Data first, interface second

Every few weeks someone shows me a slick AI interface bolted onto a business and asks why the results are mediocre. The answer is almost never the interface, and almost never the model. It's what's underneath: the data was a mess before AI showed up, and AI didn't fix it. It just made the mess faster and more confident.

Good interfaces don't fix messy, incomplete, or conflicting data. Start underneath.

AI runs on your context

An AI system is only as good as what it can see. If your customer records live in three systems that disagree, if your pricing logic exists mostly in one person's head, if "the process" is whatever the last person who did it remembers — then the most capable model in the world is reasoning from fog. You won't get insight out of it. You'll get fluent guesses.

This is why I push back when companies want to start with a chatbot or a dashboard. The interface is the last mile. The first mile is unglamorous: what are the actual things your business runs on — customers, jobs, orders, properties, claims, whatever your nouns are — and where does the truth about each of them live?

The questions that matter before any tool

• What are the nouns? The handful of things your business is actually about. Most companies have fewer than ten that matter.
• Where does each one live? One source of truth, or four half-sources and a spreadsheet named FINAL_v3?
• How do they relate? Which customer goes with which job, which invoice, which conversation. The relationships are usually where the value hides — and usually what nobody recorded.
• What's only in someone's head? That's your most expensive data store, and it has a single point of failure with a vacation schedule.

The good news

This work used to be brutal — months of migration projects and data-entry weekends. AI has changed that, in a pleasant bit of irony: the same technology that needs clean context is remarkably good at helping you create it. Extracting structure from old documents, reconciling conflicting records, drafting the standing definitions nobody ever wrote — the cleanup itself is now a fraction of the effort it used to be.

So the sequence that works: pick one workflow that matters, get its data genuinely solid, and then put AI in front of it. The companies that invert this — interface first, foundation later — end up demoing things that can't survive contact with a real Tuesday.