Agentic commerce, the idea that AI agents will shop and pay for you, is one of the most hyped themes in AI this year. Payment networks, OpenAI, Google and Shopify are all building for it. But this week a viral post from a San Francisco tech worker cut through the marketing: in her tests, AI shopping agents “basically” did not work at all. The replies turned into a broad debate about why agents keep failing at the final step.
Quick Take
- The spark: Maia B (@maiab) wrote that she had never heard so much hype about a product experience that “basically does not work at all” for her, adding: “I feel like I’m using theranos.”
- What broke: In the cases discussed, Sephora’s bot check blocked the agent, and an AutoZone order was cancelled after about 17 hours by fraud review. Some orders, such as one at Buck Mason, worked.
- Why: Checkout and fraud systems are built to stop automated buyers. Guest checkout, one-time virtual cards and mismatched names look like fraud.
- The fix in progress: Payment standards like Mastercard Verifiable Intent, Google’s AP2 and Visa’s Trusted Agent Protocol aim to tell merchants which agent is buying, for whom.
What happened
On November 3, Maia B posted her verdict on agentic shopping. The post collected more than 3,500 likes and was shared by data engineer Josh Wills among others. In follow-up replies she explained that agents work for parts of shopping, such as product comparison, and that Claude running on her own machine was less likely to be blocked because “it doesn’t ‘look like’ an agent – it’s using my IP address etc.”
The failure cases discussed in the thread and in summaries of it were specific: a Sephora bot check that stopped the agent, an AutoZone order that was cancelled after roughly 17 hours of fraud review because it combined guest checkout, a one-time virtual card and names that did not match, and errors with food delivery apps. A purchase at Buck Mason went through.
Background: the agentic commerce push
Since 2025, OpenAI has offered Instant Checkout in ChatGPT, Google has pushed its Agent Payments Protocol (AP2), Visa has introduced its Trusted Agent Protocol and Mastercard its agent payment tools including Verifiable Intent. Industry forecasts are bullish: The Paypers has projected that 25 to 30% of online purchases could involve agents by 2030. Surveys show interest but limited trust; according to Visa data cited on X, 54% of Americans have used AI to help them shop, but only 23% trust it with payments.
How it works, and where it breaks
Most shopping agents today operate a browser like a human: search, click, fill forms, pay. That works for browsing and comparison. Checkout is different. Retailers run bot detection, device fingerprinting and fraud scoring designed to block exactly this behavior. An agent coming from a cloud data center, using a disposable card and a guest account, triggers many red flags at once.
Fraud teams face the opposite problem in built-in checkout flows. A The Paypers report with Riskified from November 3 said up to 34% of the fraud signals merchants normally rely on were missing in ChatGPT Instant Checkout orders. Chargeback specialists have warned that disputes over “my agent bought that” will be hard to resolve.
| Shopping step | Agents today | Main obstacle |
|---|---|---|
| Product research and comparison | Works well | Hallucinated specs or prices |
| Large repetitive carts (e.g. groceries) | Promising | Substitutions, errors |
| Checkout on retail sites | Often fails | Bot detection, captchas |
| Payment and fraud review | Often fails | Virtual cards, name mismatch, guest accounts |
| Built-in checkout (e.g. ChatGPT) | Works for listed merchants | Missing fraud signals for merchants |
Pros and cons
For users
- Pro: Agents are genuinely useful for research: comparing products, prices and reviews, and filling large repetitive carts.
- Con: Payments may fail, orders may be cancelled hours later, and dispute rights for agent-made purchases are unclear.
For developers
- Pro: A huge opportunity for agent identity, payment tokens and merchant APIs, the missing layer the thread kept pointing to.
- Con: Browser automation against hostile checkout pages is brittle, and getting blocked is often the intended outcome for the site.
For companies
- Pro: Retailers that expose agent-friendly APIs and verified agent checkout could capture early agent-driven demand.
- Con: Letting agents through raises fraud exposure; blocking them risks losing sales as agent traffic grows.
What X is saying
re: agentic shopping. I have NEVER in my life heard so much hype and discussion about a tech/product experience that basically does not work at all for me. I feel like I’m using theranos
— Maia in SF (@maiab) November 3, 2026 · View on X
current take: cart size is the deciding factor for agentic commerce. agents will win the 40 item grocery cart. they won’t win the one-item shop pay checkout
— @mattrbahr November 3, 2026 · View on X
The hard part behind that door is trust. A merchant needs to know which agent is buying, for whom, and what it’s allowed to spend, or every agent checkout looks like fraud. That layer decides who wins agentic commerce.
— @AndresDancausa November 3, 2026 · View on X
Maia B also pointed to the underlying problem when asked why agents use web pages at all: “it’s because it’s the universal interface and the infrastructure isn’t there to support anything else. but obviously an API would be much better and faster for an agent to interact and transact with.”
What Grok says
We did not find a direct Grok reply on this specific thread at the time of writing. The consensus on X, which Grok’s broader commentary on agents also reflects, is that agents are good at research and repetitive tasks but still need human supervision for anything involving money.
Critical analysis
One person’s experience is not a market study, and the comparison to Theranos, a fraud case, is rhetorical. But the failures are structural, not random. Merchants spent two decades teaching their systems to stop bots and suspicious cards; agents look like both. Until agent identity and payment credentials are widely accepted, the browser-automation approach will keep failing at checkout. The more interesting signal is the split @mattrbahr describes: high-effort, repetitive carts are where agents add value; one-click checkouts are already fast enough that agents add little.
What’s still open
- Which agent identity standard will merchants and payment networks converge on?
- Who carries liability when an agent makes an unwanted purchase?
- Will major retailers whitelist verified agents or keep blocking them?
- How will chargeback rules change for agent-initiated payments?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on posts on X from November 3–4, 2026, industry reporting from The Paypers and Riskified, and public material from payment networks. Individual failure cases are as described by the people involved; i10x did not reproduce them.
i10x Perspective
Agentic commerce will not be won by smarter browsing agents. It will be won by whoever builds the trust layer that lets a merchant say yes to an agent: verified identity, scoped spending limits and clear liability. Until then, treat AI shopping agents as research assistants, and keep your own hand on the pay button.
