How to Tell If Your Data Is Ready for AI

Language models keep getting better and cheaper. What hasn't changed is that they answer with what you give them: if company data is duplicated, poorly defined or spread across places that don't talk to each other, the AI will give very confident, very wrong answers. And a wrong answer delivered confidently is worse than none, because someone will use it to make a decision.

Before connecting ChatGPT, Claude or your own agent to your data, ask yourself these seven questions.

• Is your data in one place? If answering "how much did we sell this quarter by channel" means manually joining the CRM, GA4 and the ERP, the AI will have the same problem.
• Is there one definition for every number? If "sale" means one thing to finance and another to sales, the AI will pick one without telling you.
• Are the definitions written down? A model can't guess that "active customer" means someone who bought in the last twelve months. It needs a semantic layer or, at the very least, a document that says so.
• Do you trust the quality? Duplicates, empty fields and impossible dates go unnoticed in a report and get amplified in a conversation.
• Does each person only see their own data? If a rep can't see another rep's accounts in the CRM, they shouldn't be able to ask the AI about them either. Permissions have to travel with the data.
• Where does personal data end up? You need to know whether data leaves your environment, who processes it and how that fits GDPR.
• Do you know what you want to ask? "Adding AI" is not a use case. "Letting reps see in seconds which accounts have had no activity for a month" is.

If you said yes to almost everything, you're in a good place for a first, narrow use case, such as a sales copilot or conversational analytics.

If the first three fail, the work starts before the AI: unifying sources and writing definitions. That's what we do in data strategy, and it isn't wasted time: the same organised data also improves your dashboards.

If permissions or privacy fail, don't connect anything yet. It's the one thing you can't safely fix afterwards.

Many AI projects start by choosing the model or platform and only then look for something to do with it. It works better the other way round: first the business question, then the data needed to answer it, and finally the tool. That way the first project is small, measurable and, if it works, easy to expand.