Agentic commerce: what AI agents buy (and what they don't)

A person in a beige knit sweater checks a phone while leaning on a desk in a brick office, next to a small white delivery robot, a stack of cardboard parcels and a monstera plant, in warm afternoon light
Written by
Eduardo Garcia Garzón Follow

Head of AI en Shakers. Diseña e implementa sistemas LLM para matching de talento a escala en entornos enterprise. Especialista en IA en producción, evaluación de modelos y aplicaciones de alto riesgo, con experiencia liderando programas de cumplimiento del AI Act.

• ...

Everything cited here takes a minute to check. The DoorDash figure is in its 30 October 2026 earnings call, with the 3-million-deliveries-a-day denominator. The bot ratios come from TollBit's State of the Bots report, which declares its methodology and warns it counts identified bots only. TIME and Perplexity sit in two Digiday pieces from July and August 2026. And the ACP reads on a single page, its official site. One figure circulating in trade press appears in no primary source, so we leave it out.

The Spain figure is our own analysis, and it counts what we could not find counted. We measured the active postings mentioning "agentic commerce" (16, of which 11 are distinct by company and title), checked that no posting uses a native-language term for the channel yet (0), and sized the whole market the same day: 44,803 active tech postings in Spain (Shakers market analysis, 3 November 2026).

Agentic commerce is the channel in which AI agents act as buyers or shopping intermediaries on an ecommerce site: they read the catalogue, compare options, start the checkout and, through the Agentic Commerce Protocol that OpenAI and Stripe released as an open standard, can complete the payment inside the conversation. The first data on the channel already exist, and they are uncomfortable reading. Agents visit a lot. They buy very little. And meanwhile, advertising built specifically for them is already running. Spain, the first market where hiring for the channel has been quantified, already has teams on the way: 16 active job postings mention "agentic commerce", according to the Shakers market analysis of 3 November 2026. This article puts those numbers in order, explains what they mean, and what an ecommerce team should actually do now.

Agentic commerce is already measured, and it does not yet buy at volume. DoorDash has received "several hundred" orders from external agents against roughly 3 million deliveries a day (earnings call, 30 October 2026). European publishers get 179 AI-bot visits for every human referral, and the ratio worsened from 150 to 227 per 1 between Q1 and Q2 2026 (TollBit, State of the Bots). TIME already serves ads only agents read. In Spain, 16 active job postings ask for "agentic commerce" out of 44,803 on the tech market (Shakers market analysis, 3 November 2026), and no posting has adopted a native-language term for the channel yet.

What agentic commerce is

A shopping agent is a program that acts on a person's behalf when they buy online. It is not a search box that suggests links, and it is not a support chatbot. The agent receives a goal ("renew my cat food every month", "find a kids' bike under £200"), queries catalogues, compares, applies criteria and executes the transaction. The person supplies the goal and the card. The agent supplies the visits.

The missing piece that makes this work at scale is a checkout standard, and one exists. It is the Agentic Commerce Protocol (ACP), developed jointly by OpenAI and Stripe and published as an open standard under the Apache 2.0 licence. Its official description: "An open standard for programmatic commerce flows between buyers, AI agents, and businesses". The protocol lets any compatible agent start a checkout securely with the merchant, who remains the merchant of record: the agent never sees the card details, because the protocol passes them from the buyer to the business, and the customer still belongs to the ecommerce business. OpenAI launched it in ChatGPT with Instant Checkout in September 2025, starting with Etsy sellers and later Shopify merchants; Stripe enables it for its merchants with a single line of code.

The vocabulary is still unsettled outside the vendors. IBM, Salesforce and Braze use the term in its standardised sense, agents that buy, with no trademark collision forcing anyone to hedge.

The agents already visit, and ads are already running for them

Before debating whether agents buy, there is the visit data. TollBit's State of the Bots report (2026 Q1 and Q2 edition, "The Bad Bots") measures AI-bot traffic across its publishers' sites, and its European finding is blunt: it takes 179 AI-bot visits to produce a single human referral from AI applications, an exchange more than three times worse than North American sites. And the exchange keeps deteriorating quarter on quarter: from 150 per 1 in Q1 2026 to 227 per 1 in Q2. The report itself warns that its figures count identified bots only, so the true scale of scraping is higher.

For UK readers the report carries a finding of its own. In TollBit's audit of the top 20 UK sites, 19 of the 20 could be scraped by at least one of the 14 scrapers tested, and paywalled content was not necessarily protected: some tools retrieved full articles. The European load figures frame the rest. Median AI scrapes per site run four times higher in Europe than in North America, European publishers see about one AI scrape for every 33 human visits, and AI bots produce only 0.05% of the total human referrals on those sites. Traffic without a return. That is the pattern.

