The Analyst Who Measures What Automation Actually Removed

AI is the top stated reason in a year when cuts fell

Challenger, Gray & Christmas is an outplacement firm that tallies announced US job cuts every month. Its July 2026 report carries two sentences worth reading next to each other. Employers announced 33,429 cuts in July, down 27% from 45,849 in June, and 477,033 through the first seven months of the year, down 41% from the 806,383 announced over the same stretch in 2025. In the same report, artificial intelligence was the leading stated reason for July cuts at 10,970, or 33% of the month. Year to date, AI appears in 112,713 announcements, roughly 24% of the total.

Put those together and the picture stops being simple. AI became the number one stated reason in a year when the total volume of announced cuts fell by more than 40%. The other categories are not small either: market and economic conditions accounted for 90,075 cuts year to date, closings 84,630, restructuring 57,476, and loss of contract 40,758.

The nature of that tally matters. Challenger counts the reason a company wrote into its own announcement. That is a chosen explanation, not an identified cause. This is exactly what the Financial Times was probing when it asked whether AI is really responsible for recent job cuts. Write “restructuring around AI adoption” and investors read a company positioning for the future. Write “softening demand” and they read a company whose business got worse. The same layoff supports both sentences, and the two sentences are not worth the same.

So companies started hiring someone to check the sentence. The work is separating which tasks actually shrank, whether the shrinkage lines up with a tool rollout or predates it, and what would have happened to headcount without the tool. That is causal inference on organizational data, which is why the posting sits with data science rather than with HR reporting.

You learn to count tasks, not people

The unit of analysis here is a task, not an employee. “Do we need one hundred engineers” cannot be answered. “If we sort what this team ships in a month by type, how many of each, and how many hours does each type consume” starts an actual calculation. Once the work is decomposed, a tool rollout becomes visible as a change in the mix: this category of ticket collapsed, that one did not move. If review latency dropped while requirements definition still takes the same three weeks, the human bottleneck moved rather than disappeared, and headcount planning should follow the bottleneck.

A 2023 study from the Korea Development Institute puts numbers on the same distinction, and the gap it found travels well beyond Korea. It estimated that in 38.8% of jobs, 70% or more of the tasks were technically automatable. Actual adoption in the same report sat at 2.7% of private firms with ten or more employees, rising to roughly 20% among firms with 250 or more. Technical possibility and installed reality were far apart. Everything interesting about this job lives in that gap.

Note also that those figures are three years old now, and studies at that scale do not get rerun annually. Producing a current version for one specific company is the reason the seat exists.

The tooling looks familiar. SQL to join HRIS records against systems that log actual work, Python or R for before-and-after comparison, a dashboard a chief people officer opens monthly. What differs is the data itself. One company’s headcount is a small sample, so significance is hard to reach, and reorganizations chop the time series into pieces. That pushes the craft toward research design instead of model accuracy: which teams form a usable comparison group, which reorg can be treated as a natural experiment, what confound the sponsor of the analysis has not thought of.

Two adjacent roles are easy to confuse with this one. Transition advisory at a consulting firm recommends what an organization should do. This role measures what an organization actually did. Employment counsel defends the decision after it is made. This role supplies the evidence the decision rests on. All three end up in the same room often enough.

The data opens late and the conclusions cost people their jobs

Start with the drawbacks. People data is the last dataset a company opens. Privacy obligations sit on it, works councils and employment law sit near it, and a new analyst does not get query access on day one. Be clear-eyed too about what the output does: an analysis you ran can become the basis for eliminating a role. Requests to run the numbers again because leadership dislikes them arrive more often here than in product analytics.

Two entrances dominate, and they produce different analysts. One starts in data science or analytics and drifts toward organizational questions. Statistics and study design are already there, so what gets learned is compensation structure, job architecture and employment regulation. The other starts inside HR, in recruiting or workforce planning, and adds analytical skill. That person misreads the data less because they know why each field was recorded the way it was. Which background wins depends on where the company parked the team, under the CHRO or under a central data function.

Being honest about the market: dedicated postings are still uncommon, and at many companies a workforce planning manager does this in a spreadsheet alongside other duties. Google is the standard reference for a mature people analytics function, and the number of organizations operating at that level is small. The flip side is that almost no data scientist has touched HR data, so anyone who has spent even one project on it becomes the obvious pick every time a company reorganizes.

If you are still in school, there is a version of this you can practice without corporate data. Pull public labor statistics for an occupation connected to your major, plot monthly job postings or employment counts, then find the month the line bends. Go read what happened that month. Then write down what additional data you would need before claiming that event caused the bend. That last line is the entire discipline. It also exposes the current limit of the field, which is that no public dataset lets an outsider verify the reason printed in a layoff announcement. That verification is only possible from inside, and that is why the seat is inside.