How a B2B Agency Cut Manager Time by 72% with an AI Lead Qualification Agent
- Before
- 75 min
- average time to qualify one lead
- After
- 21 min
- only final check and call
- Lead throughput
- +45%
- more leads handled, same team
- Implementation
- 11 days
- audit to production
Context and challenge
Account managers spent evenings manually triaging inbound leads from 4 channels (form, LinkedIn, email, Telegram). Each lead needed ICP qualification, company activity check, services relevance. 60-90 minutes per lead, 8-15 leads per day. Hot leads waited too long and went to competitors.
Approach
Split the task into 4 agent pipeline stages:
- Collection — webhook on form, IMAP on email, polling on LinkedIn/Telegram, all funneled into Postgres with a standard JSON schema
- Enrichment — for each source company we pull public data (website, LinkedIn profile, recent news)
- Scoring — Claude 3.5 analyzes the profile against ICP criteria and returns a structured summary via a Pydantic schema
- Routing — based on score, lead lands in the right channel (hot → Slack to manager, warm → CRM with week-later reminder, cold → automated email series)
Each step idempotent, with retry logic. If LinkedIn API drops during enrichment, the lead still moves forward with a limited dataset.
What was hard
Hard part 1 — heterogeneous sources. Form gives structured data, LinkedIn almost nothing, email free text. Solved via universal lead JSON schema in Postgres and per-source adapters.
Hard part 2 — false positives. First week the agent flagged a couple of large potential clients as “cold” due to atypical company descriptions. Added a manual override flag — manager can mark a lead as “always show” and the agent becomes more lenient on similar profiles in the future.
Hard part 3 — privacy. Client works with financial companies, can’t send full email threads to Claude API. Solved via local preprocessing — extract only metadata (sender, subject, keywords) and send those structured data points, not raw text.
Business impact
3 months after launch:
- Avg time per lead: 75 → 21 minutes (−72%)
- Team throughput: +45% more leads/day
- Lead → demo conversion: +18% (hot leads get called faster)
- Managers leave on time — subjective but important retention signal
"Managers stopped spending evenings on lead qualification and started closing more deals."
Tech used
Case FAQ
- What data did the client need to provide?
- ICP criteria (target customer — revenue, industry, team size), history of last 50 qualified leads for model calibration, and channel access (form, LinkedIn API, email IMAP).
- Who makes the final decision on a lead?
- Always a human. The agent only prepares the summary and assigns priority (hot/warm/cold). The manager glances at the summary and decides whether to call. Keeps control + reduces cognitive load.
- What if the AI gets priority wrong?
- First month we calibrated on your data. Manager likes/dislikes summaries, we update criteria. After calibration, priority accuracy stabilized at 87%.
- What if our CRM isn't Slack?
- No problem. Sales summary can go to Bitrix24, AmoCRM, HubSpot, Salesforce, any CRM with an API. Manager gets notified in their preferred tool.
Want the same result?
Free 2-day audit. In half an hour we'll know if there's reason to move forward.