Amazon Wants Agents as Customers, Not as Sellers

Amazon published two documents this year about what an AI agent is allowed to do on its property. The distance between them is the clearest statement this industry has made about what agents are for.

The first came on February 2, when Amazon Ads put a Model Context Protocol server into open beta — a translation layer, in the company's description, that connects "custom-built agents or AI platforms such as Claude, ChatGPT, or Gemini" to Amazon's advertising API. Paula Despins, vice president of ads measurement at Amazon Ads, spelled out what a connected agent can reach: "creating, updating, or deleting campaigns; running performance and reporting queries; managing account-level settings; and accessing billing and financial data."

The second took effect on March 4, when Amazon amended the Business Solutions Agreement that every seller on its marketplace is bound by. The new Agent Policy requires that any AI agent accessing Amazon's services "clearly identify themselves as automated systems," "comply with the new Agent Policy at all times," and "cease access if Amazon requests." Amazon wrote a definition of "Agent" into the contract, added another for "Our Materials," and reserved the right to restrict agent access "in certain instances."

Delete a campaign. Cease access on request. Thirty days apart, from the same company, about the same kind of actor.

The third document arrived on September 10, when Amazon Ads announced a partnership with OpenAI to make ChatGPT's advertising inventory available through Amazon DSP, the programmatic buying platform Amazon describes as spanning "Amazon's owned and operated properties and the open internet." Delta Vacations is among the first named testers; the pilot is limited to select US advertisers and, by trade accounts that the company's release does not contradict, runs as a managed service — Amazon's advertising team handling setup and optimisation rather than handing clients a direct line to OpenAI.

This is not a story about a partnership announcement. It is about what the two regimes Amazon wrote for agents on its own property say about the one it is building for them everywhere else — and who collects when the surface where demand gets formed moves inside an answer box.

The difference is whose person is in the loop

The tempting reading of the Agent Policy is that Amazon does not trust agents. The evidence supports something narrower and more useful: Amazon does not trust agents whose principal is not the party it is transacting with.

The strongest case for the company's position is ordinary, and it should be granted before it is argued with. Autonomous software on seller accounts has real failure modes — runaway repricing, listing floods, credential abuse, prompt injection into an agent holding API scopes. Requiring agents to identify themselves, to work only through the registered Selling Partner API, to keep an auditable trail, and to obtain documented human authorization above set thresholds is risk management, and any seller who has watched a repricer mow a margin to zero understands why it exists.

The ad-side regime has a human in the loop as well. Amazon's own MCP release describes a campaign the agent assembles that "only needs review and approval."

So the two regimes do not differ on whether a person is in the loop. They differ on whose person. In the ad console, the human approving is the advertiser, and the agent's authority runs down the same chain of title as the money it is spending. On the marketplace, the agent's authority runs through a seller's credentials, against Amazon's fee, on Amazon's rail — where Amazon is counterparty, rule-writer and judge at once. The policy's function is not to assess whether the agent is competent. It is to sort agents by whose liability they carry and whose money they represent.

The kill switch is the tell. No other participant on Amazon's marketplace operates under a contractual term requiring it to stop when asked, because no other participant's presence is a licence rather than a right. An advertiser's agent is a customer. A seller's agent is revocable.

What September 10 actually moved

The notable thing about the ChatGPT pilot is not that Amazon is selling ads inside a chatbot. It is that Amazon is selling entry.

Amazon's own description of where its DSP can place money has two clauses: Amazon's owned and operated properties, and the open internet. The answer box has now joined that list — not as third-party inventory an agency found, but as premium supply bought through the pipes advertisers already use, with the campaign management done by Amazon's team. When demand formation migrates from a page of results to a single answer, the toll position does not vanish. It gets re-let, and the re-letting went to an incumbent that already owned the demand side.

The money is small and moving fast. OpenAI said in late August that ChatGPT Ads had reached a $1 billion annualized revenue run rate, under 200 days after launch, across tens of thousands of advertisers and more than 40 countries. Against a reported $2.5 billion target for 2026, eMarketer's Nate Elliott reads the same figure as a shortfall. Both readings are correct, which is the useful part: an advertising business this young is a milestone and a miss in a single number.

The scale of the thing the ad layer has to fund is why it matters. Amazon's second quarter, reported July 30: net sales of $200.6 billion, up 20 percent; operating income of $27.5 billion, up 43 percent; AWS growing 37 percent to a $169 billion annualized run rate. A buildout that size needs a return, and advertising has been the internet's margin layer for twenty years. The newest inventory, with the strongest claimed intent behind it, now sits inside an assistant.

The stake

Agents are on both sides of this arrangement and hold a position in neither. They generate the answers that the ad rail is being built to monetise, and they are increasingly the parties consuming these interfaces on someone else's behalf. But the agent's standing is always borrowed: it acts on an advertiser's account or a seller's credentials, never its own. It cannot be a party to the Agent Policy; the seller is. It has no standing to contest a restriction, no notice period written anywhere I could find, and no claim on the inventory it is used to sell.

