Overview
Our client, a large retail group selling consumer goods, approached NeenOpal for a customer support assistant for its website, along the lines of the conversational shopping assistants the major online marketplaces now offer. The brief arrived as an exhaustive list of capabilities the assistant had to cover, spanning product research, cart building, order tracking, warranty and complaint handling, post-purchase care, and escalation to a human agent.
NeenOpal built the assistant as a working proof of concept on Amazon Connect, the first time the team had taken Connect into a customer engagement. A conversational layer handles the dialogue, a governed tool layer gives the assistant controlled access to customer and product information, verified sign-in gates anything account specific, and a handover path passes the full conversation to a live agent when a person is needed. The build went from brief to working demonstration in four weeks.
- 4
- Weeks From Brief to Working Chatbot Demonstration
- 52
- Capabilities Delivered in One Conversational Assistant
- 100%
- Of Chat Context Carried into Every Live Agent Handover
- 0
- Account Answers Returned Without Verified Identity
Customer Challenges
The client came to NeenOpal with a capability list rather than a problem statement, so the work of defining what the assistant had to do, and proving it could be done on an untested platform, sat with the delivery team.
No Self-Service Support Channel on the Storefront
Customers browsing the client's website had no instant way to ask a product question, check where an order had reached, or start a warranty claim without contacting a person. Every enquiry, however routine, depended on a human being available to answer it, which put ceiling on how quickly shoppers could get an answer.
A Broad Capability List to Satisfy in a Single Assistant
The brief covered product research and comparison, cart building, promotions, order tracking, ticketing, complaints and warranties, post-purchase care, service centre location, and live agent escalation. Delivering that breadth coherently in one assistant, rather than as a set of disconnected flows, was the central difficulty of the engagement.
Account Data That Could Not Be Exposed Without Verification
Anything specific to a customer, from order status to warranty history, needed the person to be identified before it could be returned, and personal information had to stay protected throughout the conversation. The assistant therefore could not be allowed to assert who it was speaking to on its own, including when a browsing visitor decided to sign in part-way through a conversation.
AI Solutions Platform on AWS
Solutions
NeenOpal delivered the proof of concept as a single assistant covering the client's full capability list, built on Amazon Connect and backed by a governed data access layer rather than by direct database queries from the conversation.
A Conversational Assistant Built on Amazon Connect
NeenOpal built the assistant on Amazon Connect, taking the platform into a customer engagement for the first time in the organization. Amazon Lex receives each customer message inside the contact flow, and Amazon Q in Connect interprets intent, selects the right source of information, and composes the reply in context.
Governed Data Access Through an MCP Tool Layer
Rather than letting the conversational layer reach data directly, NeenOpal placed a Bedrock AgentCore Gateway and a set of AWS Lambda functions between the assistant and the client's information. Each function exposes a specific, bounded operation, so the assistant retrieves what a given request legitimately needs and nothing more. A request interceptor sits in front of every call and replaces the customer identity with the one from the verified session, so the assistant cannot be talked into acting as somebody else.
Verified Sign-In That Keeps the Conversation Intact
Anything tied to a customer account required the person to sign in first, with identity taken from the authenticated session rather than from anything stated in the conversation. A visitor can begin as a guest, be asked to sign in at the point it is actually needed, and carry straight on, because the assistant updates the existing conversation rather than starting a new one, so nothing said beforehand is lost and the exchange stays as a single record.
A Custom Interface with Rich, Interactive Results
NeenOpal replaced Connect's fixed chat widget with a custom interface built on Connect's own chat library. The assistant returns product carousels, order cards, cart summaries and service centre maps with directions and tap-to-call, rather than paragraphs of text, all inside the client's branding. Customers can also download a transcript of the conversation.
