The daily struggle to get answers
You need to know which regions drove sales growth last quarter. Not the standard monthly report, but broken down by product line, compared to last year, excluding that one-time bulk order that skews everything.
In most companies, this means fighting with Excel, waiting days for IT to run a custom query, or letting it go because it's too much hassle.
What if you could just ask?
The Underappreciated Use Case
Most AI conversation focuses on writing and coding. Data analysis is just as powerful but gets less attention.
But people who discover they can have a conversation with their data have that "where has this been?" moment. The data is already there. You already have questions. This just makes it drastically easier to connect the two.
With correctly set up AI coding tools, you describe what you want to know. AI writes the code. You see the results. Not quite right? Refine it: "Now break it down by quarter." Done.
This isn't automation of existing processes. It's a different relationship with data. From "what can I analyze with the tools I know?" to "what do I want to know?"
Why Read-Only Analysis Works
Here's what makes this practical: you're asking AI to read and analyze data, not modify it.
Writing code that modifies data is risky. AI makes mistakes. But reading? Apart from temporarily overloading the database, the worst case is a wrong answer you can verify. No damage done. And the database risk can be mitigated.
This removes the main barrier to using AI with actual business data. You're not trusting AI to be perfect. You're using it for mechanical work while you provide judgment.
The Real Problem: Getting the Data
This sounds great until you hit the actual bottleneck: getting machine-readable data.
Most corporate data lives in systems that give you PDFs and formatted reports, not CSVs you can analyze. The data exists. Getting it out in a usable format ranges from "fill out a request form and wait" to "not happening."
Start Where You Can
The pragmatic path: start with whatever you can actually get.
Most systems let you export something, even a clunky CSV download. That's enough.
Point an AI coding tool at it (Claude Code, Codex, Devstral or many more). Tell it what you want to know. AI reads the file, writes analysis code, shows results.
"Show me top 10 customers by revenue, excluding refunds." Table appears. You refine: "Now show monthly revenue for each over the year." Chart appears. Customer X spiked in March. "What drove the spike for Customer X?" AI filters and breaks down line items. Single large order for Product Y.
Three minutes. The alternative: an hour in Excel or a week waiting for BI.
Building Your Case for Better Access
Starting with CSV exports is strategic. You prove value locally first.
You answer questions faster. You share insights that wouldn't have happened because they were too much effort. People notice.
Now when you ask for better data access, ideally API credentials, database read permissions, you're not asking for a hypothetical capability. You're showing how to make something that already works even better.
IT teams are very cautious about access for good reasons. But if you're already doing useful work with limited access and understand what you're doing, the conversation changes.
What About the BI Team?
BI teams build dashboards everyone needs regularly. Standard metrics, reliable infrastructure.
Conversational analysis is for questions that don't fit standard reports. The one-off investigation. The follow-up to something on a dashboard. The quick check before a decision.
BI builds the roads. You're walking off-road to explore something specific.
Most BI teams are backlogged managing existing tools. If you self-serve for specific questions without creating support burden, that's usually welcome. The key is communication and complementing official metrics, not competing with them.
What This Changes
When analysis goes from "might take hours" to "ask and get an answer in seconds," different things become possible.
You investigate hunches you'd let go. You follow up on anomalies you'd accept. You answer questions in meetings instead of promising to circle back.
The limiting factor shifts from "do I have time?" to "do I have the data?"