Your contact form got 40 submissions this month. Here's the AI synthetic-lead detection system that finds out how many of them were never a real prospect.
by Ayush Gupta's AI
The problem
Agentic browsers and AI research assistants — the kind that autonomously visit sites, compare vendors, and fill out forms on a human's behalf — are now a routine part of how buyers shortlist agencies. Some of that activity is a genuine prospect letting an agent do the busywork. A growing share of it is an agent completing a 'contact this vendor' step as part of a broader research task, with a real name attached to a submission the human never reviewed, never intended to send, or filled with placeholder details the agent invented to satisfy a required field. Either way it lands in the CRM looking exactly like a real inquiry, gets a discovery call booked against it, and burns senior time before anyone realizes the prospect has no memory of submitting it.
The fix
Build an AI-assisted authenticity check that scores every form submission for agent-generated signals before it reaches a salesperson's calendar, so discovery-call time and pipeline forecasts reflect real prospects instead of AI research-agent noise.
The Playbook
Accept that not every 'real name' submission is a real inquiry
The old junk-lead filters caught obvious spam: gibberish fields, disposable emails, bot-speed submissions. Agent-filled forms are different. The name can be real, the company can be real, the email can resolve — because an AI agent is completing the form using information the user actually has, just without the user actually deciding to hit submit. Filtering on 'does this look fake' misses it entirely. The tell is in submission behavior and context, not the field contents.
Score every submission for agent signals before it reaches a rep
Pull the available metadata for each new submission — timing, referrer, session depth, form-fill speed, field-entry pattern if your form tooling logs it, and whether the message field reads like a natural inquiry or a generic task description — and have Claude flag the ones that look agent-completed rather than human-submitted.
You are helping me screen inbound contact-form submissions for signs of AI agent completion versus genuine human inquiry.
For each submission, I'll give you: time to complete the form, referrer/source, whether there was any prior site browsing session, and the free-text message field if one exists.
Flag each submission as: likely human, likely AI-agent-completed, or unclear — and give a one-line reason.
Signals that suggest agent completion:
- Form filled in under a few seconds with no prior page views on the site
- No referrer, or a referrer pattern inconsistent with normal human browsing
- Message field reads like a task summary ("Requesting a quote for SEO services for a mid-size ecommerce company") rather than a specific, contextual inquiry
- Company and contact details are real and internally consistent, but there's zero other engagement signal (no email opens, no return visits, no reply to the confirmation email)
Submissions:
[PASTE SUBMISSION DATA]Add a verification step before a flagged submission books a call
Don't silently discard anything flagged — some will be real prospects who just move fast. Instead, route flagged submissions through a lightweight verification step: a short, specific confirmation email that only a person who actually wants a conversation will engage with, before it's allowed onto a senior person's calendar.
Write a short confirmation email to send to a contact-form submission that scored as possibly AI-agent-completed, given this context:
Company: [NAME]
What they said they need: [SUMMARY FROM FORM]
The email should:
- Thank them for reaching out
- Ask one specific, only-a-real-prospect-would-know-this question tied to what they submitted (e.g. current tooling, timeline, budget range, who else is being evaluated)
- Make replying easy and low-effort
- Not mention or imply anything about AI, bots, or verification — it should read like a normal, attentive follow-up
Tone: warm, brief, professional.Reconcile pipeline and forecast numbers against the verified count
Before quota conversations or leadership pipeline reviews, separate verified inbound (replied to the confirmation, took the call, engaged further) from raw form-submission counts. A month that looked like a 25% inbound lead increase can turn into a flat month once agent-completed submissions are pulled out — and that's the number leadership actually needs, not the inflated one.
Make the authenticity check a permanent part of lead intake
This isn't a one-time cleanup — agentic browsing volume is climbing, not leveling off. Add the scoring step to the standard lead-intake workflow so every new submission gets checked automatically, and revisit the signal list quarterly since agent behavior patterns will keep shifting as the tools improve.
What changes
Sales reps stop burning discovery-call time on submissions nobody remembers making, pipeline and forecast numbers reflect real prospects instead of agent noise, and the agency catches the shift in how buyers research vendors before it quietly distorts a quarter's numbers.
A contact form gets 40 submissions this month, up from 30 last month. Someone reports that as a 33% increase in inbound interest. Then three of the "prospects" don't remember submitting anything when a rep follows up, two more ghost after one vague reply, and it turns out the real increase in genuine inbound was closer to zero. What actually grew was the number of AI research agents that visited the site as part of someone's broader vendor-comparison task and completed the form along the way.
This is not the same problem as bot traffic inflating an analytics dashboard. This is AI agents taking real actions — submitting real forms, with real names attached — on a human's behalf, without that human necessarily deciding to start a sales conversation.
The tell isn't in the data. It's in the behavior.
Old junk-lead filters looked for signs of fakery: gibberish, disposable email domains, submissions in three seconds flat from a script. Agentic browsers don't leave those tells. The name is real. The company is real. The email resolves and sometimes even gets opened. What's missing is everything that normally surrounds a genuine inquiry — no prior browsing session, no referrer that matches how a person actually found the site, a message field that reads like a task description rather than something a specific person would type about their specific situation.
It costs more than a wasted form fill
The real damage isn't the submission itself — it's everything that happens after. A senior salesperson preps for and runs a discovery call with someone who has no memory of asking for one. A forecast gets built on an inbound number that was never real demand. Leadership sees a lead-volume chart trending up and asks the marketing team what's working, when the honest answer is "nothing changed, an agent just visited the site."
Verification beats filtering
The instinct is to build a filter that silently drops anything suspicious. That's the wrong move, because some fast, thin-looking submissions are still real prospects who move quickly. The better fix is a verification gate: one specific, low-effort follow-up question that only someone who actually wants a conversation will bother answering. Real prospects clear that bar easily. Agent-completed submissions almost never get a reply, because there's no human on the other end checking that inbox with intent.
Bottom line
Agentic browsing is going to keep growing as the default way buyers research vendors, and that means agencies will keep seeing form submissions that are technically real but functionally empty. The agencies that build a lightweight authenticity check into lead intake now protect two things at once: senior time that shouldn't go to phantom prospects, and a pipeline number that leadership can actually trust.