The advertising market's answer has already arrived, and it is the clearest sign that bot audiences exist as a measurable segment. TIME became in July 2026 the first publisher to serve ads designed for AI agents: markdown-formatted pages that crawlers read, where the ads appear, and that no human ever sees rendered in a browser, using technology from the startup Mobian. It was not a quiet launch. Perplexity blocked those ads in August and called them deceptive. The dispute is still open, and it does not change the underlying fact: an entire industry is building ad inventory for audiences made of software.

And they barely buy: the first numbers

If the visits are measured, the sales are starting to be counted too, and the first nominal figure comes from DoorDash's Q3 earnings call on 30 October 2026. CEO Tony Xu sized the orders arriving from third-party agents, including Meta's Muse agent, with a sentence worth quoting in full: "we're seeing maybe several hundred orders in total from some of these third-party agents". The denominator matters, and it is not small: DoorDash completes an average of 3 million deliveries a day according to its 2025 report, the figure picked up in TechCrunch's coverage. Hundreds of orders, cumulative, against 3 million deliveries a day. The distance shows without converting units, and that is the good news: hype and till can be measured separately.

"You never want to fight gravity. But at the same time, you don't want to be delusional."

Tony Xu, CEO of DoorDash, Q3 2026 earnings call

Loosely translated, Xu's line is the thesis of this piece: do not fight the gravity of the market, but do not credit it with volume it does not have. Xu did not say the channel is irrelevant. He said agent orders count in the hundreds, not the millions, and that projection-induced illusion is the real risk of the moment.

The channel's citable numbers, with their denominators

The six numbers a team can quote in a meeting today, each with its source and period.

Data point Value Source and period
Orders from third-party agents at DoorDash "Several hundred" in total, against ~3 million deliveries a day Q3 earnings call, 30/10/2026
AI-bot visits per human referral (Europe) 179 to 1, over 3x worse than North America TollBit State of the Bots, H1 2026
Exchange trend in Europe 150 per 1 (Q1 2026) to 227 per 1 (Q2 2026) Same report
UK scrape exposure 19 of the top 20 UK sites scrapable by at least one of 14 scrapers TollBit scraper audit, 2026
Ads only agents read TIME, first publisher; Perplexity blocked them as deceptive Digiday, July and August 2026
Agentic checkout standard ACP, Apache 2.0 licence; ChatGPT the first platform OpenAI and Stripe, September 2025

One row is missing from that table, and it is labour demand: we could not find any publication that has measured hiring for this channel in any market. The section below closes that gap with the Spanish reading, the first quantified one.

What the job market asks: the teams are already being hired

Here is the data we could not find in any publication. In the Shakers market analysis (cut of 3 November 2026), 16 active job postings in Spain mention "agentic commerce", which is 11 distinct postings by company and title out of a Spanish tech market of 44,803 active postings: 0,036% of the total. Nine carry the term in the job title itself: five at Accenture (data and AI architects for agentic commerce, consultants and practice leads), three at zooplus (senior AI engineers for agentic commerce) and one at E-Frontiers (a hybrid senior AI engineer in Madrid). The rest are product, fullstack engineering and growth profiles at Capitole, pasiona, Google, Alan, commercetools and Jobgether that list agentic commerce among their responsibilities. Demand clusters geographically: Madrid (7 postings) and Barcelona (6).

Company Postings with "agentic commerce" in the title Profile sought
Accenture 5 Data & AI Architect Agentic Commerce, Agentic Commerce & AI Lead, Agentic Commerce Consultant
zooplus 3 Senior AI Engineer - Agentic Commerce
E-Frontiers 1 Senior AI Engineer - Agentic Commerce (hybrid, Madrid)

The reading for an ecommerce team is twofold. First: the channel is today a small engineering and product project, not an operations build-out. Second: the profile that builds it is a known one, the profile described in what is an AI builder (the English version of the analysis), someone who orchestrates agents on top of existing systems rather than a invented job title for the launch deck. That profile is available on AI Engineer projects, with capacity to start in days.

What to do now, and what not to

With the data above on the table, the practical agenda sorts itself. Cheap first, expensive later.

