AI Agents 10 min read 16 November 2026

Is Your Website Blocking AI Agents and Their Customers?

United now limits unauthorised AI agents and JPMorgan won't allow rogue ones. Should your site block AI agents? Here's a policy to decide who gets in.

Is Your Website Blocking AI Agents and Their Customers?
Part of the Implementing AI Agents guideHow to Implement AI Agents in Your Business

Should your website block AI agents? Last week two very different businesses gave the same answer: only the ones we trust. According to The Motley Fool (fool.com/investing/2026/11/14/jamie-dimon-just-said-5-words-blocking-outside-ai), United Airlines announced on Wednesday 11 November that it is restricting access for unauthorised third-party AI agents and assistants, citing unauthorised bookings, unintended use of frequent-flyer miles and possible breaches of FAA and DOT passenger rules. JPMorgan Chase CEO Jamie Dimon put the bank's position in five words on its October earnings call: "You don't want rogue agents." The bank won't open customer accounts to just any outside agent.

Neither company is saying no to agents. Both are saying no to agents they can't identify, authorise or hold to account. That distinction is the whole game for every other business, and most are getting it wrong in one of two directions: blocking every agent by accident, or letting every agent in by default. This guide covers both, and gives you a one-page agent access policy you can apply this week.

The pressure has been building since October, when TechCrunch reported that personal agents such as Meta's Muse and ChatGPT's Dots were running into walls across the web. Amazon blocked Muse from shopping on its store. United's terms already forbade booking with an "automatic device". eBay restricts unauthorised agents. Yelp wants a paid data licence. Meanwhile Walmart, which actually partnered with Meta, still found its own human-verification buttons turning Muse away. Meta's line was blunt: turning away a personal agent means turning away the customer behind it.

Amazon and United can afford to block agents on purpose. They have the brand pull to make customers come to them directly. A mid-sized retailer, a hotel group, or a B2B supplier mostly can't. If a customer asks their agent to reorder supplies, book a room, or request three quotes, and your site is the one the agent can't get through, the order goes to a competitor whose site it can.

The Problem: Most Blocks Are Accidents

Very few small and mid-sized businesses have decided to block AI agents. The blocking happens anyway, through defaults nobody revisited. Three sources cause most of it.

  • CDN and firewall defaults. Since September 15, Cloudflare's defaults for new domains, newly added sites, and Free-tier accounts that never changed their settings allow search crawlers but block both AI training and AI agent traffic on pages that show ads. Cloudflare sorts AI bots into three buckets: search, training, and agent. The agent bucket is the one acting for a real customer in real time.
  • Bot challenges on the pages that make money. CAPTCHAs and "press and hold" checks on checkout, booking, and contact forms stop scripted abuse, but they also stop a legitimate agent at exactly the step where the customer was ready to buy.
  • Interfaces that only work for eyes. Prices that load only after a pop-up is closed, availability hidden behind a date picker with no underlying link, forms that rely on hover menus. An agent can sometimes push through these, but it is slow and fragile, and agents tend to choose the path of least resistance.

None of these are bad decisions on their own. Together they mean a growing class of buyer gets a worse experience on your site than anywhere else, and you will never see it in your analytics. A blocked agent doesn't fill in a complaint form. It just leaves.

When Blocking AI Agents Is the Right Call

United and JPMorgan are not being paranoid. Look at what an agent can do on each site. On an airline site it can spend miles, change a ticket or book a non-refundable fare, and a wrong booking turns into a refund dispute and a regulatory question. In a bank account it can move money, and the bank carries know-your-customer duties and liability for transactions it allowed. When the action is high-value, hard to reverse and regulated, blocking agents you can't verify is the sensible default.

Most businesses are not in that position on most of their pages. A product page, a price list, a stock check, a quote form or a booking request is low risk and easy to undo. An agent there is a customer asking a question faster. The mistake is treating the whole site like a bank vault because one corner of it is.

So the question isn't "block or allow". It is: for each action on our site, what is the worst an unknown agent could do, and could we undo it?

Authorised Agents vs Rogue Agents: How to Tell Them Apart

Dimon's phrase is useful because it names the real problem. A rogue agent is one you can't tie to a known operator or a consenting customer. You can tell the two apart with tools that already exist.

  • Verified bot and signed agent programmes. Cloudflare and other CDNs keep lists of verified bots, and Cloudflare supports cryptographically signed agent requests (Web Bot Auth), so the agent's operator can prove who it is. A user-agent string alone proves nothing; anyone can type "ChatGPT" into a header.
  • Delegated sign-in for account actions. When an agent needs to act inside a customer's account, it should go through OAuth or a similar consent screen, the same way apps connect to Google or a bank's open-banking API. The customer grants scoped access, you can see which agent holds it, and the customer can revoke it.
  • Step-up confirmation for money and irreversible changes. Let the agent prepare the booking, order or transfer, then send the final confirm to the human by email, SMS or push. The agent saves the customer time; the customer still owns the decision.
  • Per-agent logging. Record which agent, acting for which customer, did what. When something goes wrong, you need an answer within minutes, not a forensic project.

With those four in place, "authorised" stops being a feeling and becomes a check your systems run on every request.

Block, Allow, or Charge: Make It a Decision

The fix is not to throw the doors open. It is to treat AI agent access as a business decision per type of traffic and per part of the site, the same way you already decide what search engines may index.

