From first message to confirmed order.
An agentic shopping assistant that understands what a customer needs, finds the right handful of products, and carries them all the way to checkout in one conversation.
- Understands intent: trip, group size, season and budget, picked up from plain words.
- Recommends, not lists: two strong options, with the better pick flagged.
- Transacts: cart, budget tracking, discounts and checkout, all in the chat.
Meet our client
- Client
- E-commerce and retail technology platform
- Industry
- Ecommerce & Retail Technology
- Market
- AI-powered & agentic commerce
In a nutshell
An e-commerce and retail technology platform engaged Paramount to design and build a production-grade shopping assistant as part of an agentic ecommerce initiative on AWS Bedrock AgentCore, focused on camping and outdoor gear. Customers describe what they need in their own words; the multi-agent assistant searches a live catalog, compares with context, manages the cart and checks out, with memory that persists across every turn and every visit.
Browse, hesitate, leave.
Customers know the goal, not the product. Keyword search only hears a word or two of what they say.
of online carts are abandoned. This assistant works on one part of the problem: helping customers find and choose the right product.
- They know the goal, not the product: customers arrive to solve something (a trip, a home office) without knowing what to buy.
- Search returns a wall of options: one keyword brings back hundreds of results, with no guidance on which is right.
- Keyword search can't converse: it can't take in context or guide a customer from a need to a decision.
- Assistants forgot everything: existing chat tools lost context between steps and couldn't run a cart or a checkout.
A conversation that never loses its place
AgentCore can route turns of the same conversation to different runtime containers. State kept only in one container's memory is not there when a turn lands in another.
Container A
Container B
Container C
Illustration. When a turn lands on a different container, only state stored outside it carries over.
- State outside the process: session, cart, history and profile persist in DynamoDB and S3, so any container can pick up a turn, and the conversation carries over through recovery and later visits.
- Structured output for the screen: every reply carries text plus signals that tell the React UI which panel to draw.
- Relevant retrieval: catalog search had to stay accurate across very different ways of asking, with a Knowledge Base fallback.
- Deterministic money moves: cart and checkout run through tools with fixed contracts, not free-form model output.
An agent, not a widget
One orchestrating agent that reasons with Claude and acts through tools. Tap any part to see its job.
- One orchestrator, many tools: Claude decides what to do each turn and calls research, profile, cart and panel tools.
- Retrieval-augmented search: a Bedrock Knowledge Base with Titan embeddings and S3 Vectors backs up catalog search.
- Text plus structure: each reply is a ChatResponse with the message, the panel to show, products, cart and context pills.
- Not a recommendation widget, not a scripted bot: it reasons, searches, compares, carts and checks out, across any catalog.
From hello to checkout, one conversation
A returning customer plans a rainy August trip to Yellowstone with a $2,000 budget. Step through each turn and see what the agent shows, thinks and sends.
- 1Understands intentUse case, budget, constraints
- 2Searches the live catalogThe right handful, not hundreds
- 3Compares with contextOnly the attributes that matter
- 4Recommends the best fitFlags the better pick, proactively
- 5Carts & checks outConversational and transactional
What the agent sent the screen this turn: a structured ChatResponse alongside the reply text.
- Context from one sentence: “August” and “Yellowstone” became pills with no form or dropdown, and the agent knew August means afternoon rain.
- Two options, not hundreds, with the better pick flagged; the comparison used only weight, capacity, season and best use.
- The cart moved with the conversation: “add two more tents” updated quantity, pricing and the live budget tracker at once.
- One last useful pick: water purification tablets, flagged as essential for backcountry safety, added with a single “sure”.
Recognised, not reset.
Most stores greet a loyal customer like a stranger. This agent loads who they are from the first message, and rewards them at checkout.
- A persistent profile: location, group size, season, experience level and past purchases, kept in DynamoDB.
- Loaded from the first message, so the agent can say “welcome back” and pick up where the last trip left off.
- Rewarded at checkout: a 10% bundle discount and a 5% loyalty discount appear as clear line items, nothing to ask for.
- Updated after every order: new purchases are added to the profile, so the next visit starts even further ahead.
Real screens from the live demo
Tap any screen to see it full size.
Not search. An advisor.
When there's a smarter choice for the customer's situation, the agent says so. It doesn't wait to be asked. Tap a card to flip it.
- Production-grade multi-agent orchestration on AWS Bedrock AgentCore, not a prototype chatbot.
- Structured UI signals travel with every conversational reply, so the storefront stays rich and in sync.
- Distributed session resilience: memory survives containers, turns and visits.
- Deterministic cart and checkout wired into the AI's reasoning, so the money side is always exact.
More carts, bigger baskets
How the basket grew in one conversation
Figures from the live demo conversation above.
Fewer abandoned carts
Customers leave with the right product, not a wall of options they scroll past and abandon.
Higher order value
The agent guides the whole purchase and adds a final complementary pick at checkout.
Always-on associate
Discovery, comparison, cart and checkout, any catalog, any vertical, around the clock.
- A full multi-agent shopping assistant deployed on AWS Bedrock AgentCore.
- End-to-end conversational shopping, from discovery to order confirmation.
- Persistent memory and loyalty-aware personalisation across sessions.
- A reusable architecture for future agentic commerce experiences, in any catalog or vertical.
Not search. An advisor.
An agent that reasons about what each customer actually needs, from first message to confirmed order.
- 01Your business objective
- 02The technical context
- 03The right next step