Agentic-ready checklist for an ecommerce

  • Machine-readable catalogue: structured feeds with price, stock and availability kept current. The agent decides on what it reads; a catalogue with stale prices excludes itself from the comparison.
  • ACP-compatible checkout: if the business takes payments through Stripe, its own announcement describes the switch-on with one line of code. The business remains merchant of record.
  • Measure the bot traffic: separate AI user agents from human visits in analytics. It is the segment TollBit measures at scale across its publishers, and that most dashboards still blend with everything else.
  • Treat agent-facing ad inventory as an experiment: TIME already serves ads only crawlers read and Perplexity blocked them. The format is in public dispute; entering now is a test, not a plan.

Source: TollBit State of the Bots Q1-Q2 2026; Digiday July and August 2026; Agentic Commerce Protocol documentation.

The do-not list is shorter. Do not rebuild the whole operation for a channel that, in the only nominal figure published, moves hundreds of orders against the millions a day of the same company's traditional channel. Do not promise internal growth curves for agent-driven sales, because the citable data does not support them yet. Do not mistake bot traffic for audience: 179 visits per referral is infrastructure cost until the protocol and the rules of the game turn it into something else. Xu's advice above applies in full to this decision. The channel moves. The volume arrives when it arrives.

For teams already running agents in production, the natural next step is the one described in the AI builder analysis: defining which decisions the software delegates and which it does not, with the same control criteria any new channel demands.

Want your catalogue and checkout ready for the agents without rebuilding your operation? Tell us which platform you run and you get access to the team that connects it.

Talk to an expert at Shakers

Frequently asked questions about agentic commerce

What is agentic commerce?

It is the channel in which AI agents act as buyers or shopping intermediaries on an ecommerce site: they read catalogues, compare and can complete payment through the Agentic Commerce Protocol from OpenAI and Stripe. The first data show heavy traffic and light purchasing: 179 bot visits per human referral in Europe (TollBit, H1 2026) and "several hundred" agent orders at DoorDash against 3 million daily deliveries.

Are people actually buying through AI agents yet?

Agents already start buying, but the volume is marginal. On the 30 October 2026 earnings call, DoorDash counted "several hundred" total orders from external agents such as Meta's Muse, against an average of 3 million daily deliveries on its traditional channel. CEO Tony Xu warned against being "delusional" about the channel. It is a measurable transition, not an installed volume.

What is the Agentic Commerce Protocol (ACP)?

It is an open standard developed by OpenAI and Stripe, published under the Apache 2.0 licence, that defines how an AI agent starts a checkout with a business securely. The merchant stays the merchant of record, and the protocol passes payment credentials without exposing them. OpenAI launched it in ChatGPT with Instant Checkout in September 2025, and Stripe enables it for its merchants.

What does agentic-ready mean for an ecommerce?

Four concrete pieces: a machine-readable catalogue with current prices and stock, an ACP-compatible checkout, analytics that separate AI bot traffic from human traffic, and a documented position on agent-facing ad inventory, which is in public dispute between publishers like TIME and platforms like Perplexity. None of this requires rebuilding the operation.

What does the Spanish job market ask for agentic commerce?

According to the Shakers market analysis of 3 November 2026, 16 active postings in Spain mention "agentic commerce" out of 44,803 active tech postings (11 distinct by company and title). Nine carry the term in the title itself: five at Accenture, three at zooplus and one at E-Frontiers, concentrated in Madrid and Barcelona. It is the first quantified reading of hiring for this channel in any market.

Is it worth investing now in ads for AI agents?

It is an experiment, not a plan. TIME has served ads only agents read since July 2026 and Perplexity blocked them in August calling them deceptive, so the format has no stable rules. With 179 bot visits per human referral in Europe, the advertising return of the segment is unproven. Cheap things first: a readable catalogue and a compatible checkout.

Sources and method

Everything cited here takes a minute to check. The DoorDash figure is in its 30 October 2026 earnings call, which TechCrunch reports with the 3-million-deliveries-a-day denominator. The bot ratios come from TollBit's State of the Bots report (2026 Q1 and Q2), which declares its methodology and warns it counts identified bots only. TIME and Perplexity are covered by Digiday, in July and August 2026. And the ACP reads on a single page, its official site. One figure circulating in trade press, the "100,000 advertisers" for OpenAI's sponsored agents, appears in no primary source, so we leave it out.

The Spain figure is our own analysis, and it counts what we could not find counted. We measured the active postings mentioning "agentic commerce" (16, of which 11 are distinct), checked that no posting uses a native-language term for the channel yet (0), and sized the whole market the same day: 44,803 active tech postings in Spain (Shakers market analysis, 3 November 2026; captures timestamped, reproducible method).