AI-Powered CRM Agent
An always-on agent that enforces deal hygiene, enriches contacts, and chases follow-ups inside the CRM, so the revenue team spends less time on manual upkeep and more time selling.
My Role
Challenge
A busy BD team's CRM naturally drifts out of shape over time. Salesforce and Pipedrive captures data effectively but lacks synchronous "block save" hooks to enforce proper processes.
Dirty Pipeline Data: Deals get pushed to later stages before they are ready, and newly created deals are often missing required fields.
Missed Opportunities: Contacts are often missing vital emails, follow-ups slip through the cracks, and client emails sit unanswered.
Revenue Impact: Manual upkeep and dropped balls quietly cost the business valuable time and revenue.
Process
Iterative AI Development: Developed collaboratively with Claude Code as an engineering pair, favoring durable architecture, git-backed continuous deployment to Railway, and tests alongside features.
Engine Architecture: Built a pure, side-effect-free rule engine (
evaluate(context) -> { allowed, failures[] }) isolated from I/O. This allowed the entire business rulebook to be mapped one-to-one to the spec and verified by fast unit tests.Hybrid Orchestration: Combined real-time event-driven webhooks for stage-gate validation and enrichment with strict scheduled cron jobs (timezone-aware) for AI task generation and SLA nudges.
System Hardening: Implemented strict safeguards based on real rollout incidents, including a self-write loop guard (to prevent infinite revert loops), rate-limit circuit breakers, notification deduplication, and retry backoffs.
Shadow Testing: Created a global
DRY_RUNswitch to log external writes as no-ops, successfully shadow-testing the bot's behavior against the live CRM before turning on live writes.
5
Automated Capabilities
24/7
Pipeline Observability
10+
Hours Saved Weekly on Manual CRM Keepup
Results
I built a highly observant, always-on CRM Agent that successfully runs five core capabilities in production without failing silently.
Stage-Gate Validation (Auto-Revert): Automatically enforces a detailed rulebook. If a deal moves to an unqualified stage, the agent instantly auto-reverts it and notifies the owner via Slack and deal notes with precise instructions on how to fix it.
Intelligent Contact Enrichment: Seamlessly integrates with lead enrichment tools to find and attach verified work emails based on role and company size whenever a relevant organization is added.
AI-Generated Follow-ups: A daily sweep reads recent correspondence and utilizes Claude to propose deduplicated, high-confidence follow-up tasks with due dates inferred from email urgency.
Unanswered-Email SLAs: Automatically flags deals where client emails sit unanswered past designated SLAs (24-48h depending on pipeline stage), creating actionable tasks and pinging owners
