Every ML system should start on one page.
The Machine Learning Canvas is a one-page tool for laying out what to predict, what decision it feeds, and what a mistake costs, before anyone builds anything. I'm Louis Dorard, and I created it in 2015. We run it live with you, on the project you're scoping right now.
The project is yours to land. The definition isn't yours alone to fix.
Three weeks in, the team still doesn't agree what counts as churn, who acts on the score, or what a false alarm costs. You can't ship past a disagreement like that, and you can't settle it on your own: the people who own those answers have to be in the room.
You inherited the scope, and you're the one accountable for it landing.
"AI-driven churn prediction" meant five different things to the five people who approved it.
It shipped. Nobody agrees whether it worked.
1 page. 10 boxes. Everything that has to be agreed before anyone builds.
From what's being predicted through to what happens after the system ships.
Filled in together, in one place, before the first line of code.
What running the canvas gets you
-
Precise enough to work from directly, instead of reconstructing it later from a deck and a chat thread. Contracts and schemas are the next step, once your data people are in the room.
-
Left with a one-line prompt, an assistant has to guess at the details that decide whether an ML system works, and nothing in the prompt tells it what to guess. A canvas-derived spec answers them directly: what's knowable at decision time, what counts as success.
-
What a false alarm costs, what a miss costs, and the accuracy a model would have to reach before it's worth building at all. That last number is a decision you can take before anyone writes code: if nothing realistic clears the bar, you've saved the build. Checking whether a real model clears it comes later, once there's one to test.
-
A consultancy building anomaly detection for a US telecom manufacturer stopped the modeling after several rounds: the problem wasn't defined well enough to build against. Fixing the definition first is what got the system live.
-
What counts as a hit, who acts on it, what threshold to use. Those normally surface halfway through a build. Settle them first and NYK Group's R&D lead describes what happens: the whole process from use case idea to deployment gets expedited.
-
The canvas holds the intent, data, decision and monitoring documentation that EU AI Act, SR 11-7 and ISO/IEC 42001 reviews ask for. Written during scoping, instead of reconstructed months later.
Scales from a single project to a department’s portfolio
"Louis offers a unique systems perspective, so that you can see how all the pieces of the puzzle fit together. This is ideal for consultants, or for people who have to plan or manage a machine learning project in their organization."
— Dr Zygmunt Szpak, co-founder, Insight Via Artificial Intelligence (2020)
Bring the project you're scoping
20 minutes on it, and you'll know where it's likely to break.
Before you book
What do I actually get?A few hours of live time on a project you're already scoping. We arrive with a first draft, so the time goes on judgement calls, not blanks. Price comes on the call.
What do I keep?The filled canvas is yours, and your client's numbers stay yours. We sign your NDA. Co-branding on client-facing work, and your team can get the method too.
Why not just fill it in myself?A finished-looking canvas can still be wrong. "Reduce churn" passes any review. But if nobody named who acts on the prediction, seven of the ten boxes quietly break. That's what we catch in the room.