What's actually true, what to fix first, and what you can send to your boss on Monday.
There's the one you drafted at 9pm, where you wrote down what's actually going on with your data. And there's the one you sent the next morning, which was true in a different way.
The gap between those two documents is the reason I built this.
If you're the person who owns AI at a company too small to have a data team, you've been handed a mandate and a set of expectations that came from somewhere else. A conference. A newsletter. A competitor's press release. And you're the one who has to say what's possible, on a timeline that was set before anyone looked at what's underneath.
You're not confused about AI. You're stuck on it. Those are different problems, and almost everything on the market is built for the first one.
I've spent fifteen years doing this work from the inside, and I've run the enterprise version of this exercise plenty of times. It asks about model registries and deployment pipelines and whether your CMO and CIO have a formal partnership. Run that at a twenty-person company and half the questions don't apply, the scale tops out at something only a Fortune 500 can reach, and a well-run small team scores badly at everything and gets told it's in crisis. That's not a diagnostic. That's a company being penalized for its size.
So I rebuilt it from nothing, starting from the ways this work actually falls apart. Definitions living in one person's head. Consent that stops traveling the moment data leaves the system it was collected in. Pilots that die because a priority got reopened.
The top of the scale is what a small team can genuinely reach. Not enterprise-grade. The standard is whether something holds when you're out for three weeks.
Everything is scored on what's operational today. Not what's planned, budgeted, or in progress. Plans get captured, but they don't count.
It can tell you that you're fine. An assessment that always finds a problem is a sales tool with a scoring rubric attached, and you've almost certainly sat through one. This one has a real outcome where the answer is start now, here's the narrow place to start.
Five dimensions, each named for something you want more of.
Whether there's a specific outcome this is supposed to change, whether anyone knows the current number, and whether a real person's week gets different if it works.
Whether you can get to your own data without a vendor ticket, whether your tools agree with each other, and whether you have enough history to prove something changed.
Whether marketing, product, and clinical mean the same thing by the same word. This is the quiet one. It costs nothing to ignore right up until the moment you point a model at it.
What would hurt if it left, whether consent travels with a record into the tools that use it, and what's already touching your data that nobody formally approved.
Whether decisions stay made, whether there's permission to stop something, and what breaks if you're out. No other assessment asks this, and it's frequently the dimension that decides whether any of the rest happens.
There's no single overall score. You get five, and they don't get averaged into one number, because one number turns into a target. "You're a 2.8" invites your boss to ask how we get to a 4, which is the wrong conversation. Five separate readings force a real one.
One page on what's true, in plain language, written to you. It names what's working with the same specificity as what isn't, because you're about to walk into a room and you need to know which ground holds.
Three fixes. Never more than three. Each one with the reason the other order breaks, a first action small enough to take next week, a named owner, and an observable condition that tells you it's finished. A list of nine things is a roadmap, and you already have one of those.
What to leave alone for now, what has to be true before it's worth doing, a date to revisit it, and language you can use when it comes up again. Because it will come up again.
The same findings, sequenced for leadership. It opens with what can start and when. This is the one that closes the gap between your 9pm document and your 9am one, and it's the reason the whole thing exists.
It's built for you if you own the AI question at a company under about fifty people, you're technically capable, and your real problem is sequencing and internal alignment rather than education. I work most often with healthcare and digital health teams, where the consent questions carry more weight, though the failure patterns aren't specific to one industry.
It isn't built for you if you need convincing that AI matters, or if what you want is someone to build the thing. I don't do implementation. I do the part that comes before it, and the part where you have to explain it.
You already know what's going on. This gives you something to point at.