Most 'AI projects' are data projects in a costume
If you can't answer three plain-English questions about your data, no model is going to save you, a field guide for anyone told to add AI this quarter.
Every few weeks someone asks us to “add AI” to something, a workflow, a product, a support queue. The budget is approved, the timeline is set, and the expectation is that the hard part is the model.
It almost never is.
The costume
Here’s the pattern. A company decides it needs an AI feature, so it frames the work as an AI project: pick a model, write some prompts, wire up an API. That framing feels like progress, because it’s the part everyone can picture.
Then the work actually starts, and the model turns out to be the easy 20%. The other 80% is data, where it lives, what shape it’s in, who owns it, whether it’s even true. The “AI project” was a data project wearing a costume.
Pull on that thread and most data problems turn out to be infrastructure problems wearing a costume too. The data is messy because the systems that produce it were never wired together. Start at the bottom.
Three questions
Before you scope an AI feature, you should be able to answer these in plain English, no dashboard, no caveats:
- Where does the data live, and is there one copy of it? If the answer involves three systems and a spreadsheet someone maintains by hand, that’s the project.
- Who decides when it’s wrong? Every useful AI feature makes judgments. If no human owns “that output is wrong,” you’re shipping a liability, not a feature.
- Could a careful person do this today with the data you already have? If the answer is no, the model won’t change that. AI is leverage on a process that works, it doesn’t invent the process.
If those answers come easily, the AI part is genuinely straightforward. If they don’t, you’ve found the real work, and it’s usually worth more than the feature you came in asking for.
Why we’ll sometimes tell you to wait
This is why we occasionally talk a client out of the project they walked in with. Not to be difficult, because spending the budget on a model sitting on top of broken data is the most expensive way to learn this lesson.
The good news: the work underneath isn’t wasted. Cleaning up where your data lives, wiring the systems together, giving outputs an owner, that pays off whether or not the AI feature ever ships. You end up with a business that’s ready for AI, instead of one that bought a costume.
Start at the bottom. It’s almost always the bottom.