Your Nulls Are Lying to Your AI
There’s a problem hiding in your data that your BI team learned to handle years ago. Your AI tools haven’t.
It’s your null values. And they’re more dangerous now than they’ve ever been.
What your BI team knew
When a business analyst ran a report, they worked inside a translation layer built by people who understood the data. A null in the revenue column meant “not applicable” — excluded from averages by a rule someone wrote deliberately. A null in the status column meant “pending” — counted separately, never dropped silently.
That knowledge was encoded in the BI layer. Calculated fields, business rules, documented definitions. The nulls were handled. The analyst never had to think about them.
What AI doesn’t know
When Copilot, Databricks Genie, or ChatGPT queries your data directly, it bypasses that layer entirely. It sees the raw column. It sees the nulls. And it has no idea what they mean.
A column with 30% null values doesn’t pause the AI. It doesn’t ask for clarification. It calculates an average across the remaining 70% of rows and returns a confident, fluent, authoritative answer — representing less than three quarters of your data, with no warning, no caveat, and no indication that anything is missing.
That’s not an AI problem. That’s a data problem the AI inherited without knowing it.
The three types of missing data — and why they all hurt AI differently
Missing data isn’t random. Statisticians have known this for decades:
MCAR — Missing Completely At Random. The nulls have no pattern. A sensor glitched. A form field was optional. This is the least dangerous type — the missing data doesn’t tell you anything. AI tools will still drop the rows silently, but at least the sample isn’t systematically biased.
MAR — Missing At Random. The nulls are related to other observed variables. Customers who didn’t provide income data tend to be in a specific age bracket. The AI doesn’t know that. It drops those rows and produces results that systematically underrepresent a segment of your population.
MNAR — Missing Not At Random. The most dangerous type. The reason data is missing is related to the missing value itself. Patients who didn’t report a symptom may have been too sick to complete the survey. High-value accounts may have opted out of data collection. When AI drops these rows, it’s not just incomplete — it’s systematically wrong in a direction that matters.
Your AI tool doesn’t distinguish between these types. It sees a null and either drops the row or ignores the field. The bias is invisible in the output.
What this means for your AI queries
Ask an AI tool “what is the average revenue per customer?” on a dataset where high-value customers opted out of data collection — MNAR — and the AI will return a number. It will be wrong. It will be confidently wrong. And without knowing the null pattern, you have no way to know it’s wrong from looking at the answer.
This is the gap that traditional reporting never exposed because the BI layer caught it first. AI analytics tools skip that layer. The gap is now yours to manage.
What to do before you connect
Before you connect any dataset to an AI tool, you need to know:
- Which columns have significant null rates
- Whether the nulls are likely random or systematic
- What null means in each column — excluded, pending, unknown, or something else
- Whether a high null rate in a critical column will silently distort AI outputs
These aren’t questions you can answer by looking at the AI’s output. You have to answer them before the AI sees the data.
That’s exactly what AI Data Maturity surfaces — a column-by-column null rate based on a statistically valid sample, with a warning when null rates are high enough to affect AI reliability. Not after you’ve connected your data. Before.
Because by the time your AI tool tells you the answer, it’s already too late to ask about the nulls.