Amazon Just Sold the Wall

A few months ago we wrote that Amazon was doing everything in its power to keep third-party AI from becoming the discovery layer for its marketplace. They’d blocked dozens of bots. They’d won a federal injunction against Perplexity’s Comet browser. They’d poured billions into their own assistant. The thesis was simple:

That story just got a sequel none of us expected this fast.

What Amazon Did Last Week

On May 27, AWS launched the Agentic Shopping Assistant — ASA. It takes the technology behind Amazon's own in-house shopping AI — the assistant formerly known as Rufus, now folded into “Alexa for Shopping” — and packages it up for other retailers to deploy on their own sites.

The same engine Amazon built to keep customers inside its garden is now for sale to everyone outside it.

The numbers attached to the underlying system are the reason anyone should care:

Kate Spade is the first one live, using it to power a gift concierge. And read all of this against what we said in the spring and the whiplash is almost comical. The company that sued a browser for shopping on a customer's behalf is now selling the machinery to let brands run their own AI shopping experience.

It isn't a contradiction. It's the most honest thing Amazon has done in this whole saga.

The Detail Almost Everyone Missed

Here's the part that didn't make the headlines, and it's the most revealing thing in the entire announcement.

Kate Spade's gift concierge — the flagship proof point for Amazon's new retail-AI product — runs on Anthropic's Claude Haiku 4.5. Amazon Bedrock handles the orchestration, authentication, and evaluations; AgentCore and OpenSearch wire it into the catalog. But the actual reasoning, the part that talks to the shopper, is a model built by Anthropic.

Sit with the ownership chart for a second:

• Amazon has invested ~$8 billion in Anthropic, the maker of Claude.
• It also committed up to $50 billion to OpenAI in February, plus a $38 billion AWS infrastructure deal with OpenAI that closed in November.
• And the shopping AI it's now selling to the retail industry is powered by Claude.

So Amazon is selling on-site shopping intelligence, running on a model from one company it partly owns, while simultaneously bankrolling that company's biggest rival. This is what it looks like when a company decides the layer matters more than the model. Amazon doesn't care whose brain is inside the assistant. It cares about owning the assistant — the relationship, the data, the surface the customer actually touches. The model is a commodity it's happy to rent, even from a competitor.

That's the whole strategic insight of the last two years, sitting quietly in a press-release footnote.

Why This Isn't a Reversal — It's a Confession

The walling-off and the selling-off are the same strategy wearing two different outfits.

Both moves start from one belief: the AI experience is the most valuable layer in commerce, and whoever owns it owns the customer. Amazon walled itself off because it refused to let OpenAI own that layer for Amazon's products. Now it's selling ASA because it has realized the rest of the retail industry is going to build that layer with someone — and Amazon would rather be the someone.

This is the AWS playbook, exactly. Build infrastructure to solve your own problem at a scale nobody can match, then rent it to the whole market. They did it with compute. They're doing it with shopping AI.

And the pitch only works because the thesis is now conventional wisdom. Amazon isn't out there convincing retailers that on-site AI matters — that argument is over. Walmart proved it with Sparky. Nike proved it with NikeAI. Target proved it with its own ChatGPT app. Amazon is simply showing up to a market that already agrees and saying: you need this, we have the best version, buy it from us.

When the company that fought hardest to keep AI off its own platform starts selling on-site AI to everyone else, the debate about whether brands need to own their AI experience is finished. The only debate left is who they build it with.

The Lawsuit, in the Details That Got Skipped

To understand how serious Amazon is, it helps to look closely at the Perplexity fight — because the specifics are wilder than the summary.

Amazon first detected Comet quietly operating on its store back in 2024, via Perplexity's old “Buy with Pro” tool, and says it warned Perplexity to stop at least five times. When Amazon finally deployed a technical block in August 2025, Perplexity pushed an update that circumvented it within 24 hours — and, Amazon alleged, disguised Comet's traffic to look like an ordinary Google Chrome session. That single act of concealment is what handed Amazon its legal win, because intentional evasion maps cleanly onto the Computer Fraud and Abuse Act.

