Every Monday I pull the same numbers. Weekly conversions, cost per lead by channel, email open rates, landing page bounce. I paste it into a slide deck, write some bullets, and send it to my director before our 11am. It used to eat my entire morning because the hard part was never pulling data — it was figuring out what to say about it.
A few months ago I started doing something different. After I pull the numbers into a spreadsheet, I paste the table directly into ChatGPT and say something like: "Here's this week's marketing data next to last week. Tell me what got worse, what got better, and what didn't change enough to matter. Be honest if nothing interesting happened." That last line is important. Without it, AI will manufacture a story out of noise — it'll tell you a 2% change is "notable" when it's just Tuesday being Tuesday.
Once I get that back, I'll usually ask a follow-up like "my director is going to ask why paid search cost went up — give me three things to check before I answer her." That's way more useful than "provide insights." It gives me a checklist I can actually run through in our BI tool before the meeting.
The part that saved me the most time, honestly, was formatting. I used to agonize over how to phrase things on slides. Now I paste the analysis back and say "turn this into 4 bullets for a slide — each one should start with what happened and end with what we're doing about it. Keep it under 15 words per bullet." Then I copy-paste into the deck and I'm done. It's not magic. The AI doesn't know our business. But it's really good at organizing messy thoughts into clean sentences, and that's the part that was eating my morning.
One thing I learned the hard way: don't ask AI to explain why a metric moved unless you give it context. I once sent a conversion table and asked "why did signups drop?" and it confidently told me it was probably seasonality. It wasn't. We had a broken form for two days. Now I always add "don't guess at causes — just tell me what questions I should be asking the team." That one change made the outputs way more trustworthy. Explore our prompt library for more examples →