The AI Operating System

The brand had grown past the point where any one person could hold the operating picture in their head. Inventory sat in the ERP. Customer service ran on Gorgias. Growth metrics lived in Shopify and Meta. Finance reconciled in NetSuite. Logistics talked to a 3PL. Project work moved through Asana and Slack. The leadership team spent the first hour of every Monday rebuilding context from scratch, and most decisions still waited on whoever was closest to the source.

We started with the operating model, not the tools. The first job was mapping which decisions actually got made each week, who owned the underlying data and where the friction sat. That work surfaced a hard truth: the problem was not a missing dashboard. The problem was that there was no shared knowledge layer the systems could read and write to without fighting each other.

The build started there. A knowledge graph indexed the operational records that mattered (customers, orders, SKUs, campaigns, suppliers, tickets) and kept them current with scheduled syncs from each source of truth. On top of that, we shipped 40-plus custom skills, each one a narrow capability that knew how to read the graph, perform a specific job and write the result back. None of the skills were generic chat. Each was scoped to a workflow the team already ran by hand.

Executive briefings were the first big payoff. Every morning at six, a scheduled run pulled overnight movement across orders, inventory, ad spend, returns and service queues; the daily briefing landed in the leadership channel before stand-up. The weekly review followed the same pattern with a longer horizon and a written commentary. The team stopped rebuilding the picture every Monday because the picture rebuilt itself.

Governance had to be designed in from day one. Every skill writes to an audit log. Every write-back to a system of record is reversible. The knowledge layer carries provenance: every fact knows where it came from and when it was last refreshed. Without that discipline, the operating system would have become another shadow source the team could not trust.

What survived in production over the first six months tells the real story. The skills that earned their keep were not the most ambitious; they were the most boring. A skill that drafted the supplier follow-up email got used every week. A skill that turned a CS ticket cluster into a product-team digest replaced a meeting. The team learnt to ask for small skills and we shipped them in days, not months.