PI coding agent: GTM Lead-to-Meeting Pipeline with Claude Code v2

This workflow uses PI coding agent with MCP (Model Context Protocol) tool access to rebuild a broken lead-to-meeting pipeline from end to end. PI coding agent reads the data warehouse (Snowflake) to identify new leads based on ICP filters, writes qualified leads to Salesforce CRM, enriches each record via Clearbit and Apollo.io, scores leads on a 0-100 fit-intent matrix using GPT-4o, routes leads to the correct sales rep based on territory and product line, and drafts personalized outreach sequences. The agentic reasoning step is the routing decision: Claude evaluates lead attributes (company size, industry, tech stack, recent funding events) against the sales team's capacity, win-rate history, and existing pipeline coverage to determine which rep gets the lead and what outreach angle has the highest probability of response. It runs unattended every morning via a GitHub Action cron trigger. Human approval gates are built into the pipeline at two points: outbound email send approval and meeting booking confirmation. Teams report 40-60% reduction in lead-to-first-touch time.

BUSINESS PROBLEM

Most B2B companies lose 50-80% of leads between initial identification and first sales touch. A 2025 Gartner report found that only 25% of leads are ever contacted by sales within the first 5 minutes, and response time is the single largest predictor of qualification success. For a typical Series B SaaS company generating 200-400 leads per month, the manual pipeline involves: an SDR spending 2 hours pulling a list from Snowflake, 30 minutes per lead on enrichment and research, territory assignment via spreadsheet, and outreach drafting from templates that all sound identical. At a fully loaded SDR cost of $85,000/year, the manual pipeline costs $2,500-$3,500 per week in labor just for triage. Worse, data entry errors in CRM (wrong owner, wrong contact info, incomplete enrichment) cause 15-20% of leads to fall through cracks that nobody notices for weeks. An agentic pipeline that reads directly from the warehouse and writes to CRM eliminates both the labor cost and the data integrity problem.

WHO BENEFITS

GTM operations teams at Series B to Series D SaaS companies running complex sales motions with multiple product lines, territories, and lead sources. Sales development managers overseeing teams of 5-20 SDRs who waste 30-40% of their week on data entry and territory wrangling instead of calling leads. Revenue operations leaders who are accountable for pipeline hygiene metrics but cannot enforce data standards across a manual handoff process. Each role shares the core need: automate the data-heavy front of the funnel so humans can do the relationship work.

HOW IT WORKS

1. Warehouse Scan (6:00 AM daily). A GitHub Action triggers PI coding agent at 6:00 AM EST. Claude connects to Snowflake via MCP and runs a SQL query against the leads table filtering for: created in last 24 hours, ICP score above 60, not already in Salesforce. Output: a JSON array of 10-30 new leads.
2. CRM Write and Deduplication. PI coding agent calls the Salesforce API via MCP to upsert each lead. It runs a fuzzy deduplication check against existing contacts using email domain and company name similarity. Duplicates are logged to a Slack channel for manual review. Output: Salesforce lead records with a status of Raw.
3. Lead Enrichment. For each new Salesforce record, Claude calls Clearbit API for company data (headcount, industry, funding) and Apollo.io for contact data (direct dial, email verification, role history). It enriches the Salesforce record with the combined data. Output: enriched Salesforce lead records.
4. Fit-Intent Scoring. PI coding agent sends the enriched lead data to GPT-4o via API. GPT-4o scores each lead on two axes: fit (how well the company matches ICP on a 0-50 scale) and intent (buying signals detected from Apollo intent data or recent funding on a 0-50 scale). The combined score determines the lead tier: Hot (80+), Warm (60-79), or Nurture (below 60). This is the core reasoning step.
5. Routing Decision. Claude evaluates each Hot and Warm lead against the current sales team workload (pulled from Salesforce opportunity records) and territory definitions. It assigns the lead to the rep with the lowest active deal count in the relevant territory. The assignment is written to Salesforce and posted to a private Slack channel. Human can override within 1 hour or the assignment auto-accepts.
6. Outreach Drafting. For assigned leads, Claude drafts a personalized outreach sequence: a LinkedIn connection request, a cold email, and a follow-up email. Each draft pulls specific details from the enrichment data (recent funding, role change, mutual connections). Drafts are posted to the assigned rep's Slack DM for review. Checkpoint 1: human approves the email before send.
7. Send and Track. Once approved, Claude sends the email via the sales engagement platform API (Outreach or Salesloft) and creates a task in Salesforce to follow up in 3 days. It monitors open and reply rates and escalates to the rep if a Hot lead opens an email but does not reply within 48 hours.
8. Eval Suite Logging. Every action is logged to a structured audit table in Snowflake with timestamps, tool used, input/output hashes, and error status. A weekly eval report runs every Monday comparing pipeline metrics week-over-week.