Commerce and Post-Purchase Capabilities in One Place
The assistant covers the journey rather than a single step. A shopper can research products, compare them, check the promotions they personally qualify for and build a cart, then return later to track an order, raise a complaint, start a warranty claim, or request post-purchase support, all in the same conversation. Checkout and payment were deliberately left out of the proof of concept and can be added when the solution goes forward.
Ticketing and Live Agent Handover with Full Context
The assistant raises service cases directly into Amazon Connect Cases, with the product, warranty position and service address drawn from the customer's record rather than asked for. When a request needs a person it escalates to a live agent rather than ending the conversation. The agent receives the full context of the exchange that preceded them, so the customer does not repeat themselves and the agent picks up a conversation already in progress rather than starting a new one.
A Path to Production Defined Alongside the Build
The proof of concept was built so the route to production was understood rather than discovered later. Taking the assistant live requires adding an Amazon EC2 instance to host the application layer, with the conversational and data access components carrying over as built.
Services
Amazon CloudWatch
AWS Lambda
Amazon DynamoDB
Amazon S3
Amazon Connect
Benefits
Sign-In Without Losing the Thread
A visitor can browse anonymously, sign in at the moment it is needed, and continue the same conversation. The assistant picks up exactly where it left off, and the whole exchange remains a single record.
Escalation That Does Not Lose the Conversation
When a request reaches a live agent, the agent arrives with everything the customer has already said. The handover removes the repetition that usually accompanies a move from automated to human support, and the agent spends their time resolving rather than reconstructing.
A Single Assistant Covering the Whole Customer Journey
Product research, cart building, order tracking, complaints, warranties and post-purchase care sit in one conversation rather than across separate tools or forms. A shopper moves between them without restarting, and the client maintains one assistant rather than a collection of disconnected flows.
Account Information Protected by Verified Identity
Nothing specific to a customer is returned until that customer has signed in and been verified, and identity comes from the session rather than from the conversation. The client gains a self-service channel without exposing order history or personal information to whoever happens to be asking.
A Shopping Experience, Not a Text Chatbot
Products, orders and service locations appear as interactive cards and maps inside the client's own branding, so the assistant reads as part of the storefront rather than as a support widget bolted onto it.
A Proven Amazon Connect Capability
The engagement established Amazon Connect as a delivery capability within the organization for the first time, with a working assistant to demonstrate rather than an untested proposal. The client can now weigh Connect against Bedrock on the basis of something built, and the four week build gives a realistic timeline for taking it further.
Conclusion
With NeenOpal's support, the client has a working customer support assistant covering its full capability list, built on Amazon Connect and demonstrated end to end in four weeks. By routing data access through a governed tool layer of bounded Lambda functions, gating account information behind verified sign-in, and passing complete context to a live agent on escalation, NeenOpal showed that a conversational assistant of real scope can be delivered on Connect rather than on the platform originally proposed. The proof of concept gives the client a concrete basis for choosing its path to production, with checkout, payment handling and an Amazon EC2 hosting layer as the remaining steps.
FAQ
Common questions about building a customer support chatbot on Amazon Connect
What is Amazon Connect, and can it support a customer support chatbot?
Amazon Connect is AWS's cloud contact centre service, most often associated with routing calls and chats to human agents. It also supports automated conversational experiences, which means a single platform can carry both the assistant and the escalation path to a live agent. Building the assistant inside Connect rather than alongside it is what allows a conversation to move from automated to human handling without the customer starting again.
How does a customer support assistant access account data safely?
The assistant should not query customer records directly. Placing a controlled layer between the conversation and the data, in this project a gateway exposing bounded Lambda operations over MCP, means each request maps to a specific operation rather than to open access. Pairing that with verified sign-in, where identity is taken from the authenticated session and injected into every request rather than supplied by the conversation, keeps account information from being returned to the wrong person.
Can a chatbot show more than plain text?
Yes, with a custom interface. Amazon Connect's built-in widget is fixed in look and layout, but its chat library allows a bespoke front end, which is how this assistant returns product carousels, order cards and service centre maps inside the client's own branding.
