Amazon vs Perplexity: why AI shopping agents just hit a platform wallAmazon vs Perplexity: why AI shopping agents just hit a platform wall
Amazon has reportedly told Perplexity to shut down an experimental AI shopping agent that was placing orders on its marketplace, drawing a very firm line under a question that has been building all year: who is actually allowed to act on behalf of a customer when an AI model starts pressing the buy button. Behind the legal language, the dispute is less about one startup and more about how far autonomous agents will be allowed to go before large platforms assert direct control over transactions and customer relationships.
What Amazon is objecting to
According to Bloomberg’s reporting, Amazon has sent a demand to Perplexity to stop an AI agent from making purchases on its storefront. The key points are straightforward even if the details are still emerging:
Perplexity has been testing an agent that can go beyond answering questions and recommendations, moving into the territory of actually placing orders for users when asked.
Some of those orders have been routed through Amazon, most likely by driving an ordinary Amazon account via automated means on behalf of the end user.
Amazon’s position is that this behaviour breaches its platform rules around automated access, scraping, and use of the site for commercial or intermediary purposes.
Amazon already has its own APIs, affiliate programmes, and advertising products for third party services that want structured access to the catalogue. The shopping agent Perplexity has been experimenting with appears to sit outside those sanctioned paths. That is the core of the conflict: not “AI is bad”, but “you are treating the public website as an automation surface we did not authorise you to use that way”.
Where AI agents fit in the ecommerce stack
A modern ecommerce stack has several layers that matter for this kind of dispute:
Content layer – product pages, descriptions, reviews, pricing, and availability data. This is what classic crawlers index and what large language models often scrape or train on.
Recommendation and search layer – the logic that tries to find the right product for a query. This can be internal (Amazon’s own ranking) or external (Perplexity generating comparison answers).
Transaction layer – the actual order placement, payment, and fulfilment flows. This is where money moves and where liability sits.
For a long time, third party tools were largely confined to the first two layers. Price comparison engines, affiliate blogs, deal trackers, and review aggregators all pulled data, linked back, and then handed users back to the merchant for checkout. Generative AI agents blur the line by trying to sit across all three layers at once. They ingest content, handle recommendation, and then attempt to execute the transaction without handing control back to the platform in the usual way.
Amazon’s reaction to Perplexity is a signal that it intends to keep the transaction layer tightly controlled. You can build search and advice experiences on top of Amazon data if you play by its rules, but an unsanctioned bot that logs in as a pseudo user and drives the checkout flow crosses a line.
Why platforms are uncomfortable with third party AI agents placing orders
There are several practical reasons a marketplace like Amazon is not keen on external agents acting as automated customers.
Liability and customer ownership
From Amazon’s perspective, the customer relationship is central. When an AI agent sits between the user and the platform, several questions arise:
Who is responsible if the agent orders the wrong item, or orders far more than the user intended.
Who handles disputes when a user claims they did not authorise a purchase that the agent performed on their behalf.
Whose brand takes the hit if the experience feels chaotic or untrustworthy.
Amazon does not want to be in the position of handling returns, chargebacks, and support for decisions made by a third party model it does not control, especially if those decisions were made using scraped or reverse engineered behaviour rather than approved APIs.
Policy compliance and abuse potential
Automated access to checkout flows makes abuse easier. A sufficiently capable agent could:
Probe pricing and stock in ways that look like scraping or competitive intelligence work rather than genuine purchases.
Drive patterns of orders and cancellations that stress fulfilment and fraud detection systems.
Obscure the origin of abusive or policy violating purchases behind an apparently legitimate front end.
Most large marketplaces already have rules that restrict automated ordering, scripted browsing, and scraping at scale. A shopping agent that operates by simulating user actions on the consumer interface rather than going through an official integration path will light up the same alarms as classic bots, even if its intention is to help legitimate customers.
Data control and competitive positioning
There is also a strategic angle. If Perplexity or another AI assistant becomes the user’s primary shopping interface, the marketplace risks being commoditised into a backend. Price, delivery time, and availability become features that the agent arbitrages across different retailers. The brand and differentiated experience of the marketplace are pushed into the background.
