Why AI Analytics Insights Are Often Wrong — How to Tell Grounded From Guessed

Short answer: a language model produces fluent text that sounds right, whether or not the numbers behind it are right. AI analytics is trustworthy when ordinary code computes the figures and the model only explains them, and when every insight links to evidence you can open. It is untrustworthy when the model is doing the arithmetic.

Table of Contents

1. Why Wrong Insights Sound So Convincing
2. Five Ways AI Analytics Goes Wrong
3. Eight Questions That Expose a Guess
4. What a Trustworthy Design Looks Like
5. How DeepSync Approaches It
6. Key Takeaways

Why Wrong Insights Sound So Convincing

A wrong chart looks wrong. A wrong paragraph does not. "Mobile conversions fell 18% week over week, driven by increased checkout friction on Android" reads identically whether the 18% is real or invented. People check charts and trust sentences.

That asymmetry is the core risk. A dashboard with a bug gets noticed. A plausible narrative gets forwarded.

Five Ways AI Analytics Goes Wrong

1. The model does the maths

Language models are not calculators. Given a table and asked for a percentage change, a model may produce a number that is close, or one that is not, with the same confidence.

2. Invented causes

Two things changed in the same week, so the model states that one caused the other. Analysts are trained to resist this. A model is trained to produce a satisfying explanation.

3. Small samples treated as trends

Conversion on a page with 40 visits moved from 5% to 10%. That is two people. A model will happily describe a doubling.

4. Missing context

The model does not know there was a public holiday, a deploy, a price change, or an outage, unless that information was supplied.

5. An answer for every question

Ask a leading question and you get agreement. "Why did the redesign hurt signups?" produces reasons, even if signups did not fall.

Eight Questions That Expose a Guess

Ask these of any AI insight, or of any vendor selling one.

1. Where did this number come from? A query, or the model?
2. Can I open the evidence? The sessions, events, or rows behind the claim.
3. How many visits is this based on?
4. What is it compared with? Last week, last month, a baseline.
5. What happens when there is too little data? Does it say so, or produce an insight anyway?
6. Is the same question answered the same way twice?
7. Is it describing or explaining? "Step two conversion fell" is a description. "Because the form is confusing" is a hypothesis.
8. Did anyone check whether acting on it worked?

What a Trustworthy Design Looks Like

The principle is narrow: let the model write, and do not let it count.

A second useful pattern is tool use. Instead of pasting every table into one giant prompt, the assistant chooses a specific query to run for the question asked and reports what the query returned. The figures then come from the database.

How DeepSync Approaches It

Pulse, the AI product manager on Business and Enterprise, follows that separation. In the words of the docs: deterministic detectors run over session and event data (rage-click spikes, funnel regressions, error surges, dead-end pages) and only then are those pre-computed findings handed to a model to narrate into a daily brief. The numbers come from code, not from the model.

• Each finding links to the sessions or events that triggered it.
• Findings can be reviewed, dismissed, or promoted, so a person stays in charge.
• After you mark something fixed, Pulse keeps watching the metric and reports whether it actually moved.

The AI Copilot on Pro and above is a tool-calling assistant. It queries only the data domain a question needs, such as funnels, errors, or geography, rather than guessing from a large prompt (Copilot docs).

The same caution shows up elsewhere in the product. The chat widget's uplift figure stays blank until its control group is large enough to mean something, on the reasoning that a fabricated number is worse than no number.

We wrote more about the engineering in how we keep AI insights from hallucinating.

Key Takeaways

• Fluent text hides wrong numbers better than a chart does.
• Models should explain, not calculate.
• Ask for the source, the sample size, and the evidence.
• Treat stated causes as hypotheses until you have watched sessions.

Conclusion

AI makes analytics faster to read. It only makes it more correct when the counting is left to code and the conclusions are left to you.

See DeepSync's AI features.