On 15 October 2026 Google published the agent it uses to test WebMCP, the proposed standard that lets a website tell an AI assistant how to use its forms and tools. Buried in that project’s permission rules is one sentence every business with a booking form, a quote form or a checkout should read. When a website marks a tool as consequential, “the user is always asked, even if alerts are off, even if the tool also says it is read-only, and there is no “allow for this chat” option.”
That sentence is the answer to the question I get most often about AI agents from business owners: what stops one of these things booking the wrong day, or paying twice, or sending a message to my customer I never wrote? The answer, on a site built properly, is a red box and a human. Here is what that means and what you can do about it without learning to code.
What WebMCP is, in one paragraph
Today, when an AI assistant tries to do something on a website for you, it looks at the page the way a person would and guesses: that looks like a date field, that looks like the submit button. Google calls this actuation and describes it as a long chain of steps, each one open to interpretation. WebMCP lets the website say, in a form the assistant can read, “this is my booking tool, it takes a date and a postcode, and it is a real booking, so check with the customer first.” The assistant stops guessing. The full guide is here; the short page on the consequential flag is the one this article builds on.
The flag, and what a careful agent does with it
The specification gives a tool four optional labels. One says the tool only reads. One says the tool may bring back content the site does not control. One says the tool is for developers. And one, consequentialHint, says the tool does something “significant, real-world, or non-reversible, ex: booking a flight, transferring money.”
Google’s reference agent treats those labels like this. A read-only tool runs quietly. An unlabelled tool asks once and offers “allow for this chat”. A consequential tool asks every time, in a red prompt that names the action, with no way to switch the asking off. The customer presses Allow or the booking does not happen.
Why this matters for a booking form in Southampton
Think of the counter at a good independent shop. The assistant behind it will happily tell you the price, check the stockroom, and answer questions all day. When you say “I’ll take it,” they look you in the eye, repeat the total, and wait. That pause is not friction. It is the moment the shop protects you and itself from a mistake.
The consequential flag is that pause, written into your website. An EPC assessor whose booking tool carries the flag knows that no assistant will book a Tuesday on a customer’s behalf without the customer confirming it. A solicitor whose enquiry form carries it knows a message cannot be sent in a client’s name by accident. A solar installer whose quote tool does not carry it, because nobody thought about it, has told every agent that reads the page that no confirmation is needed.
None of this requires the customer to know what WebMCP is. They see a red box that says “Book appointment at EPC Upgrade Group for Tuesday at 10am. Allow?” and they press the button. The work is on your side of the counter.
What you can do now, without code
1. Make a list of the three things on your site that commit someone to something. A booking, a quote request, a purchase, a message sent in their name. Those are your consequential tools, whether or not they carry the flag yet. Everything else (search, prices, opening hours) is read-only.
2. Ask whoever built your site one question: “If an AI assistant filled in our booking form, how would it know the booking is real?” If the answer is a blank look, the form is a guess for the agent. If the answer involves the declarative WebMCP attributes on the form, you are ahead of nearly everyone. The two-attribute version is here, and it is small enough to hand to any developer.
3. Do not pay for this as a product yet. WebMCP is a Chrome origin trial, not a standard. Almost no customer’s browser can call a tool on your site this month. What is worth doing is the honest labelling, because it costs almost nothing and it is the part that protects you when the browsers catch up.
What we are not claiming
Google’s extension is one agent, written to show the specification at its best. Other agents may handle the flag differently; the W3C draft does not force them to. Nothing here says your bookings will rise. It says that when agents arrive at your site, the difference between a confirmed booking and a wrong one will come down to a label you can set today. We are testing a booking tool with the flag on our own work and will record what happens in the observed-outcomes register when a browser first calls it.
