You Don't Have an AI Problem. You Have a Data Problem.

A company reached out wanting to build an AI-powered reporting system. They wanted natural language queries against their business data — ask a question in plain English, get a chart and an insight. Sounds great. The technology exists. It's very doable.

Then I looked at their data.

Customer records in three different systems with no shared identifier. Revenue data in spreadsheets that got manually updated every Friday. Product categories that meant different things in different departments. Date formats that varied by who entered the data. And a "data warehouse" that was really just a MySQL database with 200 tables, half of which were abandoned experiments from three years ago.

You can point the most powerful AI in the world at this data. It will give you confident, well-formatted, completely wrong answers.

Garbage In, Confident Garbage Out

The old saying is "garbage in, garbage out." With AI, it's worse. It's "garbage in, confident garbage out." The model doesn't know your data is messy. It doesn't know that the revenue numbers in Table A don't match Table B because they use different fiscal calendars. It just picks one and presents it with authority.

This is arguably more dangerous than having no AI at all. At least without AI, someone has to manually pull the data and might notice the discrepancy. With AI, the discrepancy gets buried under a polished natural language response that nobody questions.

What "Good Data" Actually Means

When I say "fix your data first," I don't mean it needs to be perfect. Perfect data doesn't exist. But it needs to meet a minimum bar:

1. Single source of truth. For any given metric, there should be one authoritative source. Not three spreadsheets that sort of agree. One table, one definition, one calculation.
2. Consistent schemas. Dates should be dates. Numbers should be numbers. Categories should be standardized. This sounds basic because it is basic — and yet it's the number one issue I see.
3. Documented lineage. Where does this data come from? How often is it updated? What transformations happen between source and destination? If you can't answer these questions, neither can your AI.
4. Accessible architecture. The data needs to be queryable. If your "data warehouse" requires a DBA to write custom SQL for every question, AI isn't going to magically fix that access problem.

The Right Sequence

Here's what I recommended to that company, and it's what I recommend to most businesses that come to me wanting AI analytics:

Phase 1: Data foundation. Consolidate sources, standardize schemas, build a proper warehouse. This isn't exciting. It doesn't demo well. But it's the foundation everything else depends on. Typically 4-8 weeks depending on complexity.

Phase 2: Traditional analytics. Build dashboards and reports on the clean data. Let the business validate that the numbers are correct. This is your quality gate — if the dashboards show wrong numbers, AI would too. Fix it here where it's visible and debuggable.

Phase 3: AI layer. Now add the natural language interface, the automated insights, the predictive models. They're working with clean, validated, well-structured data. The AI can actually be useful because the foundation supports it.

Most teams want to skip to Phase 3. I get it — Phase 3 is the exciting part. But Phase 3 without Phase 1 is a demo that falls apart the moment someone asks a question you didn't rehearse.

The Payoff

Here's the thing that makes this approach worth the patience: once the data foundation is solid, everything gets easier. Not just AI — everything. Reports are faster to build. Questions are easier to answer. New tools integrate cleanly. And when you do add AI, it works reliably because it's built on something trustworthy.

That company I mentioned? They spent six weeks on data consolidation and warehouse setup. Then two weeks on dashboards. Then when we added the AI query layer, it worked on the first try. Not because the AI was special — because the data was ready.

That's the unsexy truth about AI in business. The model is the easy part. The data is the hard part. And the companies that get the data right first are the ones whose AI actually delivers value.