Your Agent Needs One Definition of Revenue. You Have Three.

A founder asked me last month why his AI agent kept flagging the wrong accounts as at-risk. The pilot was supposed to watch revenue and surface customers slipping toward churn. Instead it named accounts that were perfectly healthy and missed two that weren't. He assumed the model was bad. The model was fine. It had pulled "revenue" from his CRM, which tracks bookings. His billing system tracks what was invoiced. His GL tracks what was recognized. Three numbers, three systems, three answers for the same word. The agent picked one and acted on it.

The blocker isn't the model

This month Experian launched something it calls an Agent Operating System, a trust and semantic layer that lets AI agents share one definition of the data they act on. Strip the branding off and it's solving a dull problem. Agents can't agree on what a number means, so they can't be trusted to act on it.

That gap is where everyone is stuck. 99% of companies plan to put agents into production. 11% have. The difference isn't model quality. Survey after survey points to the same culprits: data siloed across systems, weak lineage, no agreed source of truth. Roughly half of finance teams say wiring their data into agent workflows is the hard part.

For a $5M company this lands harder, not softer. You don't have a data team to paper over it.

A semantic layer is just a decision nobody made

"Semantic layer" sounds like infrastructure. It's really a question. What does this number mean, and who gets to decide. At most growing companies the answer is nobody, and the meaning drifts by system.

Take "active customer." Your CRM counts anyone not marked closed-lost. Billing counts anyone with a paid invoice in 90 days. Product analytics counts anyone who logged in. Ask three people for the active count and you get three numbers, all defensible, none of them wrong. A person reconciles that in their head without noticing they're doing it. An agent can't. It grabs a source and runs.

An agent acts on whatever definition it finds first. If you haven't decided what the number means, the tool decides for you, and it won't tell you which one it chose.

Decide before you deploy

The work that makes an agent useful isn't prompt engineering. It's sitting down and writing the definition for the ten or fifteen numbers that actually drive decisions. Revenue means recognized, not booked. Active customer means paid in the last 90 days. Gross margin means these costs, counted this way. One definition, one owning system, written down where the agent reads it.

Do that and the agent gets reliable fast. Skip it and you've automated the speed at which a wrong number reaches a real decision.

Pick your ten numbers. Decide what each one means. Then turn the agent on.