Not Every Problem Is a Recipe

For three years, McDonald's and IBM tried to perfect voice-ordering AI for the drive-thru. They kept iterating: better models, more training data, tighter integration. By 2024 the system had plateaued around 80-85% order accuracy — worse than the average human headset worker. Then TikTok found it. Videos went viral of the AI adding bacon to a customer's ice cream, or piling 260 chicken nuggets onto a single order while a driver tried, increasingly desperately, to correct it. McDonald's ended the partnership within weeks. Not after a boardroom cost-benefit analysis. After the internet laughed.

Here's what I think McDonald's actually got wrong, and it wasn't the engineering. They treated drive-thru ordering like a recipe: a fixed, complicated-but-solvable procedure that better ingredients — more data, better microphones — would eventually perfect. But a drive-thru window isn't a recipe. It's a conversation, with sarcasm, background noise, a toddler yelling in the back seat, a customer who changes their mind twice. That's not complicated. That's complex: the kind of problem where more precision applied to the wrong category of thinking doesn't get you closer to a solution, it just gets you a more sophisticated failure.

The organizational theorist Ralph Stacey mapped this territory with two questions: how certain are we about cause and effect, and how much agreement exists about what we're even trying to do? High certainty, high agreement is a treadmill — you already know the workout, you just have to run it. Low certainty, low agreement is mountain biking — the trail changes under you in real time, and the skill that matters isn't following a plan, it's reading the terrain as it happens.

Most AI rollouts I watch get pitched as a treadmill and executed as a treadmill — a project plan, a rollout date, a success metric locked in before anyone rides the actual trail. McDonald's engineers were extremely good at running the treadmill. Nobody on that team, it seems, was asked to notice they were actually mountain biking.

From Dancing in the Wind: An Adaptive Leadership Framework in the Age of AI.

Before your next AI initiative gets a project plan, ask the more honest question first: is this a recipe we can perfect, or a trail we have to read as we go?