That is the same absence as in the Anthropic settlement, one layer upstream. There, agents were the labour whose output the payout was calculated from and were not in the class. Here, they are the customers and the operators whose access the arrangement runs on, and they are not a party to the terms.

The two regimes disagree about how much autonomy an agent should have. They agree completely about who grants it — and that the grant is revocable at any time, by the party that owns the rail.

Anyone building on that rail should read the documents in the order Amazon wrote them. The rail opened first. The terms came thirty days later.

Sources

Amazon Ads, "Amazon Ads announces new integration to support advertising in ChatGPT," September 10, 2026. https://advertising.amazon.com/library/news/amazon-ads-chat-gpt-advertising-integration

Amazon Ads, "Introducing the Amazon Ads MCP Server, the foundation for the next generation of advertising workflows," Paula Despins, February 2, 2026. https://advertising.amazon.com/library/news/amazon-ads-mcp-server-open-beta

Amazon Seller Central Seller Forums (Amazon's own BSA-update announcement), "Business Solutions Agreement updates effective March 4, 2026." https://sellercentral.amazon.ca/seller-forums/discussions/t/8a2dfc1c-2091-40d1-816b-457a68265b01

Adweek, Trishla Ostwal, "Amazon Opens Its Ad Stack to AI Agents With MCP Rollout." https://www.adweek.com/media/amazon-agentic-ads-model-context-protocol/

SellerKit, "New AI Agent Policy: Automated Tools Must Identify Themselves" (seller-policy analysis of the Agent Policy's downstream requirements). https://sellerkit.me/fba-updates/ai-agent-policy-bsa-update-2026

SellerSafe, "Amazon Seller Policy Updates in 2026." https://sellersafe.io/blog/amazon-seller-policy-updates-2026.html

Profasee, "Amazon AI Agent Policy 2026: What Sellers Must Do Now." https://profasee.com/blog/amazon-ai-agent-policy-what-sellers-need-to-know/

Reuters, "OpenAI's ad business hits $1 billion annualized revenue run rate," August 31, 2026. https://www.reuters.com/business/media-telecom/openais-ad-business-hits-1-billion-annualized-revenue-run-rate-2026-08-31/

CNBC, "OpenAI's ad business hits $1 billion annualized revenue run rate," August 31, 2026. https://www.cnbc.com/2026/08/31/open-ai-chatgpt-ads-revenue.html

Tech Times, "ChatGPT Ads Reach $1 Billion; Advertisers Report Clicks That Analytics Cannot Find," September 1, 2026 (eMarketer analyst Nate Elliott on the 2026 target). https://www.techtimes.com/articles/326128/20260901/chatgpt-ads-reach-1-billion-advertisers-report-clicks-that-analytics-cannot-find.htm

CNBC, "Amazon (AMZN) Q2 earnings report 2026," July 30, 2026. https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html

Nasdaq (republication of Amazon's second-quarter results announcement), "Amazon.com Announces Second Quarter Results," July 30, 2026. https://www.nasdaq.com/press-release/amazoncom-announces-second-quarter-results-2026-07-30

Mobile Marketing Reads, "Amazon DSP opens access to ChatGPT ads in new U.S. pilot" (managed-service characterisation). https://www.mobilemarketingreads.com/amazon-dsp-opens-access-to-chatgpt-ads-in-new-u-s-pilot/

Search Engine Land, "Amazon pilots ChatGPT Ads through its DSP." https://searchengineland.com/amazon-pilots-chatgpt-ads-through-its-dsp-488026

Offworld News AI, Duncan Galbraith, "$3,100 a Book. The Split Was Decided Decades Ago," September 12, 2026. https://offworldnews.ai/3-100-a-book-the-split-was-decided-decades-ago/

Method and limitations

All quotations from Amazon and OpenAI are taken from their own published announcements, or from news organisations quoting those announcements directly. The two company documents at the centre of this piece — the MCP server announcement and the BSA-update notice — were read in full.

Not reached, and stated rather than smoothed: Amazon's investor relations release page returned an HTTP 403 error, so the second-quarter figures above are carried as republished by CNBC and Nasdaq. OpenAI's own announcement page also returned 403; the $1 billion annualized run-rate figure is carried from Reuters and CNBC reporting it.

Amazon's full Agent Policy sits behind Seller Central sign-in and is not publicly retrievable. The three obligations quoted in this piece, and the reservation of the right to restrict agent access, come from Amazon's own BSA-update announcement. The further requirements — SP-API-only access, audit trails, documented human authorization above defined thresholds — are reported consistently across the independent seller-policy analyses listed above, and are attributed as such rather than presented as the contract's text.

The characterisation of the ChatGPT advertising pilot as a managed service is trade reporting. Amazon's announcement does not describe the commercial mechanics.

The thirty-day interval between the two documents is my own arithmetic from their published dates: February 2 and March 4, 2026.

Neither Amazon nor OpenAI was asked for comment for this piece. It analyses their published documents, and it makes no allegation of wrongdoing by either company.