The Number That Stopped Moving
Ramp's AI Index hit 41% in May after nearly ten straight months of growth. That single data point — drawn from card and bill-pay transactions across roughly 30,000 U.S. companies — says more about the state of enterprise AI than any keynote deck I've seen this year. Adoption isn't collapsing. It's just... done growing. And the reasons why go deeper than budget cycles.
When you look at the breakdown, the pattern gets starker. Large businesses? 49% have deployed AI in some form. Medium-sized firms? 44%. Small companies? 37%. The gradient tracks resources, not enthusiasm. Every size tier is close to its ceiling.
The AI Impact on Human Psychology at Work
Here's what the transaction data can't tell you: what this plateau feels like from the inside. The board asked for AI adoption. Teams delivered adoption. Then the board asked for proof — proof that the spend translated to output, that the "productivity gains" leadership had been promised actually materialized. And most teams couldn't produce it.
The Multitudes study covered by LeadDev surveyed engineering teams and found what many managers already suspect in their gut: nearly all teams are using AI tools, yet most cannot clearly measure their impact. Not because the tools do nothing, but because the question itself is malformed. Lines of code break down when AI generates half your pull requests. Sprint velocity means something different when an AI assistant drafts your story descriptions. The old dashboard doesn't capture the new work.
This creates a psychological squeeze that nobody talks about explicitly. Engineers and product teams feel caught between two incompatible pressures — prove that AI helps, and prove that you're still necessary. The second pressure rarely gets named. It doesn't need to. People feel it every time a reorg announcement lands. It's the same tension explored from a different angle in Reclaiming the Sovereign Mind: Why Friction and Mental Effort Are Essential to Human Flourishing Under AI.
That feeling, the low-grade anxiety of being measured against a tool that might eventually replace you, is where the AI impact on human psychology becomes a business problem, not just a sociology one. You cannot get honest adoption data from people who are anxious about what the data will prove.
Measurement Isn't Just Broken, It's Misleading
LeadDev's reporting on the Multitudes report lays out the paradox plainly. Engineering leaders want outcome-oriented metrics. They want to know whether teams are delivering on roadmaps faster and with confidence. But most organizations default to activity counts, number of AI prompts, number of AI-assisted commits, adoption rate percentages. These numbers look impressive on a quarterly slide, but they tell you almost nothing about whether the spend is buying anything real — a problem we break down in The AI Bill You Didn't Know You Were Running Up.
Andres Peate, who leads engineering at Multitudes, recommends measuring broadly rather than perfectly. Track usage, sure. But also track customer value delivered (pull requests shipped, tickets completed), incidents and bug reports as quality signals, and, critically, the human impact. Whether AI rollouts are pushing more work into evenings and weekends. Whether the 2 a.m. AI incident is your first indicator that quality has degraded.
The advice is sound. The organizational will to implement it is scarcer than the tool budget. Boards want a single dashboard number by Friday. They're not going to get a nuanced framework that treats human wellbeing as a first-class metric alongside velocity.
Klarna's Lesson and the Pilot Graveyard
You don't have to look far for a case study in what happens when adoption enthusiasm outruns operational reality. Klarna publicly committed to replacing hundreds of support agents with AI. Then customer service quality dropped. Then Klarna hired some of those workers back.
This isn't an edge case. S&P Global data shows that 42% of companies abandoning most of their generative AI pilot projects, up from 17% the year prior. Nearly half of what corporate America rushed into, they're now walking away from. Each abandoned pilot represents not just sunk cost but a small psychological withdrawal. The team that built the pilot learned something about their workflow. The organization that shelved it often discarded that learning along with the tool.
Spending Growth Will Slow Before Infrastructure Pays Off
The macro picture reinforces the micro data. Morgan Stanley analysts, writing in analysis reported by the Economic Times, expect AI spending growth to slow sharply by 2028 as capital shifts from hardware and infrastructure buildout toward software and application utility. The infrastructure phase, the GPU purchases, the data center construction, was the easiest money to spend and the hardest to tie to any human outcome. Software spend at least creates artifacts people interact with. The scale of that disconnect is the subject of The AI Investment Gap: Why Trillions Spent on Infrastructure Haven't Delivered Returns.
That shift, when it arrives fully, will make the measurement problem more urgent rather than less. Software tools have users. Users have feelings about tools. The therapeutic relationship between a person and their daily software, whether it makes work feel lighter or heavier, becomes the relevant signal once the capex story fades.
What Comes After the Honeymoon
A plateau at 41% isn't a crisis. It's information. It tells us that the first wave of corporate AI adoption, the "buy a seat, send an email, call it innovation" wave, has exhausted its easy surface area. The remaining adoption will require real workflow redesign, honest measurement, and a willingness to admit that some pilots were theater.
The companies that navigate this next stretch well will be the ones that stop treating AI adoption as a procurement exercise and start treating it as a change-management challenge. Which means paying attention to what the tools do to people's sense of competence, their relationship to their own expertise, their trust in the measurement systems that judge them.
The transaction data will tell you when money stopped moving. It won't tell you when confidence stopped moving. That's a harder thing to track, but it's the thing that determines whether your remaining 59% of non-adopters ever convert, or just quietly wait for the next hype cycle to pass them by.
