On February 28 of this year, demonstrators gathered outside OpenAI's offices in London, and Pew Research Center put a photograph of them at the top of a report it published on September 17.1 The picture is scenery. The date is the useful part, because Pew's interviewers were already working by then. They had started on February 8 and stayed in the field until May 13, reaching 42,151 adults across 36 countries, with American data coming from two separate panel surveys.1 One of the things they asked was what artificial intelligence would do to the number of jobs in that country over the next twenty years.

In 34 of the 37 countries, people leaned toward fewer jobs.1 Among the eighteen high-income countries in the set, a median of 55% said AI will mean fewer jobs over that horizon, against a median of 36% across the eighteen middle-income countries, where far more people said they were unsure, 34% against 22%.2 In Australia, South Korea and the United States, Pew puts the share expecting job loss at around seven in ten or higher, and the American figure has climbed seven percentage points in two years.1 The fieldwork ran in the spring and the report landed in September, the normal lag for work this size, so read them as spring readings.

Read as public mood, this is one more entry in a long run of AI anxiety coverage, with nothing in it for an operator to do on Monday. It is also an operating condition, and in that form it interferes with numbers you are already reading every month.

The people generating your adoption data

Somewhere in your organization there is a tool pilot with a reporting line attached. Weekly active users, prompts per person, seats claimed, hours estimated. Somebody assembles those into a slide for the steering committee. Every figure on that slide was produced by a person who already holds a private theory about why the tool showed up, and Pew has now measured that theory at population scale in the countries where most of your staff live.

The finding I would put in front of a COO is the one about awareness. In around half the countries surveyed, the people who had heard or read a lot about AI were more likely than the less informed to say it will cost jobs.1 Those are your power users. The enthusiasm you are counting comes disproportionately from people who understand the technology well enough to think it is going to take work away from somebody. Pew also found, on a different question in the same survey, that higher awareness went with feeling more optimistic about AI in daily life.1 Those two results sit together without contradiction. The first question asks about outcome and the second asks about feeling, so a person can be cheerful about the technology and still expect it to reduce headcount.

An adoption rate collected while the purpose is unstated reports how safe people feel. Usefulness is in there too, underneath, and nothing on the dashboard separates the two.

The sentence nobody wrote down

Last week I wrote about a health system that measured real hours back from an AI scribe with no decision anywhere about where those hours were going. That is the downstream version of this problem, and by the time you get there the money is already spent. The upstream version costs almost nothing to fix. Before any hours exist, somebody senior has to say in writing what they are for. Redeployment into work you have been deferring. Fewer hires next year. Shorter days. Faster turnaround for a customer who is threatening to leave. Any of those is an answer a person can plan around. An unanswered question also gets answered, though, and the Pew numbers tell you which answer most people supply when nothing has been published.

In the delivery and coaching work I have done, that question comes up more than once and I have watched it go unaddressed long enough that people stop raising it. The vacancy never lasts. It gets filled in the parking lot, by whoever talks first.

What the silence costs you in the reporting cycle

The practical damage is to your instruments. If a majority of your workforce in a high-income market expects this technology to reduce jobs, a low usage number now carries at least two readings and your slide is showing one of them. It could mean the tool is poor. It could also mean people are careful about looking eager to automate their own work, which is a rational thing to be careful about if nobody has told you otherwise. A high number has the mirror problem. It might mean the tool works, or it might mean a manager made it visible who was using it, and the dashboard cannot tell you which.

A related trap shows up in how you test. Researchers took an AI assistant away from lawyers who had used it for three months to see what each person had kept, which is a different question from what the tool produces while switched on. If you want an answer about the organization, you have to change what you are measuring.

You can repair the instrument before the next cycle for the price of one memo. State what the recovered capacity is for, put a name on it, distribute it, and then read the adoption number knowing what question it is answering. If you cannot write that sentence because leadership has not decided, you do not have a pilot underway. You have an experiment with an uncontrolled variable sitting in the head of every person supplying the data.

So here is the question for your next operations review. If somebody pulled up the AI rollout dashboard and asked what the usage number is measuring, could anyone in the room point to a written answer about where the saved time goes? What has your organization actually published, and how long ago?