The client
A Spanish industrial catering company managing hundreds of B2B accounts across restaurants, corporate canteens, and institutional kitchens. Over 10,000 orders processed monthly through their MongoDB-based ERP.
The problem
Valuable patterns were invisible to the sales team. Which accounts were quietly ordering less month over month? Which products had untapped cross-sell potential given a client’s order history? Which seasonal patterns were being missed?
The data was there. The team was too busy processing today’s orders to analyze last month’s.
What we built
An autonomous AI sales agent that operates without human triggers:
- Connects directly to the production MongoDB database
- Decides autonomously what to analyze each week — not pre-built reports, but agent-driven investigation
- Identifies account-level patterns: declining volume, cross-sell opportunities, at-risk accounts, seasonal shifts
- Writes a structured weekly report with specific, named recommendations per account
- Posts every Monday morning to a dedicated Slack channel
Tech stack: Claude Agent SDK, MongoDB, Python, Slack API, VPS with weekly cron.
Key architectural decision: the agent receives a single generic mongodb_query tool rather than pre-built queries. This means it formulates its own database queries based on what it autonomously decides to investigate. It’s not a scheduled report — it’s genuinely exploratory.
The result
- +23% cross-sell revenue in the first 4 months after deployment
- Zero human triggers — fully autonomous weekly operation
- 9 hours/week saved of manual data exploration by the sales team
- 12+ at-risk accounts flagged before the sales team noticed, allowing proactive intervention
The agent now processes approximately 400 MB of operational data per week and delivers 15–25 specific recommendations each Monday — read and acted on by the sales team before their first meeting of the week.