Everyone uses AI. Very few companies can operate it.

A company where 400 people use ChatGPT is not an AI-enabled company. It is a company with 400 personal productivity tools and an operating capability that may be zero, and the two get confused on board slides.

Applied AI, the version a CEO can put in a budget, is AI that runs inside a workflow with an owner, a log and a name on the output. Personal use gives you none of those, however good the answers are.

Why the distinction shows up in the P&L

Personal use improves an individual's week. It does not change unit cost, cycle time or the headcount plan, because it leaves when the person leaves and varies with who is at the keyboard. A CFO cannot underwrite it.

Operating capability changes numbers the business already reports. Cost per order processed. Days to close. Hours per quote. That is the version worth funding, and it is the version a lot of AI spending has not bought yet.

Three tests a CEO or COO can run

Take any AI workflow someone in your company is proud of. Run it through three tests.

1. The absence test. Does it run when the person who set it up is on leave? If the answer involves that person's laptop or that person's chat history, it is personal use.

2. The audit test. Can a manager who was not involved open a record of what it did last Tuesday, including which cases it handled and which it handed to a human? A habit has no log. A system does.

3. The signature test. Who signs off on the output, by name, and has that person read the failure cases? Personal use has no signature. Institutional use has one, and the person whose name it is can describe the wrong answer they worry about most.

A workflow that passes all three can be budgeted, audited and handed to a successor. A workflow that fails any one of them is a productivity boost for one employee, which is fine, and is not an operating capability.

Run the three tests across the ten AI workflows your team mentions most. The count that pass is your institutional AI capability. A count of zero is possible, and it is useful information at planning time, because it tells you the licenses bought the first kind of use and not the second.

What the 2026 AI Index says, and what it does not

Shana Lynch's summary of the 2026 Stanford AI Index for Stanford HAI carries the numbers behind the gap. Source facts:

• Generative AI reached 53% population adoption within three years, faster than the personal computer or the internet.
• Global corporate AI investment hit $581.7 billion in 2025, up 130% from the prior year.
• Employment among software developers aged 22 to 25 has fallen nearly 20% since 2024.

The Index does not count how many companies pass the three tests above. That framework is mine, not Stanford's. My read of the first two numbers: adoption at that speed and spend at that scale should have produced visible operating change by now, and the distance between the two says much of the money bought personal use.

The third number is the one I keep returning to. The easy read is that AI took junior jobs. The read I trust more is that AI removed the tasks companies used to hire juniors for, and the training path those tasks provided went with them. That is a hypothesis built on a correlation, not a finding. If it holds, the people reviewing AI output in five years learned by approving rather than by building.

Moving one workflow across the line

Find the person on your team who gets the most out of AI on their own. Ask for the three things they do with it weekly. Pick one.

Write down what they do, in order, including the judgment calls they make without noticing. Then build the version that works for the next person without them in the room: a place it runs, a log a manager can open, a name on the output. Run it through the three tests.

Give the writing-down step to a junior person. Extracting undocumented rules from an operator and turning them into a specification is the skill that holds its value as models improve, and it is learnable at 23.

The count is the capability

Personal use is worth encouraging. Operating capability is the number of workflows that pass three tests, and a board should ask for that number before it asks for the license count.

An executive question worth asking this week: of the ten AI workflows your team mentions most, how many pass all three? I am collecting answers, and the count interests me more than the names of the tools.