Consumer Financial Protection Bureau (CFPB) published "Chatbots in consumer finance" on 6 June 2023. It reports that all of the top 10 US commercial banks use chatbots, and it identifies risks that apply directly to agentic designs: inaccurate answers from language models, failure to recognise disputes, "doom loops" with no route to a person, and privacy and security exposure from chat logs. It states that "it is not suitable for chatbots to be the primary customer service vehicle". This is a regulator's observation, not legal advice. Check the rules that apply to your institution with your compliance team.

RiskControl to design inSource for the control
Wrong or invented answersAnswer from approved knowledge bases; test before releaseCFPB report; AWS notes on generative features
Customer trapped with no route to a personClear hand-off to a human agentCFPB report
Sensitive data in logsNever place account numbers in free-form fields; encrypt text and audio logs with KMS customer managed keysAmazon Lex V2 data protection and conversation logs pages
Agent acts outside policyLimit the agent to defined action groups and add approval stepsAmazon Bedrock Agents documentation
No evidence of why an action happenedConversation logs plus agent tracesLex conversation logs; Bedrock Agents tracing

AWS describes a shared responsibility model for Lex: AWS secures the infrastructure, and you are responsible for how you use the service, including data sensitivity and applicable laws. Lex conversation logs send text logs to Amazon CloudWatch Logs and audio logs to Amazon S3, and AWS notes they cannot be used with a bot subject to the Children's Online Privacy Protection Act. Use a policy layer such as AgentGRC, which keeps autonomous agent actions inside policy with the audit trail risk teams need, and read our guide to agentic AI risk management for the control framework.

What are the trade-offs, and when should you not use an agent?

  • Keep standard Lex intents for high-volume, predictable requests such as balance enquiries. They are easier to test and explain than an agent.
  • Do not give an agent write access to payments or customer records until limits, approvals and logging are proven in read-only use.
  • Do not make the bot the only channel. The CFPB warns about reduced access to human agents.
  • Budget for generative features. AWS says enabling them may incur additional charges, and Lex itself is billed per text or speech request.
  • Plan for platform change. AWS has placed Bedrock Agents Classic in maintenance mode, so keep the agent layer behind a clean interface.

How should a bank sequence adoption?

  1. Rank intents by volume and risk: start with read-only enquiries.
  2. Build the Lex bot with standard intents and measure how often it fails to understand customers.
  3. Add a knowledge-base answer path for product questions from approved content.
  4. Introduce one agent for one multi-step task with a small action group and a human approval step.
  5. Turn on logging and tracing and review samples with risk and compliance before widening scope.
  6. Orchestrate and govern at scale. ElixirClaw is an agentic OS platform to build, orchestrate and govern AI agents and automation workflows while keeping humans in control. Where decisions need a defensible record, the ElixirData Audit Agent gathers a decision's history, the policy version in force and the approvals involved.

To map one banking-chatbot decision your agents should make, and the context and policy it depends on, book a Context Workshop. See what happens in the workshop.

Frequently asked questions

What is an agentic banking chatbot?

It is an assistant that plans multi-step tasks and calls approved bank APIs, instead of only returning scripted answers. A person stays in the loop for risky steps.

How does Amazon Lex work with Amazon Bedrock Agents?

According to AWS, AMAZON.BedrockAgentIntent routes a request to a Bedrock Agent or knowledge base. The conversation stays with the agent until it marks the conversation finished, then Lex resumes control.

Does Amazon Lex support voice and text?

Yes. AWS describes Lex as providing ASR and NLU for voice and text conversations, published to mobile apps, web apps and chat services.

What do regulators say about banking chatbots?

The CFPB's 2023 report flags inaccurate answers, unresolved disputes, doom loops and privacy risks, and says chatbots are not suitable as the primary customer service vehicle. Take advice from your compliance team on the rules that apply to you.

How should conversation data be protected?

AWS advises never putting sensitive identifiers such as account numbers into free-form fields, and says Lex text and audio logs can be encrypted with KMS customer managed keys.

Next Steps with Agentic AI

Connect with our experts to explore implementing Agentic AI for banking chatbots. Discover how AI-powered decision-making and autonomous workflows can transform customer interactions, streamline banking support, and drive operational efficiency across financial services.

Follow XenonStack on LinkedIn: XenonStack on LinkedIn for new guides on agentic AI, data and platform engineering.

Table of Contents

navdeep-singh-gill

Navdeep Singh Gill

Global CEO and Founder of XenonStack

LinkedIn profile

Navdeep Singh Gill is serving as Chief Executive Officer and Product Architect at XenonStack. He holds expertise in building SaaS Platform for Decentralised Big Data management and Governance, AI Marketplace for Operationalising and Scaling. His incredible experience in AI Technologies and Big Data Engineering thrills him to write about different use cases and its approach to solutions.

Get the latest articles in your inbox

Subscribe Now

Interested in solving your challenges with the XenonStack team?
Get Started

Tell us about you and your priorities so we can prepare the right conversation. It takes about 2 minutes.

Please Select your Industry
Banking
Fintech
Payment Providers
Wealth Management
Discrete Manufacturing
Semiconductor
Machinery Manufacturing / Automation
Appliances / Electrical / Electronics
Elevator Manufacturing
Defense & Space Manufacturing
Computers & Electronics / Industrial Machinery
Motor Vehicle Manufacturing
Food and Beverages
Distillery & Wines
Beverages
Shipping
Logistics
Mobility (EV / Public Transport)
Energy & Utilities
Hospitality
Digital Gaming Platforms
SportsTech with AI
Public Safety - Explosives
Public Safety - Firefighting
Public Safety - Surveillance
Public Safety - Others
Media Platforms
City Operations
Airlines & Aviation
Defense Warfare & Drones
Robotics Engineering
Drones Manufacturing
AI Labs for Colleges
AI MSP / Quantum / AGI Institutes
Retail Apparel and Fashion

Continue