The Data Says Wait

The question business owners are quietly asking right now: do I need to start cutting roles because of AI?

That question is understandable. The coverage of AI and jobs has been loud, and some of it is genuinely alarming. But decisions made on fear rather than evidence tend to create problems that are harder to fix than the ones they were trying to avoid.

Recently, Anthropic released a research paper specifically designed to measure AI's actual impact on employment, not theorize about it. The findings are worth reading carefully, because they push back against the narrative most people have absorbed.

The research introduced a new measurement called observed exposure. Prior studies measured which jobs AI could theoretically affect. This one measured which tasks people are actually using AI for in professional settings, and at what level of automation.

Three findings stand out.

First, actual AI usage is far below theoretical capability. Claude is being used on about 33% of tasks in the Computer and Math category, even though theoretically 94% of those tasks are feasible with an LLM. There is a large gap between what AI can do and what it is doing.

Second, the workers most exposed to AI are not who most people picture. They are more educated, higher-paid, more likely to be female, and more likely to be older. The study found that graduate-degree holders are nearly four times more represented in the high-exposure group than in the unexposed group.

Third, and most directly relevant to workforce decisions: there is no measurable increase in unemployment among highly exposed workers since late 2022. The researchers looked for it. They used multiple methods. They found nothing statistically significant.

The one signal they did find was narrow: hiring of workers aged 22 to 25 into exposed occupations has slowed slightly. A roughly 14% drop in the job-finding rate for that group, and the researchers themselves flag that it is barely statistically significant and has multiple alternative explanations.

Here is the order problem that shows up repeatedly in how businesses are responding to AI.

A business owner reads coverage about AI replacing jobs. They feel pressure to act. They start considering restructuring, changing hiring plans, or cutting roles they assume will soon be automated. They do this before checking whether the underlying conditions actually exist in their business.

That is acting on a prediction rather than a diagnosis. And the prediction, as of today, is not supported by the employment data.

The researchers are explicit about this. They built the framework now, before major effects have appeared, precisely so future changes can be identified clearly. They are not saying AI will have no labor market impact. They are saying the impact is not measurable yet, and that past attempts to predict disruption have a poor track record.

One historical example from the paper: a widely cited study identified roughly 25% of US jobs as vulnerable to offshoring. A decade later, most of those jobs had healthy employment growth.

The cost of restructuring a team prematurely is real. Losing experienced people, disrupting institutional knowledge, signaling instability to your remaining staff. These are not hypothetical costs. They happen now, while the displacement being feared may or may not happen later.

Wait. Not indefinitely, and not passively.

Wait on workforce restructuring decisions that are driven primarily by AI displacement fear. The data does not support them yet. The researchers said they would revisit this analysis as new data emerges. Follow their lead.

In the meantime, the productive move is internal: understand which tasks in your business are actually being handled differently because of AI, and whether that is creating capacity or just adding noise. That is a diagnostic question, not a restructuring question.

If AI is genuinely changing how work gets done in your operation, you will see it in output and process, not in a research paper. Start there.

Decide well,
Chuck