The map is not the territory

A company connected an AI agent to their carefully built analytics model and asked it a normal-sounding question: which customers are we about to lose?

It answered in a second, with a confident list. The list was wrong. Not because the agent was dumb, and not because the data was dirty. The model simply had no idea what "about to lose" meant. It had revenue, a customer table, a dozen dashboards. It did not track churn (when customers leave), or risk, or the steps that lead there.

The model was built to answer the questions someone already had. The agent was asking a new one.

A model is built for the questions you already had

A semantic model, a warehouse, a pile of tables: these are maps. A good map is a real achievement. But every map makes a choice about what to include and what to leave out, based on the trips it expects. Star schemas (a common table layout) are drawn for aggregation: sums by region, counts by month. They do those trips very well.

The trouble starts when someone asks for a trip the mapmaker never planned. "Which customers are hit by a freezer breach" is not a sum by region. It is a walk across freezers, shipments, orders, and customers. If the map never drew those roads, the answer is not on it.

Why this hurts AI agents the most

A person handles a missing road with judgment. They know "about to lose" means something, so they go ask around and work it out from experience. An AI agent has none of that. It takes the map literally. Ask it a question the map cannot answer, and it does not stop and say so. It finds the nearest-looking road and drives, confidently, to the wrong place.

That is the whole reason this path exists. To give an agent data it can actually reason over, you need something closer to the territory: not just tables, but the real things in your business and how they connect. That is an ontology, and the graph it forms.

Try it on a real model. Ask the map a few things and watch where it stops.

The one trap, and where we go next

The trap is believing a better map is the answer. Cleaner tables, more measures, one more dashboard: all still maps, all still drawn for questions someone already had. The shift is not a bigger map. It is modeling the territory itself, the real things and how they connect, so an agent can find its own way to a question nobody drew a road for.

The first move is the smallest and the most important one. Before you can model how things connect, you have to decide what counts as one thing. That is an entity.