Finding insights you can trust

1. Finding insights you can trust

You've cleaned your data. You've enriched it. Now comes the work every analyst wants to do: finding insights. AI gives you a powerful new way to begin — but it also makes it dangerously easy to mistake noise for signal. Let's look at how to use it well.

2. What AI is doing when it "finds insights?"

When you ask AI for insights, it isn't doing magic. It's running queries, computing aggregations, looking at distributions, and surfacing patterns it spots — often by writing Python or SQL and executing it on your data. Treat it like a fast junior analyst: prolific, helpful, in need of supervision. Everything we learned about interrogating AI's work in the data quality lesson still applies here.

3. Four useful starting points

Good insights don't come from "show me what's interesting." They come from giving AI a specific analytical lens to apply. Four starting points cover most early exploration: trends over time, distributions across categories, differences between groups, and outliers and anomalies. Pick the one that fits the question you're trying to answer.

4. Asking with intent

The framing of the prompt matters. "What's interesting about this data?" leaves it to AI to decide what matters. Specific prompts —

5. Asking with intent

"how has revenue trended over the last twelve months?", or "how is order value distributed by customer segment?" — put that decision back where it belongs: with you. AI does the work, but you decide what's worth investigating.

6. Exploration is iterative

The real power of AI in this phase is the speed of iteration. Take the Daily Grind coffee chain data. We start with "give me an initial overview" — Claude returns a dashboard, some data quality flags, and a few suggested follow-up questions. We pick one: "how do wait times vary across our stores?" — that gets us another dashboard plus narrative takeaways pointing out the slowest locations. We drill again: "which products show the biggest margin differences across stores?" — and we get more visuals, more observations. Each prompt is a starting point for the next question, not necessarily a final answer.

7. Signal vs. noise

Here's a trap. AI will surface patterns whether or not they mean anything. A "trend" can come from a sample that's too small. A "correlation" can be coincidence. A "difference between groups" can vanish when you control for something else. The harder skill in this phase isn't asking questions — it's deciding what to believe.

8. Four checks before you trust an insight

Before you act on anything AI surfaces, run it through four checks. First, the calculation: read the query AI generated. A missing filter or a wrong join can fabricate a pattern that isn't real. Second, the mechanism: is there a plausible reason this insight would be true? If you can't explain why, be skeptical. Third, stability: does the finding survive when you slice the data differently — across time periods, segments, or cohorts? Real patterns will probably hold up. Fourth, your own domain knowledge: does this match what you know about the business? A finding that contradicts years of operational experience is more likely a data problem than a hidden truth. If a finding fails any of these checks, treat it as a hypothesis worth testing — not immediately an insight worth acting on.

9. Let's practice!

Time to start finding — and interrogating — some insights!

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.