When I built an AI agent to process invoices, the demo I put online won't let you upload your own invoice. People assume that's a missing feature. It's the most deliberate decision in the whole project.
A real invoice is confidential. It carries a vendor's banking details, account numbers, pricing nobody outside the deal should see. The moment you send it to a public demo running on a third-party AI model, you've handed that document to a company whose data-retention policy you've never read, to sit on a server you don't control. So the demo runs on synthetic data. You get to watch the agent think without a single real record leaving anyone's building.
That decision didn't come from a compliance manual. It came from the same instinct that runs through everything I build: figure out what the system isn't allowed to touch before you figure out what it can do.
Most conversations about AI governance and data security have this backwards. They treat it as a review step, something legal handles at the end, a box you check before launch. By then it's too late to matter. If the confidential data is already flowing through the model, a policy document doesn't pull it back. Governance isn't a gate at the finish line. It's a set of choices you make while you're drawing the thing up.
I see the same pattern in less obvious places. I built a mobile app that household-goods (HHG) surveyors use to inventory a home before a move. It captures what's inside people's houses, room by room. That's about as personal as operational data gets. The real question there was never "are we compliant." It was who can see this, where does it live, how long do we keep it, and what happens to it when the job is done. Those aren't legal questions bolted on afterward. They shape how you build the app in the first place.
Here's the part small and mid-sized businesses miss: you are already making these decisions, whether you name them or not. Every tool you adopt makes a choice about your data by default, and most of those defaults were written to benefit the vendor, not you. Doing nothing isn't staying neutral. It's accepting someone else's governance and calling it your own.
The operator's version of this is simple, and it's the same discipline I bring to any process. Know what data you actually have. Know where it goes. Decide, on purpose, what your tools are and aren't allowed to do with it. Then build to that. It doesn't demo well, but it's the difference between a system you can stand behind and one you're hoping nobody asks about.
I've been working through Microsoft's AI Governance series, starting with its AI Strategy and Transformation Planning course, and I joined the International Association of Privacy Professionals. Not because a certificate changes how I work, but because these questions are getting harder faster than most businesses are ready for, and I'd rather be fluent in the frameworks than improvising them.
The technology keeps getting more capable. The governance question doesn't get easier as the tools get better. It gets more important. The businesses that will trust AI with real work are the ones that decided, up front and on purpose, what it is allowed to touch.
That isn't the exciting part. It's the part that lets you sleep.
See what "synthetic data" looks like
The AP invoice agent from the top of this post runs live in your browser, on synthetic data, on purpose. You can watch it work without a single real record leaving anyone's building.
See the AP Invoice Agent →