The August jobs report came out last Friday. It showed that 162,000 jobs were added, but unemployment remained steady at 4.1%. That was better than analysts expected… on paper. But against a week of enterprise AI numbers pointing out the opposite, it feels a little uneasy. McKinsey, Gartner, and Bain all published data showing the same thing: AI spend keeps climbing, but the returns are still lagging behind. That gap has been a steady thread over the last 9 months of my reporting, and I don’t forsee that changing. There are exceptions, of course. And while they’re worth digging into to understand why, it’s important to keep in mind that they stand out precisely because they’re rare.
This week’s coverage:
The ROI Gap Enterprises Can’t Close
McKinsey, Gartner, Bain, and BambooHR all published numbers this week, and none of them agree with the AI hype cycle.
The Frontier Labs Are Cutting Prices, Not Risk
Anthropic and OpenAI both cut prices this week, right as new data showed how few customers actually pay their bills.
What Working AI Actually Looks Like
One health system is doing the thing most enterprises say they can’t — scaling AI without breaking the people using it.
The Layoffs Keep Coming, Hiring Keeps Rising
Layoffs and hiring both rose in the same data set this week, and Microsoft decided investors deserved to see exactly where the AI money is going.
Here’s what I was reading.
The ROI Gap Enterprises Can’t Close
McKinsey’s latest numbers on enterprise AI are being read as good news, and by the standard of the last two years, they are. Thirty-seven percent of organizations now attribute at least some EBIT impact to AI. But only 6% qualify as high performers, and that’s the number that matters more. The gap between the 6% capturing real earnings impact and the 31% claiming “some” is where the actual story is, and it isn’t closing. A Register piece added that most adopters still can’t point to measurable earnings impact at scale, despite two years of investment. (McKinsey says enterprise AI is finally ‘on the road to ROI’)
Gartner’s numbers, published the same week, explain part of why. Fewer than 25% of enterprises have scaled AI successfully past a single business unit. A CIO Dive piece added the harder number: just 22% have scaled AI across multiple units, and the rest are stuck somewhere between pilot and rollout. Most organizations aren’t failing at AI — they’re failing at the unglamorous work of turning a pilot into an operating process, and that’s a management problem wearing a technology costume. (Fewer than 25% of enterprises have scaled AI successfully)
That failure to scale doesn’t stop the spending. It just means the spending keeps rising without the payoff catching up. A Bain analysis found that AI could push a $10 billion consumer-packaged-goods company’s annual IT spend up 75% by 2035. A second CIO Dive piece added the mechanism: even organizations investing carefully are watching costs climb faster than the efficiency gains meant to offset them. Careful investment is still investment, and the bill is coming due before the ROI is. (Even with careful investment, AI is set to boost IT costs)
So where’s that time actually going? BambooHR has an answer. Workers now spend close to 20 days a year fixing AI’s output — errors, hallucinations, rework. Nearly half the time organizations spend “using” AI is actually time spent cleaning up after it, which means the productivity math most companies are running is fiction. (Almost half the time spent on AI is on fixing its output, BambooHR says)
Four data sets, one story: enterprises are spending more, scaling less, and burning a chunk of whatever gains they get on cleanup. That’s not a technology failure. It’s an execution failure, and it’s the same one from two years ago — just with bigger numbers attached.
