Try Onsa
Vercel had 10 SDRs qualifying inbound leads. Six weeks later, they had one.
The other nine weren’t fired. They moved to outbound — the work that actually needs a human. The AI agent matched their conversion rate. One GTM engineer built it, spending about 30% of his time.
That’s not a pilot. That’s a production deployment with measurable results.
I’ve been thinking about this story since it came out, because it mirrors something I see across every sales team I talk to: lead qualification is one of the highest-volume, lowest-judgment tasks in sales. And teams are still doing it manually.
Let’s do the rough arithmetic.
Ten SDRs qualifying inbound leads. Average fully loaded SDR cost: $80-100K/year including salary, benefits, tools, management overhead. That’s $800K-$1M per year on a function that an AI agent can handle at the same quality level.
Vercel’s results after automation: - Same lead-to-opportunity conversion rate — the agent matched human performance - Faster response times — AI handles leads 24/7, including nights and weekends when interest is fresh - 9 SDRs freed up for outbound — where relationship building and creativity matter
The speed advantage alone is significant. Research consistently shows that responding to inbound leads within 5 minutes makes you 21x more likely to qualify them compared to waiting 30 minutes. Human SDRs take breaks, attend meetings, and sleep. AI doesn’t.
Not every sales task should be automated. But lead qualification has the exact profile that works:
High volume. Most B2B companies with any inbound motion get dozens to hundreds of leads per day. Each one needs the same basic research and scoring.
Structured process. Qualification follows a repeatable pattern: check the company size, verify the industry fit, assess the request, score it, route it. This is a workflow, not an art.
Speed-sensitive. The faster you respond, the higher your conversion. Every minute of delay costs you.
Low-risk errors. A wrong qualification score doesn’t lose a deal — it just gets corrected by the rep who picks it up. Compare that to automating a contract negotiation, where a mistake could cost you hundreds of thousands.
Doesn’t require relationship. The lead doesn’t care who — or what — qualifies them. They care about getting a fast, relevant response.
This is the same pattern I described in the C.H. Robinson case study. They automated 2,000 daily freight quotes — high volume, structured, speed-sensitive, low-risk. Stock jumped 20%.
The pattern keeps repeating because the economics are obvious.
Most lead scoring is broken. It’s either too simple (company size + job title = score) or too complex (50-factor models that nobody trusts).
After studying what actually predicts conversion, I built a scoring framework around three dimensions:
Does this company match your ICP?
• Company size — headcount, revenue range
• Industry — is this a vertical you serve?
• Geography — do you operate in their market?
• Tech stack — do they use tools you integrate with?
• Role — is the contact a decision maker or an intern researching for a class project?
Fit is table stakes. A perfect-fit company that isn’t looking for anything scores 40 out of 100. You need more.
What did they actually do?
• Request type — demo request vs. newsletter signup vs. pricing page visit. These are not equal.
• Message quality — did they describe a specific problem? Or just say “tell me more”?
• Engagement depth — how many pages did they visit? Did they read a case study or just the homepage?
• Channel — referral vs. organic search vs. paid ad. Each signals different intent levels.
Intent separates “I have a problem and I’m looking for a solution” from “I’m browsing.” Both are valid, but they need different handling.
Is there urgency?
• Urgency language — “We need this by Q2” vs. “We’re exploring options for next year”
• Trigger events — new funding, leadership change, competitor switch, expansion
• Budget cycle — is this the start of their fiscal year? End of quarter?
• Competitive mentions — are they evaluating alternatives? That means they’re buying.
Timing is the most underrated dimension. A warm lead with no urgency stays warm forever. A moderately qualified lead with a hard deadline closes next month.
Add the three scores:
Score: 75-100 | Tier: Hot — Route to AE immediately. Respond within 1 hour.
Score: 50-74 | Tier: Warm — SDR follow-up within 24 hours. Personalized outreach.
Score: 25-49 | Tier: Cold — Nurture sequence. Check back in 30 days.
Score: 0-24 | Tier: Disqualified — Auto-respond with resources. Don’t waste rep time.
The beauty of explicit scoring is that it’s debuggable. When a rep disagrees with a qualification, you can see exactly which dimension drove the score. Was the fit wrong? Was the intent signal misread? You can fix the specific issue instead of tweaking a black-box model.
I built this into a tool with two modes:
Build your scoring model. You describe your ICP, feed it your closed-won deals, and it analyzes your website and competitors. It outputs a scoring model customized to your business — not a generic template.
The key insight: every company’s Fit criteria are different. A 50-person startup selling to enterprise has completely different qualification signals than a 500-person company selling to SMBs. The model has to be built from your data, not borrowed from a blog post.
Qualify individual leads. Give it an inbound request — an email, a form submission, a chat message, whatever format. It researches the person on LinkedIn, scrapes the company website, applies your scoring model, and returns:
• Score breakdown (Fit: 32/40, Intent: 28/40, Timing: 15/20 = 75, Hot)
• Research summary — what the company does, how big they are, recent news
• Draft response — personalized to their specific request
• Next steps — what the rep should do with this lead
The entire process takes about 30 seconds. A human SDR doing the same research and scoring takes 15-30 minutes per lead.
At 50 inbound leads per day, that’s the difference between 25 hours of SDR time and 25 minutes of compute time.
Three things stood out about Vercel’s approach:
They automated the full workflow, not just a piece.
A lot of teams try “AI-assisted qualification” where the AI suggests a score and a human approves it. This is the worst of both worlds — you still need the human, and now you’ve added a step.
Vercel’s agent handles qualification end-to-end. A human reviews edge cases, not every lead.
They redeployed people, not fired them.
The 9 SDRs moved to outbound, where human skills — creativity, persistence, relationship building — actually matter. This is the right framing: AI doesn’t replace salespeople, it replaces the tasks that waste salespeople’s time.
They shipped fast.
Six weeks. One part-time engineer. Not a 12-month AI transformation initiative with a steering committee and a consulting firm.
The technology to do this exists today. The bottleneck is almost never technical — it’s organizational willingness to let AI handle a function that humans have always done.
“But leads are people. They deserve a human touch from the first interaction.”
I hear this a lot. Here’s the problem: that “human touch” is usually a templated email sent 4 hours after the lead submitted a form, while the SDR was in a team standup.
The AI agent responds in minutes with a personalized message based on actual research into the lead’s company and role. The “human touch” argument assumes humans are actually delivering a personal experience. Most of the time, they’re not.
The real human touch happens after qualification — when a skilled rep has a conversation about the prospect’s specific challenges, builds rapport, and guides them through a buying process. That’s where humans are irreplaceable. Everything before that is logistics.
If you want to try this approach:
1. Map your current qualification process. How many steps? How long does each take? What information do your SDRs look up for every lead?
2. Define your scoring criteria. Use the Fit + Intent + Timing framework. Be specific about what scores a 30 vs. a 35 in each dimension.
3. Start with scoring, not routing. Run the AI scorer in parallel with your SDRs for two weeks. Compare scores. Calibrate.
4. Automate routing once you trust the scores. Hot leads go straight to AEs. Warm leads get automated follow-up. Cold leads enter nurture.
5. Measure conversion rates by tier. If your Hot tier converts at 40%+ and your Cold tier at 5%, the model is working. If not, adjust the weights.
The open-source scoring plugin is on GitHub. It works with Claude Code and includes the /design-scoring wizard to build your model and /qualify-lead to score individual leads in real time.
I’m Bayram, founder of Onsa. We build AI agents for B2B sales — automating the research, qualification, and outreach that used to take hours per lead. If your team is spending more time qualifying than selling, let’s talk.