Stopping Fake Leads with Agentic Validation on Our Website

We love demoing agents. But our standard contact form wasn’t agentic—until now.

For three straight weekends we saw waves of nonsense form entries: random characters in Company/Role, one-line messages like "asdf", or numbers-only fields. Meanwhile, prospects who booked time directly on Cal.com were almost always serious and qualified.

So we did something about it.

The Problem

• Fake or low-effort submissions spiked on weekends
• Manual triage slowed us down and distracted from real opportunities
• Our demos already used explainable agents with great results, but the site’s own form had no reasoning layer

Our Agentic Solution

We added an explainable, model-agnostic validation agent to the Contact form.

What it does

1. Client-side intercept

   • On submit, an agent validates the payload before it reaches email delivery
   • If information looks invalid, we show an on-screen agent message:

     "Hi there, I am an agent within the system… I don't believe this is a valid submission, and here is why: …"

   • Users can correct fields in place

2. Server-side protection

   • The backend runs the same validator and rejects invalid posts with structured reasons
   • Prevents bypassing the client by posting directly to the endpoint or via scripts

3. Hybrid reasoning (precision-first)

   • Heuristics catch obvious junk: mostly numbers/symbols, low vowel ratio gibberish, long unbroken strings, too-short name/message, nonsense tokens in Company/Role
   • LLM pass (Together API, OpenAI-compatible) kicks in only when heuristics pass—to double-check semantics and reduce false negatives
   • If uncertain, the agent favors VALID unless there are two or more red flags

Design principles

• Explainable: Always gives reasons for rejection
• Conservative: Avoids false negatives on legit submissions
• Resilient: Fails open in development if the model is unavailable
• Consistent: Re-uses the same validator on client and server

Technical Implementation

• Client: components/lead-form.tsx
  • Calls /api/contact/validate first; shows agent message on failure
  • Proceeds to /api/contact only if valid
• API: app/api/contact/validate/route.ts, lib/lead-validation.ts
  • Zod schema + heuristics
  • Optional Together-powered reasoning with structured JSON output: { valid, reasons[], confidence }
• Email: app/api/contact/route.ts
  • Re-validates server-side before sending email via Resend
  • Dev fallback: if email isn’t configured locally, return success to keep flows testable

Results

• Immediate reduction of weekend noise
• Saved time by eliminating manual triage of junk submissions
• Preserved a friendly path for valid users with clear, actionable feedback
• Reinforced our site as a live showcase of agentic design, not just demos

Why this matters

Agentic websites aren’t just about flashy demos—they protect your team’s time and create better user experiences. By placing a reasoning layer in front of unstructured inputs (like contact forms), you keep signal high and noise low.

What’s Next

We ship intentionally and measure before changing course. Our roadmap for this initiative:

1. Measure impact first

• Track pre/post metrics: weekend submission volume, rejection rate, false positive/negative surveys, time-to-first-response, and qualified-meeting conversion.
• Use these to adjust thresholds (heuristics and LLM confidence) rather than guessing.

2. Tie in our AI Call Center for pre‑qualification and scheduling assist

• Smart handoff from the form to a voice/chat agent that can clarify intent, gather missing details, and offer calendar slots.
• Automated confirmation and reminders to reduce no‑shows.

3. Lead data enrichment

• Enrich company/person fields (domain, HQ, revenue band, headcount, tech stack) from public sources.
• Route by ICP fit and flag high‑priority accounts.

4. “Who‑Why‑What” opportunity brief (auto‑generated)

• An agent compiles a crisp brief for our sales team:
  • Who: contact + org bio, role context, relevant background
  • Why: pains/goals inferred from submission + enrichment
  • What: initial hypothesis of scope, risks, next best actions
• Delivered in the internal notification so we hit first calls prepared.

We’ll roll out these steps incrementally, keeping the same principle: conservative by default, explainable always, and measured at each step.

External signals we’re watching (sources)

From HBR — How Successful Sales Teams Are Embracing Agentic AI (Sept 15, 2025)

• Autonomous BDR agent for top-of-funnel: scored leads from behavioral signals, generated personalized outreach using purchase/usage data, maintained real-time dialogue, and handed off to humans with full CRM context — delivering a 6% lift in response rates and a projected $50M in incremental annual revenue. Source: Harvard Business Review — How Successful Sales Teams Are Embracing Agentic AI (https://hbr.org/2025/09/how-successful-sales-teams-are-embracing-agentic-ai)
• Customer summaries to accelerate prep: at a North American wealth management firm, an agent synthesized CRM + external data to produce meeting briefs, cutting prep time by 30%+ and driving a 6% revenue uplift. Source: HBR article above.
• Retention and growth agents: an AI CX agent resolved ~85% of queries (including second‑order, multi‑system questions), cutting handling time by 65%; churn‑prevention agents reason about next‑best actions and initiate interventions. Source: HBR article above.

From LangChain — State of AI Agents (2024)

• Leading use cases today: research/summarization (58%), personal productivity & scheduling (53.5%), and customer service (45.8%). These directly inform our roadmap items for enrichment and scheduling assist. Source: LangChain — State of AI Agents (https://www.langchain.com/stateofaiagents)
• Controls and guardrails in production: tracing/observability are top controls; most teams use read‑only tool permissions or require human approval for writes/deletes; enterprises pair guardrails with offline evaluation. This validates our conservative “explainable-by-default” posture. Source: LangChain article above.
• Adoption and exemplars: ~51% of respondents already have agents in production (63% for mid‑size companies). Notable production apps include Cursor, Perplexity, and Replit — demonstrating reliable agent patterns beyond simple chat. Source: LangChain article above.

Related Work

• Our broader meta-implementation: How We Built Our Own Website with Agentic Design

Want help implementing this?

We can bring the same pattern to your public site, internal portals, or apps—model-agnostic, explainable, and tuned for your thresholds.

Let’s talk or schedule directly on Cal.com: https://cal.com/mrmetaverse/virgentai