Ask most engineering leaders about their AI strategy and you'll get a list of what they bought: "We rolled out Copilot to 200 devs, we have an enterprise Claude account, and two teams are piloting an internal agent." That's not a strategy. That's a receipt.
A strategy is a set of answers, not a set of purchases. And the uncomfortable truth is that most teams can't answer the questions a real strategy requires.
The Four Questions a Strategy Must Answer
Which models, for which tasks? "Whatever the default is" is how you end up running an expensive frontier model on work a cheap model could do. A strategy names the tiering: fast/cheap for bulk work, expensive/capable for the hard 10%, and a rule for when something graduates.
What does good look like? Not "the model is SOTA." Good means: it passes your evaluation on your codebase. The benchmark that matters is yours, and if you don't have one, you're adopting models on faith and marketing materials.
What's the budget, and who owns it? Cloud spend got a FinOps function. AI spend, which is growing faster and is far more volatile, mostly gets "we'll look at the invoice later." That later is usually a bad surprise.
What's governed? Which agents can act autonomously, what data they can touch, and what happens when they blow past a threshold. Governance is not bureaucracy — it's the difference between an agent and a liability.
Subscriptions Are Inputs. Decisions Are Strategy.
Every one of those subscriptions is an input — a capability you've purchased. Strategy is the part where you decide how the capability gets used, measured, and constrained. You can buy all the inputs and still have no strategy, the same way buying a gym membership isn't a fitness plan.
The good news: the raw material is cheap. One afternoon of writing down answers to those four questions will put you ahead of 90% of teams currently "doing AI." The bad news: without those answers, you're not doing AI. You're renting it.