# AI assistants on a live storefront, with prices the browser cannot change Entry E22. Sector: Spice and rice e-commerce brand. Status: in production. Delivered: 2026-07. Source: https://ai.prospicience.in/work/storefront-ai-assistants For a spice and rice brand, Prospicience replaced a hosted store platform with a storefront the brand owns, plus two AI assistants: one helps the team with orders, stock, refunds and reports, the other lets customers manage their own orders. Prices are computed on the server and payments confirmed by the provider, so no browser can change a charge. Speed: Store, admin and AI assistants in three days. ## The challenge The store ran on a hosted platform with per transaction costs and limited control over checkout. Every order question, refund and address change went through the team by hand, and customers had no way to help themselves without emailing. ## Why it mattered Platform fees grow with every sale, and the checkout could not be shaped around the business. Manual order handling ties up the team exactly when festival sales peak, and slow answers on refunds and deliveries put repeat customers at risk. ## What we built - Built a storefront the brand owns: catalog, product pages, cart, guest checkout, customer accounts and content pages, with page addresses matching the old store so search rankings carry over. - Added a team AI assistant that answers questions on orders, revenue, stock, customers and reports, and handles fulfilment, refunds and payment status, each change only after a person approves it. - Added a customer AI assistant in the account area that looks up the customer's own orders and handles address changes, cancellations and refund requests, with the same approve step. - Built an admin area for products, images, orders and refunds, with every payment confirmed by the payment provider rather than the browser. ## The result - The brand runs its own store and checkout instead of renting one. - Customers check and change their own orders without writing in. - The team handles orders, refunds and reports by asking the assistant, with every change confirmed before it happens. - The store keeps growing: new features, including a reports page with CSV downloads, were still shipping in September 2026. ## Built for trust Prices are computed on the server at checkout and never taken from the browser, and an order is confirmed only when the payment provider says it is paid. Both AI assistants propose and then wait: refunds, cancellations, address changes and fulfilment happen only after a person clicks Approve. ## AI at work Two AI assistants work the store every day: one answers the team's questions on orders, revenue, stock and customers and prepares fulfilment and refunds, the other lets customers look up their orders and ask for changes. Every change waits for a person to approve it. What it produces: Answers on orders, revenue, stock and customers, Reports on request, Proposed refunds, cancellations and address changes, Order lookups for customers. - Customers check and change their own orders without writing in, which takes routine email off the team. - The team handles orders, refunds and reports by asking in plain words instead of working through admin screens. - At festival peaks, routine order questions are answered by the assistants, so the team's time goes to approvals and exceptions. How it works day to day: A team member or customer asks in plain words. The assistant looks up the order or the figures, answers, and where a change is needed, sets it out and waits. A person clicks Approve, and the store applies it with prices always worked out on the server. ## Figures from the delivered system - platform replaced: hosted store, now serverless - price calculation: server side only - order confirmation: payment webhook - region: functions co-located with the database - AI assistants: 2, team and customer ## The technology - A full storefront: browse, cart, checkout, accounts - Two AI assistants: one for the team (orders, stock, refunds, reports) and one for customers (their own orders, cancellations, refund requests), with every change waiting for approval - An admin area for products, images and orders - Prices worked out on the server, never in the browser AI models used: Claude (Anthropic). ### Technical notes - Server side pricing at checkout, with the client price treated as untrusted input - Order confirmation driven by the payment provider webhook rather than the redirect back to the site - Serverless functions co-located in the same region as the database to keep round trips short - Product pages rendered on the server with structured data for search - Storefront AI assistants using the same model provider as the rest of our stack ## Questions ### Is it safe to give an AI assistant access to orders and refunds? Yes, because it cannot act alone. Every change it proposes, whether a refund, a cancellation or an address change, waits for a person to approve it, and prices are always computed on the server, so no assistant and no shopper can change what gets charged. ### Why move off a hosted store platform? Control and cost at volume. Hosted platforms are the right answer early. Once transaction fees and checkout limits start shaping the business rather than serving it, owning the storefront is cheaper and far more flexible. ### How do you make sure an order is really paid? By waiting for the payment provider to confirm it directly rather than trusting the redirect back to the site. A shopper closing the tab at the wrong moment, or a crafted request, can never produce a confirmed unpaid order. ### What happens to search rankings when you move off a hosted store? Product, collection and page addresses mirror the old store one to one, and the few that differ redirect permanently, so existing links and rankings carry over to the new store. Capabilities: ai-agents-and-automation, platform-migration.