A rare return to the survey pile: 41 stats pulled from Gallup, KPMG, Ramp, OpenAI, and more that tell the adoption story the show doesn't usually tell — half of US workers now use AI on the job, ROI still isn't landing on the bottom line, and the gap between the vanguard and everyone else is getting wider, not smaller.
Adoption is past the halfway mark — and the gap is still widening.
The show stopped covering adoption surveys because they felt disconnected from what frontier AI can actually do. But check back in and the numbers tell a coherent story: everyone uses AI now (52% of US workers, per Gallup), almost nobody has translated it to established ROI (7%, per KPMG), token costs are rattling the C-suite, and the distance between the top 1% of AI spenders and the median business is phenomenal. We are still very, very early — and the people doing the hard work of adopting from within normal, non-AI lives are the ones who will translate this for everyone else.
The show quietly dropped its old habit of dissecting big professional-services reports because the stats felt disconnected from actual AI capability. Since the start of the year — new models, harnesses, real agents — the world has split into those adapting to a totally new way of working and those laboring under the old model, and the gap is getting wider, not smaller.
A mid-May Gallup poll found 52% of US workers use AI at work — the first time past the halfway mark, across all worker types. The question has officially shifted from how much adoption there is to how well AI is being used.
Domino Data Lab found 93% of enterprises report improved production capability — yet 57% say AI's ROI still fails to outpace spend, and PwC's mid-year CEO snapshot saw just 39% reporting measurable positive outcomes. KPMG's Global AI Pulse captures the same hinterland: only 7% of global leaders report established ROI, even as those saying AI delivers meaningful business value jumped 12 points in a quarter to 76%.
In an EY AI Pulse survey, 98% of C-suite leaders said token costs were forcing them to reconsider their AI plans — a byproduct of the agentic shift from seats-times-$20 to the total cost of intelligence in used tokens. Yet only 64% of them actually meter their usage.
Ramp's card data across 70,000+ businesses shows the top 1% spending about $7,500 per employee per month on AI — an enormous multiple of the median AI-buying business. And since Ramp customers skew tech-forward, the real economy-wide gap is probably even bigger.
In Ramp's July spend data, 42.4% of its business users paid for Anthropic subscriptions versus 39.5% for OpenAI — confirming what everyone has been feeling: this is a two-company race. The thing to watch next: whether the cost-conscious token-efficiency era pushes companies toward open-weights models, fine-tuning, and multi-model router strategies.
A KPMG quarterly pulse survey found resistance to AI agents jumped from 5% to 20% in a single quarter. A jump that big could be noise — but more capable, more threatening agents could plausibly generate more ire among employees not yet taking advantage of them, so it's one to watch.
Per stats OpenAI shared at the end of June, over 25% of Codex users have handed the agent a task estimated at more than eight hours of human work. The scope of work being delegated to AI is increasing significantly.
In a controlled Atlassian experiment, workers who disclosed their AI use were rated ten times lazier than identical peers who stayed quiet. That's a huge drag on adoption: AI spreads best when power users adapt it to the firm's context and evangelize outward — and they won't if they're side-eyed for it.
PagerDuty found 66% of office professionals have used AI tools they believed broke company policy. The show's contention: this isn't malicious — the tools people have outside work are often dramatically better than what's inside. As Codex and Claude Code adoption rises inside the enterprise, the incentive for shadow behavior should shrink.
OpenAI's Work at the Frontier report, examining more than 800,000 work messages, found 43.5% of occupation-specific ChatGPT use at work is for tasks belonging to a different occupation than the user's own — the marketer updating the site instead of waiting on engineering. That blurriness may have the biggest impact on how work changes over the next few years.
BCG's June AI at Work study found 47% of workers spending more time managing and supervising AI than doing "actual work." The bot-sitting problem is acute in this transition — but ironically, for many roles the future job increasingly won't be to do the work, it will be to manage the agents that do it. That's exactly the world we're headed into.
DX found that across more than 500 engineering organizations, over 50% of code was AI-generated in Q2, up from 34% a quarter earlier. With OpenAI and Anthropic basically at 100%, the notable thing is that the shift in software engineering practice is no longer confined to early adopters — it's getting everywhere.
ZipRecruiter found 38% of employers have already shifted basic data-entry work from entry-level workers to AI and 31% have raised experience requirements — yet 35% of the same sample expect AI to grow total headcount. The organizational read on labor is very messy right now.
48% of hiring managers say their company would rather invest in AI than hire and train a new graduate — but Ramp and Revelio Labs' data on 21,000+ firms shows heavy AI adopters grew entry-level hiring 12% in the two years after adoption. Serious AI spend correlates with more junior employees, not fewer; the companies using AI to eliminate juniors may eventually re-learn what these firms already know.
While overall US job postings fell 7%, Indeed Hiring Lab found 15% growth in software developer postings since February 2025. Caveat for completeness: 71% of that growth was in senior roles, so the junior question stays open — but developers are not being wiped off the face of the planet.
AI has been the number-one stated reason for US job cuts for five straight months per Challenger, Gray & Christmas — yet the Yale Budget Lab finds exactly zero clear AI fingerprints in aggregate US occupation data, nearly three years after ChatGPT. The likely explanation: AI is a politically palatable thing to blame layoffs on, and that excuse should carry less water in the months to come.
Inside Higher Ed found 55% of college students expect AI to hurt their career prospects, with only 7% all-in on AI — but in KPMG's July Summer Intern Pulse, just 5% feared job displacement, and the top worry (43%) was losing critical-thinking skills. It's the gap between imagining AI and actually using it in real work; two-thirds of those interns had AI assisting with over a quarter of their assignments.
An Anthropic study from late last year found just 15% of Americans trust AI companies to decide how AI is developed — a number unlikely to have improved. Meanwhile, a June Pew study found Americans saying by three to one that China, not the US, is more advanced in AI.
A Reuters/Ipsos poll found 77% of Americans — near-equal shares of Republicans and Democrats — worry AI will make electricity more expensive, and 57% would oppose a data center in their community. Addressing the power-price concern should be table stakes for data-center builders, and hopefully that memo lands as the backlash gets louder.
An official ECB survey found about 50% of firms plan to invest in training current staff for AI versus only 12% planning to hire AI specialists. The catch: the market has utterly failed to provide good enablement solutions, so like their American counterparts, they'll likely end up building highly bespoke ones.
Adobe found that on this year's Prime Day, AI-referred shoppers converted 40% better than non-AI-referred ones. E-commerce feels like an area headed for complete upheaval from new consumer patterns sooner rather than later.
Axiom's survey of 528 in-house legal leaders found 92% either expect or are already negotiating AI-related cuts from outside counsel. That phenomenon won't stay unique to legal — how internal teams and external firms work together is going to shift dramatically across professional services.
Elon University found 27% of US adult internet users now have social or emotional interactions with AI, with 31% of them calling it a friend — while 74% of those same users predict AI will deepen society's loneliness. If you're looking for optimism, that self-awareness could at least create alternative paths forward.
Common Sense Media found 86% of kids ages 9-17 already use AI, and over 40% say no parent has ever discussed AI safety with them. Gen Z got hammered by social media partly because parents lacked the basis to help them navigate it — we cannot afford that mistake again, and pretending the genie goes back in the bottle is toothpaste-back-in-the-tube energy.
Broad public opinion won't be shaped by people who spend all day making AI content — it will be shaped by those adopting and adapting to AI from within normal, non-AI lives, who translate it for everyone else. In other words: my conversation is with you all, but your conversations are with everyone else. No pressure at all.