I have read enough 'AI strategy' decks this year to start questioning the format. Most of them are forty slides of architecture, governance, and roadmaps. Almost none of them include a working prototype.
I keep thinking of this as a category error. Strategy in software does not get tested by Powerpoint. It gets tested by something running. Until something runs, the strategy is a hypothesis. The deck is just a hypothesis dressed up in confidence.
The pattern I have started to notice in companies where AI actually ships: every approved strategy was preceded by something the team built, badly, in a week. The ugly v0 is what makes the strategy survive contact with reality. The deck just records what the team learned from the v0.
Where AI does not ship: the deck arrives first. It is approved. Then the build team is told to make the slides true. By the time the prototype exists, the assumptions in the deck have already been overtaken by what the model can do today.
A strategy that has not survived a prototype is a wish. A strategy that has, is a plan.
This is uncomfortable for boards because it inverts the order they are used to. Approve the plan, then build. With AI, the order may have to flip. Build the smallest defensible version, watch what it teaches you, then write the plan around what is now real.
I am not saying skip strategy. I am saying that strategy and prototype might need to live in the same loop, not in separate phases. Every prototype updates the strategy. Every strategy update is tested by the next prototype. The companies that do this seem to end up with strategies their competitors cannot reverse-engineer.
What is the smallest version of your AI strategy you could actually run for a week?