Two People, 10 Weeks: Inside the AI Commerce Agent Redmond Built on Shopify's Storefront MCP

A Salt Mining Company Built a Commerce Agent in 10 Weeks

Redmond built a production AI commerce agent in 10 weeks using Shopify's Storefront MCP

How Redmond built a production AI commerce agent in 10 weeks with Shopify Storefront MCP to scale support and gain full control.

www.shopify.com

Redmond has been pulling natural salt out of an ancient deposit in central Utah since 1958. What began with two brothers digging salt has grown into a company with more than 10 brands, including Real Salt and Re-Lyte, the electrolyte supplement that took off during the pandemic. The business runs on Shopify Plus.

The case study Shopify published is unusually specific for e-commerce operators. When the managed AI customer service tool Redmond had been using was discontinued, the team faced a choice: pay another vendor, or build their own. They chose the latter, and using Shopify's Storefront MCP as the foundation, assembled a production AI commerce agent in 10 weeks, going live in February 2026.

Two people did it. Phillip Hinson, an applied AI developer who had been building internal knowledge tools with RAG (retrieval-augmented generation), handled development, while Jeremiah Payne, the development team lead who had managed Redmond's Shopify infrastructure since 2018, provided direction and strategic decisions.

They Built Rather Than Bought for Control, Not Cost

Let's clear up a likely misreading first: this build was not a cost-cutting decision.

Redmond's customers routinely ask detailed questions about ingredients and where raw materials are sourced. For a natural products company, accuracy on those answers maps directly to business risk. At the time, managed AI chat tools generally did not expose system prompt controls to merchants. That is fine for most use cases, but for a company that needed precise guardrails around product-specific claims, it was a dealbreaker, as Hinson describes in the case study.

"

With any managed tool, there's a trade-off between convenience and control. We couldn't customize the system prompt or control the data sources, which made it difficult to build guardrails around the specific product questions we knew required careful answers.

Source: Phillip Hinson (Applied AI Developer, Redmond)

Hiring was the other driver. Redmond runs a deliberate hiring culture with multi-step interviews and cultural alignment checks, which means CX headcount cannot grow as fast as demand. Payne's line about looking for ways to scale customer experience without proportionally growing the team is where the build started.

What Filled the 10 Weeks, and What the Protocol Absorbed

This is the heart of the case. The period itself matters less than what they did not have to build, because that is what determines whether anyone else can repeat it.

Storefront MCP gives every Shopify store its own endpoint at https://{shop}.myshopify.com/api/mcp, exposing product search, cart operations, and policy lookups to AI agents in a standardized form. According to Shopify's developer documentation, tools such as search_catalog and get_product conform to the Catalog capability of UCP (Universal Commerce Protocol), the standard Shopify co-developed with Google. In other words, getting product data in front of an AI has become a connection problem rather than an implementation problem.

Hinson started from Shopify's MCP reference app and needed about a day to connect it to a development store and deploy to Azure. His description of the experience, that it is like USB-C and should work almost everywhere, captures the point.

The remaining nine weeks went into what the protocol does not cover.

One technical decision stands out: implementing Anthropic's prompt caching. The system prompt and tool definitions are large and stable across requests, so caching them cut input token costs by up to 10x on cache hits and reduced overall AI spend by roughly half. The conversational layer itself runs on Anthropic's Claude, which interprets each question, decides when to search, and composes the response.

The measurement design reveals a philosophy as well. Purchases are matched to chat conversations through a multi-tier system with varying confidence levels, from direct cart tracking to time-window correlation, and the team deliberately chose the conservative end. Payne's remark that they do not want data making the tool look better than it is addresses the failure mode that usually breaks internal AI reporting first.

Before launch, the team also exported nine months of conversation logs from the previous tool and analyzed where answers had fallen short. Unglamorous work, but it is the foundation of the accuracy they got.

The Day a New Product Appeared While He Was on Vacation

The clearest demonstration of a live data protocol's value came during launch week.

