We’ve been living with that for years. As long as people retrieved the numbers themselves, they would unconsciously correct them. “Oh, you mean excluding returns.” That silent correction, based on experience and context, kept things manageable.
An AI agent doesn’t do that.
Ask an agent for the revenue, and you’ll get a single number, delivered with with total confidence, without the context that an experienced colleague would naturally provide. The agent doesn’t know that finance and sales use different definitions. It selects a table, sums up a column, and presents the result as if it were the truth. And because it looks so definitive, everyone starts to believe it. Your old definition problem hasn’t gone away. It’s just become faster and more confident.
Where your definitions live once and for all
A semantic model is where you define those terms. What is revenue, what is margin, when does a customer count as active? Written down correctly once, with the logic behind it, so that every report means the same thing.
The point is this: if you query that model instead of the raw tables underneath it, both humans and agents will get the same answer. The agent no longer invents its own definition. They use yours.
This changes three things:
Answers become reproducible. The same question always yields the same number, whether you ask it on Monday or Friday.
Trust scales accordingly. You trust the agent’s answer because you trust the underlying definition, not because it looks convincing.
Permissions remain in effect. The agent enforces the same row-level security as your reports. No one can see anything through the agent that they wouldn’t be allowed to see in a dashboard.
The pitfall
AI is not a shortcut to bypassing poor data hygiene. Unleashing an agent on a messy model only accelerates how quickly everyone comes to believe the wrong number. AI doesn’t make the proliferation of definitions any less problematic—it just makes it less visible. Get your definitions in order first, then bring in the agent.
That may sound like a damper on the hype, but it’s actually the opposite. Organizations that have their business definitions in order can safely deploy AI. The rest will lose trust faster than they gain it.
The next step isn’t chatting
The next phase of BI isn’t “chatting with your data.” It’s ensuring your definitions are correct, so it no longer matters who asks the question. A human, an agent, a new colleague opening your model for the first time: they’ll all get the same answer.
One definition of revenue. Even for your AI agent. Especially for your AI agent.
This blog post is a contribution from DataTako, the leading white-label distribution platform for Power BI. With DataTako, you can securely and scalably share dashboards and reports with employees, customers, and partners through your own fully branded portal. Find more inspiration at www.datatako.com or visit DataTako at Data Expo.