Agent swarms are coming for small business web forms

On Tuesday, Nikita Bier of X posted a prediction about who is going to be hurt by AI agents. @nikitabier wrote that bot detection and human verification will be one of the most urgent needs for businesses in the next few years, because "agent swarms will suffocate every website and form." He was specific about who gets hit first: "small companies and government websites are most vulnerable."

He also said that when X went looking for a product to handle this, it couldn't find one that pulled the current techniques together, so the team built its own.

Two days later, @paulg quoted him. Paul Graham said the problem is big enough for several startups, not one. One kind would filter agents out. Another would be "meant to welcome them in an orderly way."

Old problem, new visitors

Anyone who runs a business website has deleted a pile of fake inquiries. What's different now is that the visitors can read. An AI agent fills in a form the way a person would. It understands the fields, writes a believable message, and gets past the tricks that caught older bots. Something like it showed up on my own site in July, when bot submissions started leaving my honeypot field alone after months of getting caught by it.

Then there's volume. One person with a research task can start dozens of agents at once, and they don't get tired or bored, which is why people run them in swarms.

Ethan Mollick added a third angle on Thursday. @emollick argued that the security threat from agents may come less from bad actors than from masses of agents chasing harmless goals. His example is a swarm trying to break into your IT systems to learn what you paid for company t-shirts, all as part of a research job to "find good shirt prices."

It reads like a joke, but this week we learned that an OpenAI agent got into an Australian government Medicare portal while researching public medical spending. It was an ordinary research task, and it still ended up somewhere it wasn't supposed to be.

My read

Bier is right about who is exposed. A large retailer has a security team and a budget for bot management. A dental office, a plumber or a parish running a donation form usually has a free form plugin and a shared inbox. When agent traffic grows, those forms get hit first and the people reading the inbox feel it first.

Graham's split is the more useful idea for an owner, though. Not every agent on your site is an attacker. More customers will send an assistant to request a quote, check your hours or book an appointment. If you block every agent, you turn some of those customers away. If you let everything in, the real leads get buried in a thousand fake ones.

The better question for a small business is which parts of the site an agent should be able to use, and how. A few practical steps:

• Give agents an easy, legitimate path to what they need: clear hours, pricing ranges, service areas and a booking link in plain text on the page.
• Put limits on anything that costs you money or time, like quote requests that trigger a callback, free consultations or SMS sends.
• Check form submissions automatically before a person sees them, so staff aren't sorting junk by hand every morning.
• Keep an eye on volume. If near-identical inquiries suddenly jump, treat it as a warning before you treat it as a good week.

You don't need a new vendor for any of this. By Bier's account the right tool doesn't exist yet, and Graham expects several companies to try building it. Until then, set up your forms and intake so a flood can't cost you much.

What this means for businesses using AI

Most owners think of AI as something they decide to adopt. These posts are about AI you didn't choose: other people's agents visiting your website, in bigger numbers every month if Bier is right. I'd expect the businesses that handle it well to be the ones that already decided which inquiries get an automatic answer and which ones go to a person.

If you'd like help mapping that for your own lead flow, New Face Design's free process audit looks at how inquiries come in and where they get stuck. Start here.