What the machine can't lead
At a large technology company, 97 percent of senior leaders said in an employee survey that they clearly understood the company's strategic priorities.
Then someone asked the people responsible for executing that strategy to name three of them.
28 percent could. Around 30 percent could not name a single one.
Donald Sull, Charles Sull and James Yoder published this in MIT Sloan Management Review after analysing 124 organisations. The measure was generous — any three of the top five counted as a match. The number is flattered, not harsh.
And note the date: 2018. Long before ChatGPT. This is not an AI finding. It is the ground AI is being rolled out on.
That's what makes it the opening of this chapter. Because AI does not lead. It accelerates what is already there — including the lack of clarity.
You can see the same thing from three other directions. BCG puts 70 percent of AI success with people, process and culture, and only 10 percent with the algorithm. Deloitte finds 84 percent of companies have not redesigned jobs or workflows around AI. McKinsey finds that among companies using gen AI, only 21 percent have fundamentally redesigned even some of their workflows.
Four independent measurements. One shape.
Scott D. Anthony names the mechanism behind it precisely: by the time you have enough data to justify a decision, it is too late to make it. A model can execute a priority at extraordinary speed. Setting one is not something it can do for you.
And where leadership does set and say it, it shows up in the numbers. EY surveyed around 1,150 employees in late 2025: where AI strategy was communicated clearly, 92 percent reported positive productivity. Where it stayed unclear, 62 percent.
So the missing half here isn't a technology gap. It is the half someone has to name before a system can carry it out. 🧭
Here is what changed. That unclarity used to travel through people. Middle managers translated it, asked again, quietly corrected it on the way. A system does not ask again. It executes the instruction it was handed, fast and at scale.
So the rule I'd defend: automate what people can name consistently. What they can't name consistently isn't ready for automation yet.
💬 Would you sign that rule — or is it too strict to run a business on?