AI Canvas: structuring an AI decision across six dimensions

The business cases for AI initiatives that reach approval tables suffer from a formal disease: they are documents built to convince, never to decide. Thirty pages where the expected value occupies the cover, the risks sit in an appendix, the dependencies do not appear because nobody looked for them, and the committee receiving them can only choose between trusting and postponing. The AI Canvas was born as a countermeasure to this form: one page, six mandatory dimensions, and the rule that a dimension left blank counts as an alarm, never as a detail to fill in later.

The underlying idea is that the quality of a decision depends on the structure with which it is prepared, and that for AI initiatives the right questions are known and number six. Let us walk through them in filling order, which has a logic of its own.

The first three: what, what can go wrong, how much it matters

The opportunity opens the Canvas and demands a precise discipline: the business problem before the technical solution, with the metric that would improve and the signature of whoever owns that metric. "Reduce complaint handling time by 30%, measured by the head of operations" is an opportunity; "use generative AI on customer service" is a technology in search of employment, and the Canvas rejects it by construction.

Risks occupy the second box with the same rigour, in four families: system error and its impact on customers or people, the regulatory profile (the use case's AI Act classification, the personal data involved with its legal basis), supplier dependence, and adoption risk, meaning the concrete possibility that the system works and nobody uses it. Priority, the third dimension, positions the initiative inside the portfolio we wrote about: the score on the three axes, the time horizon, what it displaces or postpones in the backlog, because every yes said to one initiative is a no said to another and the Canvas forces you to write down which.

The second three: who, with which means, after what

Stakeholders are the dimension traditional business cases take for granted and failures do not: the executive sponsor with signing power, the process owner ceding the process, the users who will have to change habits, the control functions (risk, legal, the DPO whenever personal data enters the picture) to be involved at the beginning, when they cost hours, rather than at the end, when they cost months. Capabilities census the gap between what is needed and what exists, across data, skills and infrastructure, with an explicit choice for each gap: build, buy, or ally.

Dependencies close the loop and are the dimension that most often overturns decisions: what must be true for the initiative to start (that cleaned dataset, that integration, that approved policy) and what this initiative makes possible afterwards, because a mediocre use case that unlocks a reusable foundation can be worth more than a brilliant one that leaves nothing behind. It is the box where the portfolio stops being a ranking and becomes a sequence.

How it is used, and what it produces

The Canvas performs best filled in session, two or three hours with sponsor, process owner, technical lead and one control function in the room, because the value lies as much in the final sheet as in the disagreements that surface while filling it: the session in which it emerges that the estimated value has no signatory, or that the data dependency was known to everyone except the sponsor, has already paid for the method.

The output feeds directly into the go/no-go criteria: from opportunity and risks come the success and stop thresholds, from capabilities the true budget inclusive of prerequisites, from dependencies the sequence. And the comparison between different Canvases, all on the same structure, is what allows a committee to decide between initiatives instead of between rhetorics.

In our AI Strategy work the Canvas is the instrument through which every backlog initiative is prepared before asking for capital, and the completed version enters the deliverables together with roadmap and decision criteria: calendly.com/fabiolalli/zerofive, or hello@zerofive.ai. If you want to test the method without waiting, take the most promising AI initiative in the house and try filling the six boxes in an hour, with the right people in the room: the box that stays blank is telling you something, and it is telling you now, when it costs little, instead of mid-project. Which of the six, shall we bet, will it be?