How We'd Build an AI Lead Qualification Agent for B2B Sales

Every B2B sales team has the same dirty secret: their highest-paid people are spending the majority of their day on work that has nothing to do with selling.

When a new lead submits a form, here is what typically happens next. A rep or SDR opens the contact in the CRM. They Google the company name. They check LinkedIn for the contact's role and background. They scan the company's website for headcount, industry, and product. They check if there has been any recent news — a funding round, a product launch, a hiring surge. They cross-reference all of that against their ideal customer profile to decide whether this lead is worth pursuing. Then they draft an outreach email that references something specific enough to not sound generic. Finally they log all of this activity in the CRM before moving to the next lead.

The deeper problem is consistency. Research quality depends entirely on who does it and when. A thorough SDR on a slow morning does excellent work. The same SDR on a Friday afternoon with 12 leads in the queue does approximate work. An AI agent does the same quality of research on lead 1 and lead 40, at 3am on a Sunday, in under 2 minutes.

Here is the exact technical blueprint Codewingz would follow to build this system for a B2B company receiving 20 to 200 inbound leads per week. Every architectural decision below is deliberate and explained — not a default choice.

Stage 1 — Trigger & Data Capture

The pipeline starts the moment a lead submits any form — your website contact form, a webinar registration, a product sign-up, or a reply to a cold email sequence. We would use webhooks to feed every trigger event into a normalisation layer that extracts the core data points: full name, company name, email domain, job title, and any free-text input they provided. This normalised payload fires the orchestration agent.

We would also handle inbound leads from Calendly (demo booking), LinkedIn message exports, and direct API calls from your outbound tool. Every lead source feeds the same pipeline. The entry point does not matter — the quality of what comes out the other end is consistent regardless.

Stage 2 — Enrichment Engine

The agent then fires enrichment requests in parallel — not sequentially, which cuts total enrichment time from 60 seconds to under 15. The enrichment stack we would typically use:

Apollo.io provides company data (employee count, revenue estimate, industry classification, technology stack) and contact information (verified email, LinkedIn URL, direct phone where available). Clearbit adds funding history, tech stack signals, and growth indicators. A web search agent — built with LangChain's web search tool — scans for news from the past 90 days: funding announcements, product launches, executive changes, partnerships, or press coverage. This is the research that requires judgment and cannot be pulled from a static database.

All enrichment results flow back into a structured JSON object that becomes the context for every subsequent stage. The agent knows the company size, the contact's seniority, whether they just raised money, what technology they currently use, and what problems that technology typically creates — before a human has seen the lead.

Stage 3 — ICP Scoring Engine

We would build a configurable ICP scoring model — not an ML model, but a transparent weighted scoring system that your sales team can understand, audit, and adjust without engineering involvement. Here is a representative scoring framework:

Scores 70–100 route to your senior AEs with a "high priority" flag. Scores 40–69 route to SDRs for follow-up. Scores below 40 go into a low-touch nurture sequence automatically — a personalised email sequence that keeps the relationship warm without burning selling time. This tiering is fully configurable and adjustable by your sales leadership without code changes.

Stage 4 — Personalised Email Draft

We would use GPT-4o for the draft generation step, not Claude, because GPT-4o's instruction-following is particularly strong for constrained writing tasks where the format must be exact. The system prompt would encode your brand voice, your email structure preferences, character limits, and the specific types of personalisation that your team has found effective.

The model receives: the contact's full enriched profile, their score breakdown, any recent company news, and your product's documented pain point map — the specific problems your product solves that correlate with the signals found in the enrichment data. From this it generates a first email that references something real and specific, not a generic opener. A lead whose company just raised funding gets an email that references the growth challenge that typically follows a fundraise. A lead whose tech stack includes a tool that creates the specific problem your product solves gets an email that names that tool and the problem.

The draft goes to the rep for review, not to the customer directly. One-click send if it looks right. Easy editing if it needs adjustment. The rep's judgment is the final gate — the AI is doing the research and the first draft, not making the sending decision.

Stage 5 — Routing, CRM Logging & Alerts

The moment the pipeline completes, three things happen simultaneously. First, a Salesforce (or HubSpot) record is created or updated with the full enriched profile, the score, the scoring breakdown, and the draft email stored as a note. Second, a Slack alert goes to the assigned rep with a summary card: name, company, score, one-sentence summary of the most relevant finding, and three action buttons — view profile, send draft, or reassign. Third, the complete pipeline run — every enrichment call, the scoring rationale, and the draft — is logged in LangSmith for quality review and continuous improvement.

• We would not let the AI send emails autonomously. The draft is for the rep to review and send. Fully autonomous outreach removes the human judgment that catches the 5% of cases where the personalisation misfires — and one bad email to the wrong person at the right company causes more damage than 50 perfectly crafted emails cause good.
• We would not use a black-box ML scoring model. If your sales leadership cannot look at a score and immediately understand why a lead scored 82, they will not trust the system and will stop using it. The scoring model must be transparent. A weighted rule system your team can read and adjust is more valuable than an ML model that scores 2% more accurately but nobody understands.
• We would not enrich synchronously. Calling Apollo, then Clearbit, then running a web search in sequence takes 45–90 seconds. Running them in parallel takes under 15. We would always design the enrichment stage with parallel API calls and async results assembly.
• We would not hardcode the ICP criteria. Your ideal customer profile changes as you learn which customers succeed. The scoring weights must live in a configuration file that your sales ops team can adjust without a code deployment.
• We would not skip fallback logic. Apollo will occasionally return no data. The web search will occasionally find nothing relevant. The agent must handle missing data gracefully — scoring with what it has, flagging data gaps to the rep, and never producing a blank draft because one enrichment call returned empty.

• Research time per lead drops from 35–60 minutes to under 2 minutes. The agent does not get tired, does not cut corners on lead 40, and never skips the news search because it is Friday afternoon.
• 3× more qualified conversations per rep per week because reps are spending their time on selling, not on researching. The capacity freed by automation converts directly to pipeline.
• Consistent scoring across all leads regardless of volume, day, or time. A lead submitted at 3am on a Sunday gets the same quality of research and the same scoring rigour as a lead submitted at 9am on a Monday.
• Reply rate improves by 20–35% on AI-drafted outreach versus generic templates, because the personalisation references something real — recent news, specific technology, or a known industry pain — rather than a generic opener.
• CRM data quality improves automatically. Every lead enters the CRM with a complete, enriched profile rather than just a name and email. Pipeline reporting, forecasting, and territory analysis all improve as a downstream effect.