zcom.

zcom is a multi-tenant e-commerce platform. Merchants could see their sales, but couldn't answer the questions that matter — am I making money, what's my break-even, is my marketing paying for itself? There was also no AI on the dashboard at all.

Layer 1 of ZCOM AI: a metrics engine that computes about 50 business metrics per store on a schedule; a Business Insights tab that renders them at zero AI cost; Ask ZCOM, a dashboard chat that answers questions from those metrics; per-plan AI limits; a super admin AI section for usage and cost; and metering for the zcom MCP connector so merchants' own Claude uses the same tools and limits.

Trust and cost. The model never sees raw orders and never does arithmetic — fixed, tested formulas produce the numbers, and Claude only explains them. Every store's data stays inside its own tenant, and answers respect each staff member's permissions, so an owner's profit figures never reach someone who can't see profit.

On cost, merchants see questions per week rather than tokens, with a global daily spend cap behind every plan.

Node.js services behind a federated GraphQL gateway, with a database per store in MongoDB. The metrics engine runs as scheduled jobs and writes snapshots; the chat uses the Claude API with a cached, byte-stable system prompt built from the snapshot, plus small read-only tools. A daily “What to do next” summary runs through the Batch API.

Layer 1 gives merchants decisions instead of graphs, at a predictable cost to the platform. It also lays the data groundwork — bought-together and return-reason aggregates, attribution and customer segments — for the next layer: a storefront assistant and semantic product search.