Amazon has no incentive to accelerate that shift. It wants to remain the primary shopping front end, not just a fulfilment layer serving API calls from AI agents that sit between it and the customer.
Perplexity’s likely view of the experiment
From Perplexity’s side, the logic for experimenting with a transactional shopping agent is simple enough:
The product is already positioned as an AI answer engine rather than a traditional search engine.
Users will increasingly expect “do this” behaviour, not just “tell me about this” answers.
The easiest way to prototype that is to drive existing consumer interfaces on major platforms, because that is where the products and fulfilment already live.
In that sense, Perplexity is pushing toward the same kind of agent vision that OpenAI, Google, and others have been talking about, where models can browse, reason, and act. The friction arises because acting in this context means driving a third party web property in a way the owner does not endorse.
Even if Perplexity applies guardrails, logging, and user confirmations, the marketplace still ends up with an opaque agent performing actions that look like they came from a normal user account. That is not a comfortable position for a platform that is responsible for fraud control and regulatory compliance around payments and consumer protection.
Why this matters beyond Amazon and Perplexity
It is tempting to read this as a one off clash between a big retailer and a smaller AI company, but the pattern will repeat across sectors.
Other marketplaces and platforms
Once one major marketplace publicly pushes back against unsanctioned AI agents making purchases, others are likely to take a similar stance. You can expect more explicit language in terms of service about:
Restrictions on automated use of consumer interfaces for transactional purposes.
Requirements to use official APIs and partner programmes for any automated purchasing or quoting.
Explicit bans on acting as an intermediary agent without a formal agreement.
That will apply not just to general retail, but also to travel sites, ticketing platforms, and subscription services. Anywhere that an AI assistant might be tempted to press the buy button on a user’s behalf is now on notice that platforms will want explicit control over that behaviour.
Regulatory context
Regulators are already looking closely at AI transparency, consent, and accountability. An AI agent ordering goods and services raises obvious questions:
Did the user give informed consent for the agent to spend their money.
Is the decision making process explainable enough to resolve disputes.
How are data protection and profiling rules applied when a model is acting across multiple services.
If incidents start to pile up where AI agents place problematic orders, marketplaces and AI providers will both be drawn into regulatory conversations. Amazon’s demand to Perplexity can be read as a preemptive move to keep that risk surface as small as possible on its side.
What a more sustainable model could look like
None of this means that autonomous shopping agents are impossible. It does mean that the spontaneous, screen-scraping, “we will just drive your checkout flow for you” approach is unlikely to survive at scale. A more durable model would include:
Formal API based integrations – agents that want to place orders use documented APIs with rate limits, authentication, and explicit permissions, not browser automation against consumer UIs.
Clear user consent flows – users know when an agent is about to commit to a purchase, which account it is using, and under what constraints.
Shared logging and dispute handling – both the AI provider and the marketplace can see what was requested, what was executed, and who authorised it.
Brand and UX agreements – marketplaces can set minimum standards for how their offers are presented inside third party agents, avoiding misleading representations.
The common theme is that the marketplace remains a full participant in the loop, rather than being treated as a passive website that an external agent can simply drive like a human user.
Takeaways for AI and platform engineering teams
For AI teams working on agents, the Perplexity and Amazon situation is another data point that the frontier of “let the model act on the web” is not just a technical problem. It is deeply entangled with platform policies, commercial incentives, and regulatory expectations.
For platform teams, it is a reminder that their consumer interfaces are likely to attract more and more agent traffic, whether they like it or not. Waiting until that traffic arrives to decide what is acceptable is not ideal. Clear statements around automated access, and dedicated integration paths for serious partners, will be needed if they want to harness AI driven demand without losing control of the customer relationship.
Either way, the message from Amazon is blunt. AI agents that handle recommendations are one thing. AI agents that quietly take over the checkout process without explicit, structured agreements are another, and large platforms are not going to ignore that boundary being crossed.
Leave a Reply Cancel reply