TOOL INTEGRATION

PI coding agent (CLI): Main orchestrator. MCP is configured via a .mcp.json file in the project root. Gotcha: MCP tools are callable in any order, but the workflow fails silently if Snowflake MCP returns an empty result set. Add a validation step that checks result set size > 0 before proceeding. Snowflake (MCP): Read-only queries against the leads and audit tables. The MCP-Snowflake connector requires a Snowflake user with SELECT on the relevant schema. Gotcha: Snowflake MCP does not support multi-statement queries. Each query must be a single SQL statement. Salesforce (via REST API MCP): Write operations for lead creation and update. Requires OAuth 2.0 with refresh token. Gotcha: Salesforce API write limits are 1,000 records per hour on Enterprise Edition. If the warehouse batch exceeds this, splits the insert into chunks with a 3-second delay between chunks. Clearbit API: For company enrichment. API key at dashboard.clearbit.com. Free tier: 50 lookups/month. Paid: $0.01/lookup. Gotcha: Clearbit returns null for bootstrapped companies without strong digital footprints. Add a fallback to Apollo.io company data when Clearbit returns incomplete records. Apollo.io API: For contact enrichment. API key at apollo.io/settings. Free tier: 1,000 credits/month. Gotcha: Apollo's email verification endpoint returns a verification status but does not guarantee deliverability. Verified emails still bounce at 1-3% rates. Slack API: For human approval gates and notifications. Gotcha: Slack's Block Kit has a 4,000-character limit per message block. Long outreach drafts need to be split into multiple blocks. GitHub Actions: Cron trigger for the daily run. Gotcha: GitHub Actions routinely has 2-5 minute scheduling delays on the free tier. The 6:00 AM cron may trigger at 6:03-6:05 AM. Do not set hard SLA expectations on trigger time.

ROI METRICS

1. Lead-to-first-touch time: industry average 42 hours → projected under 15 minutes for Hot leads. Measurable in week 1. 2. SDR hours on data work: 20-30 hrs/week → 2-4 hrs/week for exception handling only. 3. Lead volume processed per day: 10-15 leads manually with enrichment gaps → 30-50 leads fully enriched, scored, and routed. 4. CRM data error rate: 15-20% of leads have incorrect owner or missing enrichment → under 2% with automated CRUD. 5. Pipeline velocity: 60-90 days from lead to SQL → projected 35-50 days with sub-5-minute first touch. (Source: Gartner Lead Response Management Study, 2025.)

CAVEATS

1. MCP tool failures cascade: If the Snowflake MCP connection fails, the entire pipeline stops with no partial processing. The workflow does not gracefully fall back to a CSV import. Monitor MCP connection health with a heartbeat check. 2. Over-enrichment risk: Enrichment APIs return large payloads that can bloat Salesforce record sizes. Salesforce has a 128KB limit per record. Claude does not truncate automatically. Cap enrichment field count to the top 15 data points. 3. Cost spikes from enrichment volume: Clearbit and Apollo bill per lookup. A batch of 50 leads costs $0.50-$5.00 depending on enrichment depth. At 1,000 leads/month, enrichment alone costs $10-$100/month. Add to the eval suite. 4. This workflow does not handle inbound lead routing from web forms, chatbot conversations, or event attendee lists. It processes only leads that already exist in the data warehouse with an ICP score.