On January 20, 2026, eBay quietly rewrote its user agreement. The change looked small on paper: a tighter anti-scraping clause, now explicitly covering AI. In practice, it was eBay drawing a hard line in the middle of the biggest shift in online commerce since the checkout button. Starting February 20, 2026, eBay prohibits “buy-for-me” agents, LLM-driven bots, or any end-to-end flow that places orders without human review, unless the operator has eBay’s prior permission. The company had already blocked AI crawlers in its robots.txt file the previous fall. This update made the ban contractual, not just technical.
If you run an auction platform, this is not a story to skim past. It is the first major test case of a question every marketplace with timed, structured bidding is going to face this year: what happens when the bidder on the other side of the screen isn’t a person?
What eBay actually did, and why
eBay’s new terms target three overlapping categories: third-party chatbots and shopping agents acting on a buyer’s behalf, LLM-driven bots that scrape or interact with listings, and any automated flow that completes a purchase without a human reviewing the final decision. Any AI tool that wants to operate on the platform now needs eBay’s explicit sign-off first.
The reasoning eBay has given, through its policy update and public statements from its product leadership, comes down to three things: bid integrity, accountability, and preserving a human-controlled transaction. An auction is a trust mechanism. It only works if all the bidders play by the same rules, at the same speed, with the same information. In eBay’s own framing, unsanctioned agents could snipe bids unfairly, scrape pricing data at a scale no human competitor could match, and increase disputes and fraud when a purchase happens without anyone actually deciding to make it. In the same update, eBay is also rewriting its arbitration and class-action language, which tells you that the legal team is thinking about liability and not only product experience. Who’s liable when an autonomous agent buys the wrong item, or bids above a limit nobody set: the platform, the agent’s developer, or the user who deployed it? Nobody has a clean answer right now, and eBay decided the safest move was to shut the door until one exists.
Crucially, eBay has not called this an anti-AI stance. Executives have talked openly about agentic commerce as the near-term future of the platform, including a shift where agents negotiate, verify provenance, and execute cross-border transactions end to end. The message is closer to “not like this, not yet, not without permission” than “never.”
The bigger question this raises
Here’s the thing eBay’s policy exposes, whether or not you sell collectibles online: auctions are the single best-suited commerce format for AI agents to dominate and also the format where that dominance costs the most.
Think about what an auction actually is. It’s a structured, machine-readable signal (current price, time remaining, bid increment) tied to a clear, binary win condition, updated in real time. That is close to a perfect specification for a piece of software. A human bidder has to notice the listing, do the math, decide their ceiling, and physically place a bid, often against a countdown clock designed to create urgency and a little bit of adrenaline. An AI agent does none of that under pressure. It doesn’t get emotionally invested at minute 58 of a 60-minute auction. It just executes: monitor the price, compare it to a pre-set limit, bid at the optimal moment, repeat across a thousand listings simultaneously if needed.
That’s the technical problem. The harder problem is what it does to the format itself. Three things are at stake:
Fair competition. If some bidders can deploy software that reacts in milliseconds and never sleeps, and others are refreshing a browser tab on their phone during a lunch break, the contest stops being fair the moment agents outnumber humans in a given category. Sniping already causes friction on platforms like eBay. A fleet of tireless, emotionless agents sniping every auction in a category is sniping at industrial scale.
Price discovery. Auctions are supposed to reveal what something is actually worth, based on what real buyers are willing to pay in the moment. If AI agents are bidding based on scraped comparable prices and a stop-loss rule, the “price discovery” becomes an artifact of whatever pricing data the agents were fed, not a genuine read of demand. That can just as easily suppress prices (agents refusing to bid past a hard-coded fair-value ceiling) as inflate them (agents in an unintentional bidding war against each other).
The emotional and social experience. This is the part platforms tend to underweight. People don’t just want the item; they want the win. Auctions are entertainment as much as they’re transactions. eBay’s live auction and collectibles categories exist because bidding is fun, tense, and a little addictive in a good way. Hand that experience to software running on both sides, and you haven’t just changed the mechanics; you’ve hollowed out the reason a lot of buyers showed up in the first place.
The scale of what’s coming
This isn’t a hypothetical arms race. It’s already underway, and the numbers are moving fast enough that “wait and see” is not a neutral choice; it’s a decision to fall behind.
Estimates of the size of the agentic commerce market vary wildly depending on what each analyst counts as an “agentic” transaction, but the direction is the same everywhere you look. Grand View Research put the global agentic commerce market at roughly 7.7 billion dollars in 2026, projected to grow past 65 billion dollars by 2033. On the retail and e-commerce side specifically, Mordor Intelligence estimates the market at over 60 billion dollars in 2026, expanding at close to 30 percent per year. Gartner’s B2B forecast is the most aggressive: it expects 90 percent of business-to-business purchasing to run through AI agents by 2028, moving more than 15 trillion dollars through machine-to-machine exchanges.