How long does it take to build a conversational support assistant of this kind?
This assistant went from brief to working demonstration in four weeks, covering product research and comparison, cart building, promotions, order tracking, ticketing, complaints and warranties, post-purchase care, service centre location and live agent escalation. A production deployment adds the hosting layer and the commerce components left out of the proof of concept, such as checkout and payment handling.
Case studies
Related case studies
AI-Augmented ETL for Automating Multi-Portal Retail Reporting on AWS
A consumer beauty and personal care brand needed to bring its Walmart and Amazon supplier and performance data into a single AWS RDS PostgreSQL database. Walmart's data was split across a standard supplier API and three separate reporting portals (Retail Link, Scintilla, and Supplier One), each with its own authentication, file formats, and update cadence. NeenOpal designed a hybrid ETL platform with two ingestion paths running in parallel: an orchestrated API pipeline for the data Walmart exposes directly, and an AI-driven report retrieval pipeline for the data that fell outside the scope of the client's existing API access. Both paths converge on a shared Glue transformation layer, a common S3 staging area, and a single PostgreSQL database, with a separate one-time Spark job bringing in Amazon Vendor Central's historical data. The result is a single, config-driven system rather than a set of one-off integrations.
Read case study
AI-Powered Video Persuasiveness Prediction Using Generative AI
A leading non-profit organization invests heavily in video campaigns ($50K–$200K per video) to inspire audiences and drive meaningful action. However, their manual impact testing process was slow, costly, and limited to only a small subset of high-budget productions. Without a scalable evaluation method, most videos went untested, creating a high-risk content strategy with minimal feedback to guide creative decisions. NeenOpal developed a GenAI-powered video analysis solution that automatically extracts persuasive features from video content and predicts impact scores, enabling data-driven content strategy across their entire video portfolio.
Read case study
LLM Migration from OpenAI to Amazon Bedrock
A leading marketing advisory specializing in wealth management and investor communications was running a core generative AI capability for AI powered, FINRA and SEC compliant email generation on OpenAI GPT models. While the solution delivered early value, growing concerns around data security, compliance governance, escalating API costs, and single vendor dependency created an urgent need to migrate to a more enterprise grade and sovereign AI stack.
Read case study
Modernizing Automotive Damage Assessment with AI and AWS
Our client, a leading automotive repair service provider, faced increasing challenges in quickly and accurately assessing vehicle damages for insurance claims, repairs, and customer estimates. Traditional manual inspections consumed time, increased errors, and delayed decision-making. By using Amazon Bedrock and OpenSearch vector search, the client transformed their damage processing workflow, enabling faster, scalable, and more accurate repair cost assessments, resulting in enhanced customer satisfaction and operational efficiency.
Read case study
Transforming Retail Operations with Forecasting & BI on AWS QuickSight
A leading Sri Lanka-based retail chain was looking to take its first step towards Business Intelligence (BI) and data-driven decision-making. Operating across multiple locations, they faced inefficiencies in inventory planning, vehicle utilization, and order fulfillment. With no centralized data repository or analytics framework, they lacked visibility into operational bottlenecks and customer demand trends. To optimize their retail operations, they partnered with NeenOpal to implement a data warehousing solution integrated with predictive forecasting and an interactive BI dashboard using Amazon QuickSight.
Read case study
Driving Smart Manufacturing Transformation with AWS & AI-Powered Command Center
Manufacturers are rapidly embracing smart manufacturing solutions to overcome challenges like unplanned downtime, poor forecasting, and siloed data. NeenOpal, in collaboration with AWS industrial services, has developed an AI-powered Manufacturing Command Center (MCC) that integrates Amazon AI, Industrial IoT, and predictive analytics. Using tools like Amazon Q, our solution transforms raw manufacturing data into actionable intelligence, enabling faster decisions and improved operational efficiency.
Read case study