A few more details that say everything about how tangled this market is:

Perplexity's public response was a blog post titled “Bullying is not innovation,” arguing that Comet only does what the customer told it to do — that blocking it is like a store refusing to let you bring a personal assistant. Amazon's counter, which the judge found persuasive, is quietly radical: an AI agent acting on a customer's behalf does not inherit that customer's right to be there. The customer can shop on Amazon. The agent cannot — not without permission.

If that holds, it's the load-bearing wall of the whole agentic-commerce era. Once an agent needs permission, it has to negotiate. Once it negotiates, the platform sets the terms.

The Tell Inside the New Press Release

There's a line in Amazon's ASA messaging that deserves more attention than it got. AWS went out of its way to promise that retailers keep control of their own customer data, product catalogs, and business rules.

Sit with why that promise had to be made at all.

Amazon's retail arm competes directly with most of the retailers it's now pitching. Asking a brand to run its AI shopping experience — the layer that sees every customer question, every hesitation, every intent signal — on infrastructure owned by its largest competitor is a genuinely strange ask. The data-control language exists because Amazon knows it.

It's worth noting that the smartest retailers are already hedging. Tapestry, Kate Spade's parent, didn't just adopt Amazon's tool — it also built and patented its own internal AI system, “Mira,” to keep its merchandising and customer intelligence in-house. Use the giant's recipe for the storefront; keep your own brain for the business. That instinct is the whole game.

What This Means for Everyone Who Isn't a Giant

Here's where the ASA launch is genuinely useful as a signal — and genuinely incomplete as a solution.

It validates the thesis completely. The most sophisticated commerce company in the world just declared, with a product, that brand-owned on-site AI is the future and that building it from scratch is too slow for almost everyone. That's the entire argument, shipped by the one company whose opinion the market can't dismiss.

But look at who ASA is actually for. A 60-day enterprise engagement with the AWS Generative AI Innovation Center, custom-architected on Bedrock, with system-integrator partners — that's a fit for Kate Spade and the tier of retailer that can staff and fund a project like that. It is not a fit for the thousands of mid-market and D2C brands that make up most of commerce — the ones who need a salesperson on every page but don't have a quarter to spend on a custom build, an enterprise cloud contract to negotiate, or a reason to route their crown-jewel customer data through their largest competitor.

So the launch sharpens the exact gap we've been pointing at all along:

That second gap is the one that matters now. And it's wider than it looks, because the brands on the wrong side of it are exactly the ones with the most to lose. They don't have Amazon's leverage to push back against the platforms. They don't have Walmart's budget to build Sparky. If a customer asks an AI for the best of something in their category and the answer sends them nowhere — or sends them to a 2015 product page — they don't have a customer relationship anymore. They have a wholesale order, if that.

The Pattern Holds

Step back and the move fits the pattern we've tracked for months. Amazon, Walmart, Nike, Target — every giant converged on the same conclusion: the AI experience can't be outsourced and the customer relationship can't be rented. The only thing that changed last week is that Amazon went from defending that conclusion for itself to selling it to the market.

The brands that internalize this don't wait. They make sure their products show up in the AI conversations happening off their site, and they make sure that when a customer lands on their site, the experience can hold its own against the AI conversation that brought them there. Amazon just made the first half easier to believe and the second half easier to buy — if you're big enough.

For everyone else, the lesson is the same as it's always been, only louder. The most ruthless platform in commerce just told you, by reaching for its wallet, that owning your AI experience is non-negotiable. The open question is no longer whether to do it. It's whether you'll do it on infrastructure built by the company you compete with — or on infrastructure built for you.

Amazon already decided the AI experience is worth owning. They've decided it twice now, in two opposite directions, and landed in the same place both times.

We're building the infrastructure that lets every brand own their AI shopping experience — not just the giants, and not only the ones who can commission a custom build. If that resonates, we'd love to talk.