Do we need clean data before starting AI?
No. You need a few hundred representative real examples for the target workflow and permission to use them. The first build often includes the data plumbing for that workflow.
Language models are built to handle messy text, so perfect data is rarely the blocker. What matters is that you can export a few hundred real examples of inputs and the decisions or outputs that followed, and that your policies allow you to use them with an AI provider. A prototype on a random sample of real data reveals the edge cases early, which is far cheaper than a data-cleaning project that runs first.