The AI & Work Brief: better tools, harder decisions

The week’s research signals

Anthropic’s September 24 book-trading study tests agent delegation in a controlled employee experiment. It offers a narrow setting for examining a broad workplace question: how should a person evaluate a decision made on their behalf?

SHRM’s September 22 release examines IT and computer-science job postings across 27 countries, using July 2025–June 2026 data. It reports rising AI-skill demand across the countries studied. That is a hiring-advertisement signal, not a count of people hired.

IBM’s September 21 release reports that 80% of surveyed CHROs believe AI adds work such as checking outputs and handling exceptions. The study surveyed 1,500 CHROs and 8,800 employees; this percentage describes executive responses, not measured hours.

Fed Communities’ September 22 article brings together existing Federal Reserve research on AI and work. It emphasizes that occupational exposure can lead to different outcomes; exposure does not establish that a job will disappear.

Sources: Anthropic; SHRM; IBM Institute for Business Value; Fed Communities

What the evidence does—and does not—add up to

Our synthesis is that these sources are most useful as separate prompts for investigation. A successful experimental task is evidence about capability under specified conditions. A posting describes what an employer advertises. A survey records what respondents report. Combining them does not produce a single displacement rate, and this briefing makes no such estimate.

The Conference Board’s four labor-market scenarios and Anthropic’s economic explorer add longer-term planning context. They are conditional frameworks, not new counts of people losing work this week.

Sources: The Conference Board; Anthropic

Our interpretation: evaluate the whole assignment

We would make the complete assignment the unit of evaluation. For example, a hypothetical analyst producing a customer report should count time gathering inputs, resolving unclear definitions, checking calculations, and explaining the result. Measuring only the drafting stage would leave the rest of the assignment invisible to that experiment.

This changes the portfolio we would recommend to someone looking for work. Include a short before-and-after account of one bounded task: the starting material, the desired outcome, the checks performed, and the remaining mistakes. Explain which choices you made and why. Keep confidential data out of the example. A reviewer should be able to understand your contribution without needing access to your employer’s systems.

For managers, our proposed test is similarly concrete. Ask whether the experiment improves a result the team actually owns: a resolved case, an accepted deliverable, or a decision made with sufficient evidence. If the output is faster but requires another team to repair it, record that transfer of work before calling it a gain.

A practical agenda for the coming week

Pick one recurring assignment and describe a satisfactory result before trying a tool. Compare a few similar examples, recording elapsed time, your active effort, review effort, and corrections. Keep the result descriptive if the sample is small. Write down what you would change for a second attempt and what evidence would persuade you to stop.

Then examine a handful of current openings for your target role and location. Note the specific responsibilities behind broad AI language. Choose one skill to demonstrate through the assignment you just evaluated. This connects learning to work you can explain, while leaving room to change direction as better evidence arrives.