21,559 Companies Adopted AI. The Heaviest Adopters Hired Most.

This series has been about three walls standing between what AI can do and AI actually running a company. Cost. Trust. Jobs.

Jobs was always the wall people were most confident about, and the one with the least evidence underneath it. That changed.

The study

Ramp's Economics Lab and Revelio Labs did the thing nobody had done: linked observed corporate spending to observed workforce records. Not a survey about intentions. Actual AI vendor spend from Ramp's corporate card and bill pay data, matched against Revelio's monthly workforce records, across 21,559 US firms from 2021 to early 2026.

What they found:

• Companies in the high-intensity AI adoption group grew headcount 10.2% in the two years after adoption.
• Entry-level roles grew about 12%.
• The hiring showed up across sales, marketing, administration, finance and customer service. Not just engineering.
• Only the heavy spenders posted gains that were statistically significant.

Entry level is the number that matters, because entry level was supposed to be the casualty. The reasoning was clean and everyone believed it: junior work is the most codifiable, so it goes first. The firms automating hardest hired 12% more of it.

The honest caveats, because this post lives or dies on them. The researchers are explicit that this is correlation, not causation. AI adopters were already larger, faster-growing and more technical than average. The dataset covers Ramp-linked companies meeting activity filters, so it describes a tech-forward slice of the economy rather than the whole thing. What the study rules out is a simple, visible, large-scale substitution of workers by AI inside the firms doing the most AI. It does not prove AI created those jobs.

It is not alone. Anthropic's head of economics has reported that jobs have been less negatively impacted than expected. a16z's data shows the biggest AI adopters increasing entry-level hiring. Sam Altman, who had every incentive to claim otherwise, said in July that so far AI appears to have been net job-creating, and that this was not what he expected at this level of capability.

Three years of confident predictions. The firm-level data arrived and pointed the other way.

So where did the damage go

Something did break. It just was not the layer everyone was watching.

Through July, Gergely Orosz documented a pattern across his network and then confirmed it was global: CTOs, Heads of Engineering and VPs of Engineering are burning out and leaving. Hiring for those roles is hard, and even after the role is filled, people leave again. The reasons, from the people themselves: the role they were hired into is a bad one to hold right now, largely because AI has made expectations unrealistic. Or they arrive and realise the company has no future.

StaySaaSy put the market in one line: 90% of engineering leadership roles hiring right now are one of two bad options. Come fix a company that is dying in the AI era after all the good people left. Or something equally unappealing.

Mark this clearly. This is a well-observed trend from practitioners, not a study with 21,559 firms behind it. It is directional. The headcount data is rigorous, the leadership churn is anecdotal. Treat them accordingly.

But the shape is consistent, and it makes sense.

Why the middle broke and the bottom did not

A junior's job is to do defined work under supervision. AI makes that person more productive at defined work, so a company that is growing wants more of them, not fewer. The economics of a junior improved.

An engineering leader's job was to manage people doing work. That job has quietly been replaced by a different one: designing systems in which agents do the work, deciding where humans stay in the loop, owning what happens when an agent is wrong at 2am, and answering to a board that read one article and now expects a 40% cost reduction by Q3.

Nobody was trained for that job. It is not the job they were hired into. There is no playbook, the expectations arrived before the methods did, and the person holding it is accountable for an outcome nobody has reliably produced yet.

So the org chart did not shrink. It hollowed at the level where the work has to be redesigned, which is precisely the level where AI transformation either happens or does not.

What this means if you run the company

Three things follow.

Stop budgeting for headcount reduction. The firms getting the most out of AI are growing. If your AI business case is built on cutting people, you are targeting the outcome the data does not support, and you will spend the year defending a number that never arrives.

Your bottleneck is the layer you are not protecting. The person who has to redesign how a function works is the person most likely to quit this year. Hiring a Head of AI on top of them makes it worse, because it splits accountability for the redesign away from the person who owns the function.

Junior hiring is a live advantage. Entry-level talent plus agent leverage is the combination the data actually rewards. Everyone else froze their graduate pipeline on the strength of a prediction that did not happen.

The wall was never the jobs. It was the dozen people expected to redesign the work, with no method, no cover, and a deadline.

Protect the middle.