Week 6 · Beyond
One number, three meanings
Reconcile revenue, then make the definitions machine-readable.
The brief
Thursday, 16:00. Tomás sends a one-line email: "Board deck says revenue 7.77M. Sales deck says 6.84M. Which one do I put on the slide?"
Neither is wrong. One is the sum of transactions (cash movements, including renewals and upsells). The other is the sum of closed_won deal amounts (bookings). They diverge exactly where a business is healthy: existing customers paying more. But if nobody writes that down, an AI agent someone is piloting for "revenue questions" will pick whichever definition it finds first.
That is the ontology gap, and it is your last deliverable: a reconciliation that explains every dollar of the difference, a running revenue total by month, and one capstone query that computes the whole metric tree from a single set of definitions.
This week: reconcile, then write it down in a form both people and agents can consume. You are no longer fixing a dashboard. You are defining what Meridian means when it says a word.
Lessons
Read
Ontology as Operational Infrastructure for AI Agents
As LLM agents act autonomously, the ambiguity humans used to resolve silently (what is 'revenue'? what is a 'customer'?) becomes a hard failure point.
Agent Observability ≠ Agent Correctness
Most teams running AI agents have monitoring (did it run?) but not evaluation (was it right?) — and standard software testing assumptions don't hold for agents.
Lean AI: Matching Model Size to Task Difficulty
Small language models can be ~100x cheaper than large ones per conversation; mature teams route by difficulty rather than defaulting to the biggest model.