Do we trust the numbers our AI tools produce — and if not, why not?

Most AI distrust traces back to inconsistent definitions, not flawed models. When the AI and the CFO disagree on revenue, the problem is usually that they are using different definitions of revenue.

If we asked an AI to calculate our profit margin, would it know which costs to include and which to exclude?

Without a semantic layer, the AI has to guess which cost categories apply. The semantic layer removes the guessing by codifying exactly what each metric means in your business.

What would it take to make our data definitions consistent enough that AI agents could use them without human translation?

The investment is in documenting and codifying your business definitions — what each metric means, which source systems are authoritative, and how edge cases are handled. Once codified, every AI agent inherits the same understanding.

Data & Analytics Intelligence

Semantic layer as AI trust foundation

By Mark Ziler · Last updated 2026-04-05

The semantic layer is what makes AI trustworthy on your data. When you ask an AI agent "what was our margin last quarter?" the semantic layer tells it: margin means gross_margin_pct, it is stored as a percentage not a decimal, it lives in the financial table not the jobs table, and it should be filtered to completed jobs only. Without this, the AI guesses — and confidently gives you the wrong answer. With it, the AI answers like your best analyst who has been at the company for ten years and knows every quirk of the data. The semantic layer is not a technical luxury. It is the difference between an AI you can trust and one you cannot.

Go deeper

Your regional director asks the AI agent: 'Show me our worst-performing location by margin.' The agent returns Location 7 at 12% margin. Your director knows Location 7 is actually your best performer. What happened? The AI used gross revenue instead of net revenue in the margin calculation because nobody specified which one 'margin' means in your business. The director tells her team the AI tool is unreliable. Word spreads. Six months of investment in AI-powered analytics gets shelved because of a missing definition.

The trap most companies fall into is investing heavily in the AI interface — the chatbot, the voice agent, the fancy UX — while underinvesting in the semantic layer underneath. It is like building a beautiful house on a cracked foundation. The AI will confidently answer every question. Confidence is not accuracy. Without a rigorous semantic layer, you have built a system that is articulate and wrong — which is worse than a system that is slow and right, because people act on confident wrong answers.

Questions to ask

Can we produce a document that shows, for every metric in our system, the exact calculation, the source table, and the business owner who approved the definition?
What is our testing process before we let an AI agent answer questions about a new metric?
If the AI gives a wrong answer, can we trace exactly which definition or mapping caused the error?