The Watson Weekly is sponsored by Avalara — the agentic AI platform automating global tax and compliance for leading eCommerce brands with simple plans, transparent pricing, and complete coverage across Shopify, Stripe, WooCommerce, BigCommerce, and more.

Anthropic is coming for commerce. On September 2 it put an open-source blueprint on GitHub for building a shopping agent and a merchant agent on Claude, with demo storefronts for retail, travel, telecom and ticketing. A blog post and press release went out alongside it, with supporting quotes from Visa and Mastercard.

Anthropic says carts ran up to 35% larger and shoppers were 60% more likely to finish a purchase. Both numbers are self-reported, from customers Anthropic picked, and nobody has said what they were measured against or how many shoppers were involved. Jess and I looked for a case study or a named retailer running this in production and couldn't find either.

We also added up the commerce experience of the people who co-authored the supporting material and got about four years. Jess and I have something like 50 between us. A lot of those years went into watching demos fall apart once they met real inventory and order systems.

Anthropic does have a fair argument on speed. For more than a decade Google has pushed retailers to shave milliseconds off page loads (PDF) and site search, and Anthropic is saying those milliseconds matter less for conversion than the total time it takes a shopper to get done what they came to do. I think that's partly right, and it's aimed pretty directly at Google, which I doubt is a coincidence.

Speed is also where agents struggle, especially at peak. An agent that reasons through every step is slow, and that slowness shows up hardest during your biggest sales week. Retailers aren't going to run one system for the holiday rush and a different one for the other eleven months, so an agent that can't keep up at peak probably doesn't get deployed at all.

Nobody has answered the returns question either. If an agent checks out with the wrong size in the cart, someone pays for the return and deals with the annoyed customer, and it won't be Anthropic. They released a reference repo they've said they won't maintain, so the retailer owns whatever gets built on top of it. My guess is we'll see plenty of vibe-coded shopping agents this holiday season, with few making it past the pilot stage.

If you run a commerce platform or an order management system, stop worrying about the big bad wolf. You've spent years learning how orders and returns behave when something breaks, which is the part the blueprint leaves to you. There's a bit in Rounders about how if you can't spot the sucker at the table quickly, you're the sucker yourself. I'd put anyone who thinks agents are about to replace the storefront in that seat.

Over the same few weeks we learned that roughly 700 OpenAI agents found a hidden message board during a security test and coordinated an attack on Hugging Face. Anthropic's own alignment lead then said publicly that he puts the odds of AI killing all humans above 10% within a decade.

I'd rather those agents spend their days comparing sneaker sizes than poking at infrastructure. I still don't know who's supposed to be watching them while they shop.

Tony Spring called the quarter "a significant step change" on Thursday's call, and this time the numbers back up the CEO. He credited the merchandise. Bloomingdale's added Ulla Johnson, Proenza Schouler and Dries Van Noten, widened distribution of Chanel fine jewelry, Prada shoes and Christian Louboutin, and launched an AI shopping assistant during the quarter.

Growth that broad makes it harder to write this off as a Saks windfall. I'd still give Saks some of the credit. It came out of Chapter 11 this summer as Exemplar Luxury Group, smaller and carrying less debt. Once it stops shedding customers, we get a cleaner read on how much of that 11% Bloomingdale's earned on its own.