Rahul Fernandes
Rahul Fernandes Rahul Fernandes

Every trade outlet covering retail's AI moment has landed on the same phrase: the "decision layer." It's the recommendation engines, conversational agents and pricing logic that now sit between a shopper's intent and their purchase. That framing is useful for a boardroom slide. It's useless for an engineering team that has to ship something.

Strip the strategy language away and the decision layer is just a routing problem: which retailer's product data gets pulled into an AI agent's context window. Whoever gets pulled into that context window gets considered. Whoever doesn't, doesn't. That's a plumbing problem, and the plumbing is being standardized right now.

The stack forming underneath the hype

Four protocols have emerged in the last twelve months to govern how agents talk to retailers, and each solves a different layer of the problem:

  • Model Context Protocol (MCP), Anthropic's open standard, gives an agent a live interface into a merchant's systems instead of forcing every AI platform to build a bespoke integration per retailer.
  • The Agentic Commerce Protocol (ACP), built by OpenAI and Stripe and now underpinning ChatGPT's Instant Checkout, standardizes two things: a product feed spec (how a merchant describes what it sells) and a checkout spec (how an order actually gets placed). Notably, ACP's checkout spec is deliberately kept out of PCI scope — cardholder data never transits the model.
  • Agent Payments Protocol (AP2), from Google and more than sixty payments partners, sits on top of MCP and the Agent2Agent protocol and solves authorization: it uses cryptographically signed "Intent Mandates" and "Cart Mandates" so a merchant can prove an agent actually had a user's permission to buy, at that price, at that moment.
  • Universal Commerce Protocol (UCP), which Google launched in January with Shopify, Etsy, Wayfair, Target and Walmart as co-developers, tries to cover the whole journey — discovery through post-purchase support — with a live, queryable product API rather than a static feed.

None of this is theoretical. It's shipping documentation with required fields, refresh intervals and error codes. That's the actual battlefield.

Catalog sync is the first gate, and most retailers miss it

An agent can't recommend a product it can't see accurately. The emerging convention is that price and availability need to be correct within a window measured in minutes, not the daily-batch cadence most retail systems were built around. A stale price or a "ships in 2 days" claim that's wrong doesn't just cost a sale; it can get a merchant's feed deprioritized or flagged by the agent's trust layer entirely.

Latency isn't a UX nice-to-have, it's a placement requirement

Amazon's own account of building Rufus is instructive here, and it's an engineering-heavy story: a custom LLM trained on catalog, reviews and Q&A data, retrieval-augmented generation to pull in fresh product facts, continuous batching so new requests don't wait behind slow ones, and a streaming response architecture that starts rendering an answer before it's fully generated. That's the level of infrastructure investment now required to be the agent doing the recommending.

For everyone else — the retailers being recommended, not building the assistant — the requirement inverts but doesn't shrink: your systems need to answer an external agent's query about stock, price or delivery estimate fast enough that it doesn't time out of the agent's decision window. If your inventory takes 800 milliseconds to answer, you're not in the running; the agent has already ranked a competitor that answered in 80.

Two live architectural bets worth watching

Amazon has blocked OpenAI's ChatGPT-User and OAI-SearchBot crawlers — its listings largely can't surface in ChatGPT's shopping results at all. That's a bet that owning the assistant (Rufus) beats being discoverable inside someone else's. Walmart has made the opposite bet: Sparky, its own assistant, is being built to plug into ChatGPT and Gemini rather than wall itself off. One architecture optimizes for a closed, vertically integrated stack; the other optimizes for being a well-behaved, protocol-compliant node inside everyone else's agent. Both are defensible engineering strategies.

The takeaway for anyone not named Amazon or Walmart

Most retailers aren't going to build the agent. They need to be legible to one. That means: a machine-readable product feed with the required price and availability fields; an API layer that can serve live inventory and pricing queries in double-digit milliseconds, not seconds; and a checkout endpoint that can accept and honor a signed mandate rather than assuming a human clicked "buy." None of that shows up on a strategy slide. All of it determines whether you show up in the answer.