The Frontier Labs Are Cutting Prices, Not Risk
Anthropic’s new Fable release undercuts its own pricing while promising less restriction and zero data retention — a combination clearly aimed at enterprise buyers who’ve been sitting on the sidelines over compliance concerns. A TechCrunch piece added that zero data retention lets clients run Anthropic’s models entirely on their own infrastructure, with no data leaving the building. Anthropic is selling control as much as it’s selling capability, and for a lot of IT leaders, that’s the more persuasive pitch. (Anthropic’s new Fable release is cheaper, less restrictive)
OpenAI moved the same direction days later, rolling out GPT-6 Astra to top-tier ChatGPT plans at half the rate of GPT-5.6 Sol. A Decoder piece added that the $200 Pro plan now includes 200 GPT-6 Pro messages a week — a meaningfully better deal than the prior tier offered. Both labs cutting prices in the same week isn’t a coincidence. It’s competitive pressure, and someone is going to lose that pricing war. (OpenAI rolls out GPT-6 Astra to top-tier ChatGPT plans at half the rate of GPT-5.6 Sol)
A PYMNTS piece added that OpenAI and Anthropic both get roughly 80% of their revenue from the top 1% of customers — the same week both labs pushed price cuts into the mass market. The mass-market pricing push and the whale-dependent revenue base pull in opposite directions, and only one of those strategies wins. (OpenAI and Anthropic Get 80% of Revenue From 1% of Customers)
If the enterprise whales are what actually pays the bills, cheaper consumer tiers look less like strategy and more like scale-chasing while the real business sits somewhere else, mostly unaffected by any of it.
What Working AI Actually Looks Like
A health system built AI into its clinical workflow from day one, and the results look different from most of what showed up in Section 1. The system serves 240 physicians and roughly half a million patients, using a tool called ScopeAI to deliver clinical background before a patient consultation even starts. An HR Executive piece added that the goal was never to replace physicians — it was to cut the cognitive load of chart review so doctors could spend consultation time on the patient instead of the screen. This is what AI adoption looks like when it’s built around a specific job to be done instead of a mandate to “use AI” — narrow, embedded, and judged by whether physicians actually want to keep using it. (Adoption lessons from the nation’s first AI-native health system)
The piece doesn’t offer hard outcome metrics, and that’s a real gap — a case study without a number is a story, not proof. But the contrast with Section 1 matters: fewer than 25% of enterprises have scaled AI successfully, and here’s one that did, in one of the more change-resistant, high-stakes environments in the economy. Scale doesn’t require every function in the company to touch AI at once. It requires one workflow, done well, that the people using it don’t want to give up.
The Layoffs Keep Coming, Hiring Keeps Rising
September’s WARN notices are already stacking up, with another wave of major companies confirming workforce reductions this month. An IBTimes UK piece added a detail that cuts against the doom narrative: hiring is up 25% compared to last year, even as the layoff headlines keep coming. Layoffs and hiring are climbing at the same time, which means the labor market isn’t shrinking — it’s reshuffling, and reshuffling is harder to plan around than either growth or decline on their own. (September Layoffs 2026: Full List of Major Companies Facing Workforce Reductions)
Microsoft made its own kind of disclosure this week, announcing it will start breaking out Azure’s quarterly revenue as part of a broader segment reporting change. Azure grew 42% to $29.42 billion last quarter, now about a third of Microsoft’s total revenue. A CNBC piece added that the change is meant to give investors clearer visibility into how much of Microsoft’s growth is actually coming from AI infrastructure. Microsoft is choosing transparency because the Azure number is now good enough to be a selling point on its own — companies don’t volunteer to show their homework unless they like the grade. (Microsoft to start disclosing Azure quarterly revenue as company consolidates business units)
Put the two together and the picture is a labor market absorbing AI-driven change unevenly — jobs disappearing in some places, appearing in others, while the infrastructure spend behind all of it grows too large for Microsoft to keep folded into a single line item.
The AI ROI story is not fictional — McKinsey’s 6% of high performers prove that much. What’s fictional is the idea that spending your way into that group is enough. The organizations still stuck below 25% scaled aren’t behind on technology. They’re behind on the operating discipline that turns a pilot into a habit, and no amount of cheaper tokens from Anthropic or OpenAI closes that gap for them. Are most leaders running an AI strategy right now, or a series of expensive experiments they’re calling one? In my opinion, most are running experiments — and the teams and functions that stop doing that first are the ones McKinsey will be writing about next year.