Your data is worth more than you think, and in worse shape than you think
Everyone wants AI. Almost nobody wants to look at the real state of their data, which is exactly what decides whether the AI will work.
There's a conversation that keeps repeating. A company wants to use AI for something specific, shows us the idea, and it sounds good. Then we ask to see the data it's supposed to run on, and the room goes quiet. It turns out the data lives in five different places, in formats that don't match, with fields each person fills in their own way.
AI learns from what you give it. If you hand it a file where the same customer appears three times spelled three ways, don't expect magic. What it does is repeat your mess at scale. The boring work of getting the data clean and in one place is what decides whether the rest is worth anything.
What's odd is that almost no company knows the real state of its data until it looks closely. People assume it's fine because the business runs. And the business runs because people are patching the gaps by hand, fixing what the system gets wrong, remembering exceptions written down nowhere.
That imperfect data is worth a lot of money, even if it doesn't look it. Years of orders and resolved incidents are exactly what would make an AI useful for your specific business, something your competitor doesn't have. The trouble is that the value is buried under the mess.
Before thinking about models, it's worth doing something less glamorous: pulling the data into one place and agreeing on what each thing is called. Then fixing the worst errors before going further. It's not exciting and nobody shows it off in a demo, but it's the work that makes everything else possible.
Whenever someone asks me for AI, before talking about models I ask something else first: 'can I see your data?'. The look on their face usually tells me how much road is left before the fun part.