  1. 1Separate training from agents. Blocking crawlers that harvest your content to train models is a reasonable default for many publishers. Blocking an agent that is trying to buy from you is usually not. Most bot tools now let you set these separately; check that yours does.
  2. 2Open the money paths, protect the rest. Product pages, availability, pricing, booking, and quote requests are where agent access earns revenue. Account settings, admin areas, and anything that changes stored customer data are where strict limits belong.
  3. 3Rate-limit instead of hard-blocking. A per-agent request ceiling stops abuse without stopping a customer. Verified agents can get a higher limit than anonymous traffic.
  4. 4Give agents a front door. A clean product feed, a structured booking or quote endpoint, or an MCP server gives agents a reliable way in that you control, instead of having them drive your website like a confused human. We covered why this matters for your internal systems at wizeb.com/blog/ai-agent-ready-business-systems-headless-2026.
  5. 5Charge where it makes sense. If what an agent wants is your data rather than your product, pricing it per request is now practical. We put a paid tool on our own MCP server to test exactly this: wizeb.com/blog/mcp-server-x402-agent-payments-2026.

A One-Page AI Agent Access Policy

Here is a starting policy most small and mid-sized businesses can adopt as is, then tighten where their risks differ. Write it down, agree it with whoever owns your website and security, and configure your CDN and app to match.

  1. 1Public information (product, service, pricing, availability, FAQ pages): allow search and agent traffic, rate-limit anonymous agents, and decide separately whether AI training crawlers get in.
  2. 2Lead capture (quote requests, contact forms, booking requests): allow verified agents, rate-limit the rest, and score submissions for spam instead of putting a CAPTCHA in front of everyone.
  3. 3Checkout and paid bookings: allow verified agents to build the basket or booking, use payment risk scoring rather than blanket bot challenges, and require customer confirmation above a set amount.
  4. 4Customer accounts (saved details, loyalty points, order changes): agents only through delegated sign-in with scoped, revocable access. No password sharing, no screen-scraping logins.
  5. 5Admin, staff and internal tools: block all outside agents. Your own agents get their own credentials and permissions, set the way we describe at wizeb.com/blog/ai-agent-security-guide.

Review it every quarter. Agent traffic is changing month to month, and the policy that was right in October may already be leaving orders on the table.

Standards Are Arriving, Slowly

The industry knows this is messy. According to TechCrunch, Meta, Walmart, Stripe, Sierra, Genesys and others have been working on an open standard for how agents talk to businesses for commerce. Separately, the Universal Commerce Protocol, backed by Google, Shopify, Amazon, Microsoft and Stripe among others, covers discovery through checkout, and Stripe and OpenAI publish the Agentic Commerce Protocol. These will make agent shopping smoother over the next year or two.

But standards help businesses that are already reachable. If your firewall drops agent traffic before it reaches your site, no protocol fixes that. The groundwork, deciding what agents may do and making sure they can actually do it, is yours either way.

A Realistic Example

Consider an online office-furniture retailer doing about $4 million a year, around a third of it from repeat business buyers. Its site runs on Cloudflare's Free plan and shows a small ad banner for a financing partner, so after September 15 the agent-blocking default applied without anyone noticing. Its checkout also carries a bot challenge added after a card-testing attack two years earlier.

The first sign of trouble was anecdotal: two account customers mentioned that their assistants reported they "couldn't complete the order" and they had gone elsewhere. A review found agent traffic being rejected at the edge and the checkout challenge failing every agent that did get through. The fix took a week. Agent traffic was allowed on product, stock, and checkout pages with a rate limit; training crawlers stayed blocked; the checkout challenge was replaced with risk scoring that only escalates on suspicious payment patterns; and a simple reorder endpoint was published for account customers. Within six weeks, orders placed through agents went from effectively zero to about 3% of repeat orders, card-testing attempts stayed blocked, and the team finally had a log showing which agents visited and what they tried to do.

Three percent is small. It is also the start of a channel that will keep growing, and the retailer is now on the right side of it while competitors are still invisible to it.

Where This Fits in Your AI Agent Plan

Most AI agent planning looks inward: which jobs your own agents should take on, what they cost, how to govern them. Customer agents arriving at your site is the outward half of the same shift, and it belongs in the same plan. If you are mapping out where to start, our full guide is at wizeb.com/blog/how-to-implement-ai-agents-in-business. The same shift is happening inside AI assistants too: ChatGPT now lets customers talk to a business's agent directly, which we cover at wizeb.com/blog/chatgpt-business-agents.

How WizeB Helps

WizeB audits how AI agents experience your business today: what your CDN and firewall let through, where bot challenges block buying journeys, and which pages agents can't parse. Then we fix it. That means setting agent access rules by section, adding rate limits and logging, and building the structured front doors, product feeds, booking and quote endpoints, or an MCP server, that let customer agents buy from you reliably. The same team builds your own internal agents, so both halves of the plan line up. See how we work at wizeb.com/services/ai-agents.

Check your agent readiness

Score how ready your business is for AI agents, inside and out, with the free AI-Native Index at wizeb.com/ai-native-index. Or book a call at wizeb.com/services/ai-agents and we will test your site with real agents and show you where they get stuck.

Three Questions to Ask Your Team This Week

  1. 1What do our CDN and bot-protection settings currently do with AI agent traffic, and did anyone choose that on purpose?
  2. 2Can an AI agent get from a product or service page to a completed order, booking, or quote request on our site without a human stepping in?
  3. 3Do we have any log that tells us how many agents tried to buy from us last month, and how many failed?

Ready to act on this?

We build exactly what this article is about.

Tell us about your situation — we'll come back with a realistic assessment.