The day after the agent went live, Redmond released a new protein powder. Hinson was on vacation in Mexico with his family. A colleague messaged to congratulate him on updating the agent's product knowledge so quickly. He had done nothing. Storefront MCP had pulled the new product from Shopify's Catalog automatically, and the agent could answer questions about it the moment it was published.

With the previous managed service, every new product required manually initiating a content update. That difference compounds into monthly operational hours. Shopify's reason for deploying an MCP server to every store is precisely the disappearance of that sync work.

There were secondary effects too. When Payne consolidated four Shopify stores, he had been evaluating paid apps to migrate historical customer and order data, then completed it himself using the Shopify MCP and Claude. He went further and built middleware that listens to webhooks and automatically migrates new customers and paid orders. The fact that the protocol spilled over into operations beyond the chatbot is an easily overlooked benefit at evaluation time.

Can Other Teams Copy This Directly

From here a different angle is needed, separate from the vendor's narrative.

Start with the numbers. The case study describes post-launch Anthropic API token costs as minimal, but no figure is disclosed. Azure hosting costs, the price and name of the previous managed vendor, the specific monthly conversation volume, and any accuracy metric are all undisclosed as well. The case study itself notes that the 10 weeks of internal development time is not included in the cost. Because the previous service was an all-in subscription covering hosting, model, support, and updates while the custom build separates those components, the comparison is not apples to apples.

Then there is the success rate. Sinch surveyed 2,527 senior decision makers across 10 countries for its 2026 report The AI Production Paradox and found that 74% of enterprises rolled back or shut down a live AI customer communications agent. The rollback rate rose to 81% among organizations with fully mature guardrails, which the company reads as evidence that better monitoring surfaces failures sooner. Combined with MIT's Project NANDA finding that 95% of generative AI pilots never reach production, Redmond getting to production in 10 weeks puts them on the exception side.

The case against building is worth reading too. Fin, the AI customer support product from Intercom, estimates in its build versus buy analysis that a production-grade AI customer service agent runs $300,000 to $500,000 in first-year cost with $150,000 to $250,000 in annual maintenance. That is a vendor arguing its own book and should be discounted accordingly, but its citation of Gartner research, that 62% of underperforming AI projects trace back to insufficient data preparation versus under 15% for technology limitations, aligns neatly with why Redmond succeeded. Hinson says as much himself: the hard part is not the AI, it is knowing your product well enough to teach it.

Security is a live issue as well. MCP spread fast, and prompt injection, tool poisoning, and credential misuse have been flagged repeatedly. It is no accident that Redmond made PKCE authentication, log sanitization, and rate limiting an explicit production hardening phase. A 10-week build that skips that phase is a different thing entirely.

What This Means for E-Commerce Operators

The decision sequence looks like this.

First, ask whether the need for control is real. Redmond built because it had to govern the exact wording of answers about ingredients and sourcing. If order status inquiries make up most of your volume, the case for giving up a managed tool's convenience is thin. Projects that start from "it will be cheaper" tend to land on the failure side of the statistics above.

Second, even when building, cut the scope first. The reproducibility of this case rests on Storefront MCP eliminating product data integration and the reference app eliminating authentication and scaffolding. The move toward MCP servers from commercetools and other platforms is spreading beyond Shopify, so whether your platform offers an equivalent connection layer is the first thing to check.

Third, recognize that curating the knowledge base is the actual work. Stripped down, what Redmond built is a sync pipeline for approved content, embeddings to retrieve from it, and a system prompt that constrains the answers. All three hold their value regardless of which model is fashionable, and they survive a model swap. Conversely, building the agent first while product information remains scattered across the organization produces a machine that is confidently wrong.

Conclusion

Redmond's 10 weeks is a concrete example of protocols lowering the difficulty of building AI agents. It also shows that what dropped was the difficulty of connection, not the difficulty of operation. The next thing to watch is the transactional capability the company has signaled, starting with refunds and cancellations for unfulfilled orders. Moving from informing to transacting raises the weight of permission design and exception handling all over again. Read alongside Shopify's figures on surging AI-driven orders, this is a case worth following.