Your Next Customer Is an AI Agent Asking About Returns at 3 a.m.

At 2:47 a.m. Mountain Time, a shopping assistant acting for a runner in Denver opens two conversations at once.

To Brand A, it sends a structured query against the merchant's commerce endpoint: size 9.5 trail shoe, ship to 80203, needed by Friday, and what happens if the fit is wrong. The reply lands in roughly 300 milliseconds: 14 units in stock, ships today, arrives Thursday, 60-day returns with a prepaid label. Signed, structured, cacheable.

To Brand B, it asks the same question. There is no endpoint. The agent falls back to the product page, where the size selector renders in JavaScript it cannot fully resolve, the delivery estimate hides behind a zip-code modal, and the returns policy lives in a PDF last touched in 2023. The agent scores its own confidence in the answer, finds it too low to present, and quietly drops Brand B from the shortlist.

The runner wakes up, sees one recommendation, and buys it before coffee. Brand B's analytics record nothing. No session, no bounce, no abandoned cart. From the dashboard's point of view, the sale never existed.

That silence is the whole story. Agent-to-agent pre-purchase queries are already a conversion surface. Most brands just haven't wired it up, so it behaves like a leak instead.

The demand your analytics never sees

Zoom out from that one exchange and the volume is startling. Adobe measured traffic from AI sources to U.S. retail sites growing 393% year over year in Q1 2026, with March alone up 269%, continuing a holiday season where AI-sourced visits had surged 693%. And these visitors close: Adobe's own conversion data shows assistant-referred shoppers arriving late in the decision with the comparison work already done.

The money is no rounding error. Salesforce tracked a record $336.6 billion in global Cyber Week 2025 spend and attributed $67 billion of it to AI and agents, which influenced 20% of all orders worldwide.

Here is the uncomfortable part. Those figures count the queries that succeeded. When an agent asks about stock and gets no parseable answer, nothing enters your funnel to be counted. Human failure modes leave fingerprints: a rage-click, a bounce, an abandoned cart email trigger. Agent failure modes leave a clean absence. The demand existed, the question was asked, and your stack never registered either.

The rails are already bolted down

The transaction layer stopped being hypothetical in late 2025. Stripe and OpenAI co-developed the Agentic Commerce Protocol, an open standard that lets a ChatGPT user buy inside the chat, launching with Etsy sellers and a promise of a million-plus Shopify merchants. Rollout has been bumpier than the press releases implied: CNBC reported that only around 30 Shopify merchants were live via Instant Checkout as of February 2026, while Walmart had made roughly 200,000 products purchasable inside ChatGPT.

Google answered at NRF in January 2026 with the Universal Commerce Protocol, co-developed with Shopify, and by April Google had shipped Cart and Catalog capabilities that let agents retrieve product data and assemble multi-item carts directly.

Picture the diagram: three vertical lanes labeled discover, decide, pay. The pay lane is filling up with protocol logos. The decide lane, where stock checks, delivery promises, and returns terms live, is still mostly hand-drawn. That middle lane is the one this article is about, because checkout standards make purchasing a commodity while answering remains a competitive act.

Build the answer layer before the checkout layer

Now the architecture. Sketch it as a funnel turned on its side: agent query on the left, a thin answer gateway in the middle, and three systems of truth behind it, inventory, delivery promise, and policy. Everything else is plumbing.

Three moves matter most:

Promote answers from page furniture to services. Stock levels, delivery estimates by destination, and return terms usually exist as rendered UI, legible to humans and opaque to machines. Adobe's AI Content Visibility Checker found the average U.S. retail homepage is only 75% machine-readable, meaning a quarter of the content simply does not exist for an LLM, and category pages score worse. Run your own pages through an AI visibility checker before assuming agents can see what shoppers see.

Encode policy as data. Schema.org's Product and Offer types support shippingDetails and hasMerchantReturnPolicy, which means your delivery window and return terms can travel inside the markup an agent already parses. A schema generator gets the skeleton right in an afternoon; keeping it synchronized with your actual OMS and policy engine is the real work, and the part most retailers skip. When we audited the homepages of 54 well-known B2B SaaS vendors with our AI Visibility Checker in September 2026, half carried two or more high-value schema types and 15% had no JSON-LD at all; those are well-resourced software companies audited on the homepage alone, and a retail catalog with per-variant stock and shipping fields is a harder thing to keep marked up.

