Amazon Gave Its AI Agent Two Speeds

The most important AI feature shipped this month is not a smarter model. It is a dropdown with two options.

On 23 September, Amazon expanded Seller Assistant from a tool that answers questions into one that runs persistent workflows: monitoring pricing, inventory, ratings and account health around the clock and acting, in Amazon's words, "within seller-defined parameters." The part worth pausing on is how those parameters work. For each workflow, the seller chooses whether the agent only recommends or is authorised to act. Actions land in an audit trail. Some can be set to require approval first.

That is it. That is the feature. And it is the clearest signal yet about how AI actually enters a business that has money on the line.

Why a boring settings field is the news

Think about how you onboard a new hire who is good but unproven. You do not hand them the company card on day one and you do not lock them in a room either. You let them draft the email and you press send. After a few weeks of watching their drafts, you stop reading every one. The trust is not a binary. It is a dial you turn, one task at a time, and you keep the receipts while you turn it.

Until now, most AI products offered the opposite: a product-wide on switch. Either the assistant was a chatbot that suggested things into the void, or it was an automation that did things and you found out afterwards. Amazon split the difference and made the split a per-workflow setting, which is how permissions have always worked in serious software and how they never quite worked in the first wave of AI tools.

The numbers Amazon put behind it explain the confidence. Seller Assistant is rolled out to more than 90% of sellers, with hundreds of thousands of active users, and Amazon says its recommendations are accepted more than 90% of the time. It also says 90% of sellers already use third-party AI tools. Read those together and the strategy is obvious: the acceptance rate is the argument for turning the dial from recommend to act. You earn the promotion with a track record, same as the new hire.

The interesting product is no longer the agent. It is the record of what it was allowed to do, and who said so.

Amazon opened the door, and kept the key

The same announcement included a Selling Partner plugin that exposes listings, metrics, inventory and sales data to outside AI tools. It launches with Amazon Quick and is in beta with Anthropic's Claude, US stores first, with limits on data access, human approval for actions, and audit trails again. Primary account holders get 12 months of Quick Plus free through 31 December.

So outside agents are welcome, through a gate Amazon built and controls. That is the shape of the whole moment. Platforms are not refusing agents. They are refusing unauthorised agents, and the authorisation is becoming the product.

The card networks arrived at the same answer

Payments got there independently, which is what makes it convincing. On 17 September, Alchemy added Mastercard Agent Pay credentials to its AgentCard: an agent pays with a one-time tokenised Mastercard credential linked to your existing card, and you or your issuer set limits by amount, merchant category and location. It supports Verifiable Intent, Mastercard's framework for proving that an agent acted within the instructions the user actually authorised.

Strip the jargon and Verifiable Intent is a receipt for permission rather than for money. Not "this card was charged 240 euros" but "this human authorised an agent to spend up to this much, at this kind of merchant, and here is the proof." It is Amazon's audit trail wearing a card network's uniform.

Europe got the plumbing to match. On 14 September, Worldline launched a payment handler for the Universal Commerce Protocol on its Global Collect platform, among the first in Europe, covering Visa, Mastercard, Amex, Apple Pay, Google Pay and SEPA bank transfers. Agentic checkout had mostly been running on rails that suit American shoppers. SEPA support is what lets a European shopper's agent pay the way Europeans actually pay.

The dashboard era is ending

Then there is Noibu, which launched six AI agents for ecommerce stores covering conversion optimisation, A/B testing, ADA compliance, bug resolution, site speed and return on ad spend. Each one runs the same loop: analyse the store data, diagnose the problem, draft a fix, deploy it once a human approves, then measure the result.

Their own framing is the sharpest line in the whole story. For a decade, they said, tools like this "ended at a dashboard." A dashboard is a very expensive way of telling someone they have a problem. What closes the gap between knowing and fixing has always been a person with the credentials and the nerve to push the change. Noibu replaced most of that person with software and kept exactly one human step: the approval.

What to take from it

Three different industries, three weeks, one design. Marketplace, card network, storefront tooling. None of them shipped "trust the AI." All of them shipped graduated authority plus a written record.

If you are evaluating any agentic tool right now, that is your checklist, and it is short. Can I set what this thing is allowed to do, separately for each job it does? Can I start it in suggest-only mode and see its hit rate before I promote it? When it acts, is there a record of what it did and which human's authority it was acting under?

If the answer to all three is yes, you are looking at a product built by people who expect to be audited. If a vendor can only offer you on or off, they have not thought about the day it gets something wrong. And it will, eventually, on a Tuesday, at scale. The approve button is not friction. It is the only part of this that is load-bearing.