Layoffs Hit a Two-Year Low While AI Stayed the Top Stated Reason

33,429 and 477,033

Challenger, Gray & Christmas tracks job cuts that US employers announce, month by month. The July tally, published August 6, came to 33,429. That is 27% below June’s 45,849 and 46% below the 62,075 announced in July 2025. It is the lowest monthly figure in two years.

The cumulative picture points the same way. Announced cuts from January through July total 477,033, down 41% from 806,383 over the same seven months last year. If you read layoff headlines every week, that number runs against the feeling.

In the Same Report, AI Led for a Fifth Straight Month

July cuts attributed to AI came to 10,970, or 33% of the total. Market and economic conditions came second at 7,960, followed by closings at 6,060, restructuring at 2,815, and loss of contract at 2,003. That is five consecutive months with AI at the top. Year to date, 112,713 cuts carry the AI label, 24% of all cuts, and the count since 2023 stands at 184,538.

Stop there and you have the familiar story. Put both figures in one table and it changes.

The peak month for AI-attributed cuts was May, at 38,579 against a monthly total of 97,006, which lands close to 40%. June came in at 14,029 and July at 10,970. In absolute terms the count fell to roughly a quarter of its peak in two months.

The share barely moved: 31% in June, 33% in July. Do the division yourself and the reason shows. July is 10,970 over 33,429. May was 38,579 over 97,006. When the numerator shrinks but the denominator shrinks faster, the quotient holds or climbs. “AI is the leading reason for job cuts” is currently counting a label on a pile that is getting smaller.

It also helps to remember what this dataset is. Challenger does not investigate causes; it collects the reasons companies write in their announcements. In CBS News coverage of the April figures, Andy Challenger said that regardless of whether individual jobs are being replaced by AI, the money for those roles is. Budget moved. Where it moved to is recorded in a different table, not in the reason column.

The Other Half of the Same Report

Announced hiring plans in July came to 16,095. That is 47% above June and 400% above the 3,200 announced in July 2025. Year to date hiring plans total 107,500, up 25% on last year. Challenger’s own framing is that AI is shifting the labor market rather than dismantling it.

Counting live postings gives a rougher but more specific view of what is being hired. In a scan startupxo ran on August 20 across 17 public job APIs, covering 110 postings, the boards with the most open roles were Stripe at 12, Anthropic at 11, Databricks at 9, Robinhood at 9, and Cursor at 8. Three of the five most common titles sit against customers or operations: solutions architect at 5, security engineer at 4, and forward deployed engineer at 4. That last title originated at Palantir in the early 2010s and describes an engineer who goes into a customer’s environment and leaves working code behind.

The sample has obvious limits. It is a single point in time across 17 specific boards, not a census of the labor market. Within that sample, though, the roles being added outnumber the ones being described as automated away.

Two Things to Check in a Layoff Announcement

When you read a competitor’s or a portfolio company’s cut announcement, look at the count before the share. A sentence like “AI accounted for 40%” only means something next to the month’s total. Forty percent of a month that halved and forty percent of a month that grew are different events entirely.

Second, hold the stated reason against the org chart. A company that cites AI will also have posted roles that quarter, and those postings are public. If the roles cut and the roles opened cover the same function, that is redeployment. If they cover different functions, the business changed direction. One word in a press release does not distinguish the two.

One limit worth stating plainly: Challenger counts announced plans, not people who actually left. How far announcements diverge from execution is not visible in this dataset.