Context
Niche: B2B SaaS platform (international market).
Client: requested anonymity (NDA).
Starting point: 3 sales reps, a steady flow of inbound leads, qualification taking two business days. Only a small share of them reached a meeting.
Goal
Cut time to first contact and automate qualification: initial classification, profile enrichment, calendar booking — without sales involvement.
What we did
1. Lead data map
Defined required signals for qualification: company size (Clearbit/Apollo), industry, contact role, tech stack (BuiltWith). 14 scoring parameters.
2. AI agent on LangGraph
Architecture: orchestrator → 4 specialized agents (research, scoring, outreach, calendar). Uses GPT-4o + Claude 3.5 for different tasks. Decision logging for auditability.
3. Integrations
HubSpot CRM (lead writes), Cal.com (booking), Slack (notifications), Sentry (agent error monitoring).
4. Testing on 200 leads
Parallel A/B: half of leads via the agent, half via reps. Compared conversion, meeting quality (sales NPS), time to first contact.
5. Gradual rollout
Month 1: 30% traffic. Month 2: 70%. Month 3: 100%. Reps freed for mid- and large-deal negotiations.
Results
Цифры по проекту мы не публикуем. Часть данных клиент просил не раскрывать, а проценты без доступа к его системам ничего не доказывают: рост с двух визитов до шести тоже выглядит внушительно. На встрече показываем проект с доступом к Метрике и Вебмастеру — и объясняем, за счёт чего пришёл результат.
Timeline & budget
3 months build + 2 months A/B validation. Stack: TypeScript + LangGraph + Vercel. LLM costs ~$700/mo at 3,000 leads.
What was tough
First two weeks the agent confused roles (startup CEO vs enterprise CEO). Fix: extra LinkedIn check + company size. Also added escalation: if agent's confidence is below 70% — the lead is handed to a rep.