Consumer behavior is already shifting too. During Cyber Week 2025, Salesforce identified AI agents impacting 67 billion dollars in global sales, or approximately 1 in 5 orders worldwide. AI-referred traffic to U.S. retail sites increased by nearly 400% year over year in Q1 2026, and by more than 1,300% since the company began tracking the channel in late 2024, Adobe said. Orders that came through AI-powered search on Shopify were almost 13x the volume from the previous year. eMarketer’s more conservative estimate still sees AI-platform checkout at nearly four times what it was the year before, at $20.5 billion in US retail for 2026.
Meanwhile, OpenAI, Google, and Amazon are all scrambling to launch general-purpose shopping agents that can discover, compare, and complete purchases without any human clicking “buy.” None of these companies are building agents to browse politely. They’re building agents to win, and an auction, with its clean signals and clear win condition, is exactly the kind of environment where “winning” is easiest to automate.
The three decisions every auction platform owner has to make
eBay chose to block first and figure out permissioned access later. That’s one legitimate path, but it’s not the only one, and it won’t be right for every platform. There are really three separate decisions hiding inside “should we let AI agents bid,” and they deserve to be made separately rather than as one blanket policy.
This is the defensive baseline, and honestly, it should happen regardless of your longer-term stance. Bot detection for auctions isn’t new (sniping tools and scripted bidding have existed for years), but general-purpose LLM agents are harder to fingerprint than old-school bots because they can browse and interact more like a human, including solving CAPTCHAs and mimicking natural pacing. You’ll need behavioral signals beyond user-agent strings: bid timing patterns that are too consistent, response times that don’t match human reaction speed, accounts that bid across an implausible number of categories at once. Rate limiting and step-up verification at bid time (not just at login) become table stakes, not nice-to-haves.
This is the harder, more interesting decision, and it’s the one eBay has publicly hinted it’s leaving open. Fighting agentic traffic indefinitely is a losing game of whack-a-mole against companies with far more resources than most marketplaces. Another approach is to get ahead of it: publish a permissioned agent protocol with clear rules (disclosed identity, rate limits, spending caps set by the human principal, mandatory human-confirmation thresholds above a certain bid amount) and make unsanctioned, undisclosed agents the only thing you’re actually blocking. That turns an enforcement problem into a product and revenue opportunity, and it lets you set the terms, rather than react to whatever comes next from OpenAI or Amazon.
This is the question that will matter most in a dispute, a fraud investigation, or a KYC review, and almost no platform has a good answer for it yet. You don’t want to be in the position of saying “we have no way to know who or what placed that bid” during arbitration if a bidder disputes a charge and says an agent bid without authorization, or if a seller claims a winning bid was fraudulent. That means baking bid provenance into your data model today: capturing device fingerprints; disclosed agent identity where permitted; explicit consent flows when a user authorizes an agent to bid on their behalf; and audit trails that can distinguish a scripted click from a human one. This is the unsexy work of infrastructure, but this is the difference between a platform that can defend a contested auction and one that cannot.
Where we land: don’t stay neutral on this
Our view is that outright, indefinite bans are the wrong long-term strategy, and permissive, unmonitored access is worse. The right move is closer to what eBay is signaling it might do eventually, just faster and more deliberately: block unsanctioned agents now, because the fraud, dispute, and fairness risks are real and immediate, while building toward a permissioned agent framework as a first-class feature, not an afterthought.
Platforms that treat this purely as a security problem will spend years playing defense against companies with far bigger AI budgets, and they’ll lose. Platforms that open the gates without any bid-provenance or accountability infrastructure will get burned by the first serious dispute or fraud case that lands in front of an arbitrator. The platforms that win the next five years will be the ones that figure out, on purpose, who gets to bid, how that’s proven and what happens when something goes wrong, not the ones who figure it out after a crisis forces them to.
That means a 2026 roadmap for most auction platforms should have four things, not just one policy update. For example, a detection layer to catch unsanctioned agent traffic today. A public statement of intent, so buyers and sellers know where you stand while the rules are still being written. A pilot permissioned-agent program with hard spending caps and human-confirmation thresholds for at least one category. And a provenance and audit system that can answer “was this bid human or machine” with evidence, not a shrug, should a dispute ever reach arbitration or a regulator.
None of this is optional forever. The volume moving through agentic channels is compounding by the quarter, not the year, and the platforms still debating whether to have a policy in twelve months will be negotiating from a position of catch-up rather than control.
If your platform doesn’t have a documented position on AI bidding agents yet, that’s the gap to close first. Bid integrity policy isn’t a compliance checkbox anymore. It’s quickly becoming core product strategy, and the platforms that get there first will be the ones setting the rules the rest of the industry ends up following.
