Your next customer might be an AI assistant

For twenty years, the visitor to your contact form has been a person with a browser. That is starting to change, and faster than most small businesses have noticed.

AI is already asking businesses for prices

In 2025 Google added a feature to Search that lets someone ask an AI to phone local businesses for them. The AI calls, asks about price and availability, and sends the person a summary by text or email. It launched for a handful of categories: auto repair, tires, hair salons, wellness, dry cleaning, vets, pest control and plumbing.

Invoca, which tracks phone calls for businesses, measured what happened next:

In November alone, calls to plumbers rose more than 650% and calls to vets more than 1,700%. And the businesses were not ready:

• 26% of these calls were not answered at all.
• Of the ones that were answered, 49% did not get a price.

The person who asked never hears a ringing phone. They see a summary. A business that did not answer, or would not give a price, is a line in that summary with nothing in it.

And agents are filling in forms, too

Phone calls are one door. Web forms are the other.

Browsing agents such as ChatGPT agent can open a website, find the form, fill it in and press send on someone's behalf. In August 2025 Cloudflare started a "signed agents" program: agents like ChatGPT agent sign each request with a cryptographic signature (a standard called Web Bot Auth), so a website can check who is really knocking.

That matters because most websites today treat any automated visitor as spam. CAPTCHAs, honeypots and bot filters were built to keep machines out. Now some of those machines are carrying a real customer's request.

What goes wrong when an agent meets a normal form

Most contact forms were built for a person who can guess. An agent has to read. Here is where they trip:

• Unlabelled fields. A box with placeholder text "Your message…" and no label is a guess for a machine.
• Everything in one text box. "Tell us about your project" gives an agent nothing to fill in properly, and gives you nothing structured back.
• No price anywhere. The agent's whole job is often "find out how much." If the only answer is "call us," it moves on.
• A thank-you page and silence. The agent submitted the form, but where does it check for the answer? Usually nowhere. Your reply goes to an email inbox the agent cannot see, a day later.
• Blocked at the door. A CAPTCHA built to stop bots stops this one too, along with the customer behind it.

What an agent-ready form looks like

None of this means building a separate site for robots. The things that help an agent are the same things that help a person in a hurry:

1. Clear, labelled fields with the right types: an email field that is an email field, a date that is a date, a short list of job types instead of a blank box.
2. A price range and a next step, so the answer to "how much and when?" exists at all. (We wrote about how to give a range safely.)
3. A machine-readable description of the form, so an agent does not have to guess what each field means.
4. A status link. When the form is submitted, the agent gets an address where your answer will appear, so it can come back and tell the person.
5. Verification instead of blocking. Check signed agents and let them through; keep the spam filters for everything else.
6. A human decision on what goes back. An agent asking for a quote should get the same answer a person would, approved by you.

How OnDot forms handle it

Every OnDot form is built this way from the start. The same form your customers see is also published as a tool an AI assistant can use: a description of every field, a tool definition, and an MCP endpoint for assistants that speak that protocol. When an agent sends a request, it must say who it is acting for, it gets a request id and a status link, and your answer, once you approve it, appears there as well as in the customer's email.

Signed agents are checked. Unsigned automation is treated with more caution. And in your inbox, a request from an assistant is marked as one, so you always know who you are talking to.

Nothing about this changes the rule we started with: bots draft, you approve. Whether the customer typed it or their assistant did, nothing goes out until you say so.

Where this is heading

Nobody knows yet how much of a typical small business's enquiries will arrive through AI assistants in two years. Invoca's numbers are early and come from one feature. But the direction is clear, and the businesses that answer clearly, quickly and with a price will be the ones that show up well in whatever summary the customer reads.

The good news: the work that gets you ready for agents is the same work that wins more human customers today.