Measure: Test Your AI Before Your Customers Do

You wouldn't roll out a new payroll system without testing it. AI deserves the same suspicion, maybe more, because it fails in ways that look confident. A chatbot that invents a refund policy sounds just as sure of itself as one quoting the real one.

Measure is the NIST framework's answer: evaluate your AI before it meets the public, and keep evaluating after. That means checking accuracy on real examples, checking whether it treats different groups of people differently, and watching for drift — the slow slide when the world changes and the model doesn't.

You don't need a lab. Start with twenty realistic test cases and a spreadsheet. Run them monthly. Write down what you find, including the uncomfortable results, because a test you won't look at is theater. Testing you actually do beats a perfect testing program you never start.

I explain frameworks; I don't give legal counsel.