Agentic commerce is a new customer, not a new channel

The work to be ready for the agent customer sits in product, infrastructure, and commercial strategy. It is upstream of the transaction. Catalogues, structured data, machine-readable terms, inventory truth. The slow, unglamorous stuff that decides whether an AI agent can even see you, let alone buy from you.

What changed

Three things happened in the last twelve months that decide the next decade of commerce.

1. Google launched the Universal Commerce Protocol, co-developed with Shopify, Etsy, Wayfair, Target, and Walmart, and endorsed by more than twenty payments and retail partners.
2. Stripe announced 288 agentic commerce products at Sessions 2026, including the Agentic Commerce Suite, a Link wallet now extended to 250 million users with agent-ready virtual cards.
3. The Machine Payments Protocol: Adobe Analytics measured AI-driven traffic to US retail sites up 693 percent year on year across the 2025 holiday season. More importantly, AI-referred visits now convert better than non-AI traffic.

Four signals from the last twelve months: Live infrastructure

Two things are worth noticing across these announcements. The protocol race is converging, not fragmenting:

• Stripe co-authored OpenAI’s ACP and is live in Google’s UCP.
• Payment networks are endorsing both.

The infrastructure layer is settling faster than anyone expected, which means the differentiation will not come from which protocol you support. It will come from how legible your business is once you support them.

Both UCP and ACP keep the retailer as the merchant of record. This means the customer data, the dispute, the loyalty signal, and the refund liability still sit with you. An agent is a buyer and not an intermediary trying to disintermediate you. This is good news, but only if you are set up to act on it.

The infrastructure is being built by the companies that will own the rails for the next decade of commerce.

Why the "new channel" breaks

A channel is a place where you meet your customer, and it adds a surface. You build a presence on it, measure attribution and report performance. The customer is the same customer, and the job is to win them in a new location.

An agent is not a location: an agent is a buyer with completely different properties. It does not browse. It doesn’t remember your brand campaign. It doesn’t see your homepage and doesn’t give you the benefit of the doubt on a slow page load or a missing returns policy. It reasons over structured data, evaluates across sources you do not control, and acts within delegated permissions from the human who sent it.

Picture what that buyer actually does: A user asks their agent for a pair of running shoes under $120, returnable within thirty days, in stock for next-day delivery. The agent pulls structured product data from a dozen retailers in under a second. It rejects four because shipping terms are buried in PDFs the agent cannot parse. It deprioritises three because availability is marked as "usually in stock" rather than a number. It scores the remaining five on price, returns clarity, and delivery confidence, and presents two to the user. The user picks one. The other ten retailers never appeared. They were not outbid on price or beaten on brand. They were filtered out for being illegible.

If you sell to humans, you optimize for attention, persuasion, and conversion. If you sell to agents, you optimize for discoverability, evaluability, and execution reliability. Different jobs. Different talents. Different metrics. Different parts of the org.

The companies framing this as a channel are putting agentic commerce within their digital marketing function. The companies treating it as a customer are putting it inside product and infrastructure. The first group is optimising the funnel. The second group is building for someone who is not in the funnel at all.

The proof is in what companies are hiring for

We pulled 1,291 job postings across 197 companies and 43 countries in May 2026 to see how the work is actually getting organised.

Ulta Beauty has a Senior Manager Agentic Commerce GTM who is accountable for revenue growth, margin performance, and commercial outcomes across AI-driven shopping platforms. That is a P&L, not a channel team. Apple has ten dedicated agentic commerce roles, framed around how Apple participates in a world where AI agents mediate commerce. Not how Apple sells through AI agents. Walmart has a Distinguished Data Scientist for Agent-Led Engagement, building models for when, how, and why to engage. With whom? Not the customer. The customer’s agent. JPMorgan Chase has a VP-level role for Generative Engine Optimisation as a deposit acquisition strategy, not a marketing tactic.

Hiring volume by sector, May 2026. Jobs are organised around the customer, not the channel.

These are not channel teams reporting into digital marketing. They are product, infrastructure, and commercial strategy functions. The companies that have decided already are organised this way. The ones still framing it as a channel are eighteen months away from realising they have built the wrong thing.

What this looks like in practice

We use a five-pillar model for what a retailer or brand needs to be ready for the agent customer. The pillars are Findable, Trustable, Transactable, Answerable, and Continuous. They map to the customer journey for an agent, the same way a funnel maps to a human journey.

• Findable: This is more about giving an agent a defensible reason to pick you over fourteen identical alternatives in a structured comparison, rather than SEO. The retailer with the cleanest product feed, the most accurate stock signals, and the most machine-readable terms wins the slot. Think: more citations, less keyword stuffing.
• Trustable: This layer addresses consent and authorization. The agent acts on behalf of a human with delegated permissions: spend ceilings, category limits, and preferred merchants. If a user has told their agent "no purchases over £200 without confirmation", your system needs to know that, respect it, and not treat the agent like a logged-in user. Stripe, Visa, and Plaid are building the rails for retailers to build on top of.
• Transactable: This is the execution layer with tokenised credentials, agent-initiated payments, real-time inventory and pricing. An agent reaches your checkout, finds inventory is stale by ten minutes, and the transaction fails. Here, the agent doesn’t retry - instead it buys from the competitor whose inventory was accurate. If your transaction layer is not agent-native, you get filtered out before you are even compared.
• Answerable: How you handle the questions an agent asks before it commits. This could include: return windows, delivery terms, warranty handling. A returns policy written in plain English on a help page might be invisible to an agent. The same policy in structured data, with concrete day counts and category exceptions, is readable. Ambiguity is a rejection signal. If the agent cannot resolve a question in machine-readable form, it picks someone who can.
• Continuous: This is the post-purchase relationship. The agent monitors tracking, delays, return windows, and resolves issues within scope. When a delivery slips, the agent files the complaint and requests the refund without the human ever opening your app. Loyalty no longer happens in your app. It happens in the agent’s memory of you.

These pillars all sit inside product, infrastructure, and commercial strategy.

What to do next

A few things matter in the next ninety days, in order:

1) Name the customer: If your org chart does not have a function accountable for the agent customer, you do not have a strategy. This is an organisational decision before it is a technical one. Pick the person, give them a P&L, put them in product or commercial strategy.

2) Audit your legibility: Walk through your own product, returns, and delivery information the way an agent would. How much of it is in structured data versus prose, images, or PDFs? That ratio is your starting score. Everything illegible to an agent is invisible to one.

3) Pick a pillar and ship in six weeks: Run a six-week sprint on Findable, then a six-week sprint on Transactable and so on.

The retailers who got mobile right in 2010 were not the ones with the best apps. They were the ones who understood early that the customer's behaviour had changed and reorganised around it. The agentic customer is the same kind of moment, but the window is shorter and the filtering is harsher. You won’t lose to a competitor with a better brand. You will lose to one whose data an agent could read. If you take one thing from this piece: name the customer, then go and look at your own business the way an agent would.

In the next blog post we’ll explain all the different types of protocols for building an agentic customer.