Light AI spenders show no hiring gain. The heaviest grew headcount 10%

Revelio Labs and the finance platform Ramp matched AI spending records at more than 21,000 US companies against their workforce data. Revelio published the result on 30 June. Stanford's Institute for Economic Policy Research (SIEPR) cited it again in July, and The Daily Upside did the same on 28 September ahead of the September jobs report. The headline is that AI adopters hire more. The useful finding sits underneath it. Companies in the top third of AI spending per employee kept employment about 10.2% higher than comparable firms that had not yet adopted. The other two-thirds of adopters showed no statistically significant gain.

The average hides two different companies

Revelio counts a company as an adopter once it records at least three consecutive months of $100 or more in monthly payments to AI vendors such as OpenAI and Anthropic. By that measure, roughly a quarter of the sample had adopted by the end of 2025. The researchers then split adopters by spending per employee in the first three months after adoption.

That split carries the whole result. The heavy spenders grew. The light spenders stayed flat.

Entry-level hiring follows the same line. Among the heaviest spenders, the entry-level share of the workforce rose by 1.15 percentage points against firms that had not yet adopted. Light adopters moved the other way, with a modest fall in their entry-level share. The authors add that earlier studies unable to separate the two groups "may be picking up the signal from low-intensity adopters".

Revelio is careful about cause. Adopters were different before they bought anything: median year-on-year headcount growth of 6.0%, against 1.6% at companies that never adopt. The authors say plainly that the data does not show AI mechanically creating jobs. It shows that the firms going deepest are the firms growing. The gap also opens late. Employment paths look alike around the adoption date and only separate several months on, as AI is worked into how the company operates.

Most employers sit in the flat group

SIEPR's July brief supplies the base rate. The Census Bureau's Business Trends and Outlook Survey puts AI use at about 20% of firms. Even among adopters, SIEPR finds use is usually confined to one or two business functions, or spread thinly across tasks at low frequency. Full integration is the exception. Only 5% of firms in the Census data report any employment effect from AI. Eighty per cent of executives surveyed by the Federal Reserve Bank of Atlanta said their AI investments had yet to change headcount or improve productivity.

Put that beside the Revelio split. The typical employer is in the group the data shows going nowhere on headcount and thinning slightly at the junior end. SIEPR also records where HR executives say the effect actually lands: role consolidation, and hiring avoidance in roles where AI can automate many tasks. That is the hollowing pattern. No announcement, just fewer replacements.

The same brief shows why depth matters more than access. In one study SIEPR cites, an AI assistant given to Kenyan entrepreneurs lowered revenue and profit for the less-skilled among them, because they acted on generic advice that did not fit their business. Stronger performers drew out suggestions tailored to their situation. The tool was identical. The judgement applied to it was the variable.

The comfortable reading of this summer's research is that AI adoption and hiring go together. The accurate reading is narrower. Buying the tools, on this evidence, delivers neither headcount growth nor a stronger junior pipeline. The gains show up where companies rebuild work around the tools and keep people on the decisions. AI does the routine work. You do the thinking. The firms that are growing are organised that way.

Your Next Move

Audit your team's depth this week. List every place AI output is a standard step in a recurring process, with a named owner and a defined check. Then list the licences used ad hoc. If the first list has one entry or none, your team belongs to the flat group, whatever the vendor invoice says.

Move one workflow from trial to standard before the month ends. Choose a recurring task, write the procedure with the AI step and the human review point, and run it for four weeks. Report the hours and the error rate to your manager in writing. The person who documents the shift from licence to process is the person the deeper version of the company needs.

Ask the depth question in every hiring conversation. Whether you are recruiting or being recruited, ask which processes run with AI built in, who owns them, and what has happened to junior intake since adoption. An answer given as a licence count tells you which group the company is in.