How We Built an Automated B2B Lead Research Pipeline Using AI

Lead research was killing us. Not the strategy. Not the outreach. The research — finding the right companies, qualifying them, identifying the right contact, writing a personalised first message. Four hours a week, minimum. All manual. All repetitive. According to HubSpot’s 2025 State of Sales Report, B2B sales representatives spend an average of 21% of their week on research and prospecting — time that could be reclaimed through automation without sacrificing quality. At Varnan, we measured this problem, automated it, and measured it again. Here’s the exact system we built.

“The bottleneck in B2B outreach isn’t sending messages — it’s knowing who deserves one, and having enough real context to make the first one worth reading.”

— Amal Jandheer, Founder, Varnan

What Problem Were We Solving?

For B2B agencies running cold outreach, the bottleneck isn’t sending messages. It’s knowing who deserves a message — and having enough context to make the first one worth reading.

Generic outreach fails because it’s generic. “Hi [Name], I help businesses like yours grow…” gets deleted in under a second.

Good outreach requires real research: What stage is this company at? What are they struggling with publicly? Who is the actual decision-maker? What’s the right angle for this specific person?

Done manually: 20–30 minutes per prospect. Done at volume: impossible.

What Does the B2B Lead Research Pipeline Actually Do?

Our AI lead research pipeline runs daily. Every morning it:

1. Pulls a batch of target companies from Apollo.io based on our ICP filters
2. Enriches each company with publicly available signals — LinkedIn presence, recent posts, hiring activity, funding stage, tech stack
3. Scores and filters leads against our ICP criteria automatically
4. Generates a personalised opening angle for each qualified lead
5. Pushes qualified leads to ClickUp with all enrichment data and the suggested first line
6. Notifies the team via WhatsApp that leads are ready to review

What used to take four hours now takes 20 minutes — and most of that 20 minutes is reviewing outputs, not producing them.

The Technical Stack: Apollo → Clay → Claude API → Make → ClickUp

Data Source: Apollo.io

We set up saved searches in Apollo filtered by industry, company size, job title, and technology used. Apollo exports these as CSV or via API on a schedule.

Enrichment Layer: Clay

Clay is the core. It pulls each company from the Apollo export and enriches it — LinkedIn company page, recent posts, hiring signals, news mentions. It also has a built-in AI column that writes personalised first lines based on all the enrichment data combined.

AI Reasoning: Claude API

For harder qualification decisions — “is this company actually in our ICP, or do they just look like it?” — we pipe the enrichment data through a Claude API call with a structured prompt. The model reasons through the ICP criteria and returns a confidence score with a brief rationale. This filters out leads that pass surface-level checks but wouldn’t convert.

Workflow Automation: Make (formerly Integromat)

Make ties everything together. It polls for new rows in Clay, applies our final filter criteria, pushes qualified leads to ClickUp as tasks with all enrichment data in the description, and sends a WhatsApp summary to the team each morning.

What We Learned Building This at Varnan

The prompt is everything. Our first ICP-qualification prompt was too broad. It was passing leads that were technically in our ICP but practically impossible to convert. We rewrote the prompt three times over the first two weeks. False positive rate is now under 10%.

Data quality determines output quality. If your Apollo search is poorly configured, the pipeline just automates bad research. We spent more time tightening the source filters than building the automation itself.

The personalisation angle is worth the effort. The AI-generated opening angles aren’t perfect — we edit about 30% of them before sending. But even the unedited ones outperform anything we wrote manually, because they’re grounded in actual research rather than guesswork.

Review doesn’t disappear — it shifts. We still have a human review every lead before active outreach. The difference: reviewing 12 pre-qualified leads in 20 minutes vs. researching 40 leads from scratch in four hours.

The Results: What Changed After 3 Months

Three months in:

• Lead research time: down from ~4 hours/week to under 30 minutes
• Qualified leads per week: up from 8–12 (what we had energy to research manually) to 25–35
• Reply rate on cold outreach: up ~40%, which we attribute to better personalisation quality

The time we recovered went back into follow-up, relationship building, and closing — the work that actually moves revenue.

Should You Build This Pipeline for Your Business?

If cold outreach is part of your acquisition strategy and you’re spending more than three hours a week on lead research: yes, this is worth building.

The barrier is lower than most people think. Clay handles enrichment without code. Make handles the workflow without engineering. The only real skill you need is knowing how to write a good prompt — learnable in a week.

The bigger risk is not building it. Competitors who’ve automated this are researching three times more leads with the same headcount. That compounds fast.

Frequently Asked Questions

How much time can AI lead research automation actually save?

In our case, it cut weekly research time from roughly four hours to under 30 minutes — and most of that remaining time is spent reviewing outputs, not producing them. The pipeline runs daily, pulling, enriching, scoring, and drafting outreach angles automatically, so the team’s only job is a quick review before sending.

Do you need to be a developer to build an AI lead research pipeline?

No. Apollo handles sourcing, Clay handles enrichment without code, and Make ties the workflow together with minimal setup. The only real skill that matters is writing a good qualification prompt for the AI step — and that’s learnable in about a week, not a specialised engineering skill.

Why does the AI qualification prompt need multiple rewrites to work well?

Because an overly broad prompt will pass leads that look right on paper but never convert in practice. Our first version did exactly that, so we rewrote it three times over two weeks until it reliably filtered out low-fit companies — bringing the false positive rate to under 10%. The prompt is the part of the system that determines whether the rest of the automation produces good leads or just automates bad research.

Does automating lead research remove the need for human review?

No — it shifts where review happens rather than eliminating it. A person still checks every lead before outreach, but instead of researching 40 prospects from scratch in four hours, they’re reviewing 12 pre-qualified, enriched leads with a drafted opening line in about 20 minutes. We still edit roughly 30% of the AI-generated opening lines, but even the unedited ones outperform what we wrote manually.

How do I know if my business is wasting too much time on manual lead research?

If cold outreach is part of how you find clients and your team spends more than about three hours a week researching prospects, it’s worth automating. The barrier to entry is lower than most people assume — the bigger risk is standing still while competitors who’ve automated this are working through three times as many qualified leads with the same headcount.