Make the delivery promise deterministic. "Arrives by Friday to 80203" is a computable claim from carrier cutoffs, warehouse location, and live inventory. If your stack can compute it at checkout, expose the same computation upstream where the agent is deciding. An assistant will always prefer the brand that commits over the brand that says "3-7 business days."

This is genuinely a build, and it touches commerce, data, and support systems at once, which is why we treat it as an agentic AI automation engagement rather than an SEO tweak.

Count the questions like you count clicks

Instrumentation is what turns this from infrastructure spend into a conversion surface you can manage. The diagram here is a plain event pipeline: agent request in, classifier tags it (intent, SKU, geography, protocol), the gateway answers or fails, and every event lands in the warehouse next to your orders table.

From that pipeline, three metrics fall out:

1. Agent Answer Rate. Of all machine queries about stock, delivery, or returns, what share got a complete, structured answer? This is your new conversion rate, one level up the funnel.
2. Answer latency. Agents rank options under time budgets. A response in 400 milliseconds and a response in 9 seconds are different products.
3. Agent-assisted revenue. Match answered queries to subsequent orders by SKU, geography, and time window. Imperfect, but it converts the whole program from faith to arithmetic.

Route these events into the same first-party data foundation that feeds your ad platforms, because agent queries are astonishingly honest intent signals: exact SKU, exact zip, exact deadline. That is richer than most clickstream data, and it sharpens the segmentation work we cover in our ecommerce personalization blueprint and feeds directly into the audience math behind our ROAS playbook for AI audience targeting.

One more integration point hides in plain sight: your support stack. Pre-purchase questions about sizing, compatibility, and returns are exactly what support automation already answers for humans. Expose that same knowledge through the agent gateway and you get double duty from one investment, a pattern we mapped in detail in our piece on AI customer insights for support automation.

When does the answer layer cost more than it returns?

The gateway is the right build for most catalogs with live stock and a delivery promise worth making. That said, it is the wrong first build in a few situations.

• Slow-moving, low-SKU catalogs. If stock rarely changes and shipping is the same everywhere, a real-time endpoint adds infrastructure while machine queries stay rare. In practice you get an answer rate with a denominator too small to trend. Start with Product and Offer markup carrying shippingDetails and hasMerchantReturnPolicy, and log agent hits in the server logs first.
• Answers fed from a stale sync. The mistake most teams make first is exposing an endpoint that reads a nightly inventory snapshot. An agent hears "in stock", the shopper lands on a sold-out page, and the brand has now failed twice. Unless the endpoint can read the OMS live, return a range with a confidence flag.
• Delivery promises only part of the catalog can keep. Dropship and made-to-order lines have no carrier cutoff to compute from. The exception is a line with a contractual handling window; commit there and label the rest as estimates, because a broken promise is worse than a vague one.
• Matched revenue that would have arrived anyway. The catch with SKU, zip and time-window matching is that it claims brand-search orders too. If answer rate rises while total revenue stays flat, the model is over-crediting; compare against a period or region where the endpoint is switched off. We would not recommend reporting matched revenue to finance before that holdout exists.

How we build the answer layer on client engagements

We treat this as an agentic AI automation build. The first week is an inventory of questions: every stock, delivery and returns query the store currently answers through rendered pages, support macros or a person, ordered by how often agents already ask it in the server logs. The gateway then ships one service at a time, availability first because it is usually the query with the most log hits, each reading from the OMS, carrier cutoffs or the policy engine directly rather than from a page. Before any service goes live, the event pipeline is in place: classifier tags, answer-or-fail status and latency landing in the warehouse beside the orders table, so answer rate has a baseline from its first day. A change to published return terms needs a human sign-off, answer quality is spot-checked against what checkout itself computes, and the client keeps the code and the documentation.

Run the 2:47 a.m. test

Back to Denver. The difference between Brand A and Brand B was never product quality, price, or brand love. It was that one of them could answer three questions, in structure, in under half a second, while nobody was watching.

So run the test on yourself. Ask an assistant whether your bestseller in a mid-range size can reach a specific zip code by Friday and what happens if it fits badly. Time the answer. Read what the agent actually retrieved. If the result would embarrass you, good; you found the leak while it is still cheap to fix, and months before your competitors think to look. The next sale is already being negotiated between two pieces of software. Make sure one of them is yours.

Sources

• Adobe: AI traffic grows but retail sites lag in AI search visibility
• Salesforce: AI and Agents Propel Cyber Week to Record $336.6B in Global Spend
• CNBC: OpenAI's first try at agentic shopping stumbled. It's trying again