Global AI spend continues on pace to nearly double this year. That’s a lot of money that’s, potentially, being wasted — at least that’s what a lot of the data shows. Last week, the majority of noise came from the AI companies demanding a slow-down. This week, what I read pointed the opposite direction. Employees, under pressure to use more AI, are inflating how much they do, at pay structures that reward the wrong skills, and at workflows built before any of this existed.
Last week’s coverage:
Workers Are Performing AI Use, Not Doing It
Employees felt pressure to use AI, weren’t sure what was expected, and plenty admitted to inflating how much they were doing it.
Employers Are Paying for AI Skills While Judgment Erodes
Pay premiums are landing on tool fluency, and IBM’s data showed the judgment that makes the output trustworthy slipping.
The Bottleneck Is the Workflow, Not the Model
AWS, Fortune and Microsoft’s own write-up all put the constraint in how work is organized, not in what the models can do.
The Frontier Debate Keeps Skipping What’s Already Deployed
Gates wanted mandatory oversight, a lawsuit called the labs’ slowdown talk collusion, and Mollick said the models already deployed were the real story.
Here’s what I was reading.
Workers Are Performing AI Use, Not Doing It
Organizational demands for AI aren’t translating to the real-world work. In a survey, 44% of workers said they feel the pressure to use AI, 42% weren’t confident about folding AI into their processes, and only 22% called themselves very confident. Sixty percent said they either aren’t expected to use AI or don’t have clear expectations. Workers are being pushed to use AI without anyone defining what using it well looks like. (44% of workers feel ‘AI-whelmed’ as pressure to use AI grows)
Nearly half of employees in a Visier study reported pressure to use AI, and a similar share admitted they exaggerate their use and expertise to colleagues and managers. Seven in 10 worried adoption could hurt their careers or cost them their jobs. When employees can’t safely say what isn’t working, they report that everything is. Visier’s Andrea Derler put the cost bluntly: leaders who don’t allow that transparency get blanket answers, which means “you’re wasting time.” (Employees report performative AI use amid role changes)
Usage is the easiest AI number to report... and (perhaps more importantly) the easiest to inflate. Derler said the pressure to show results runs through every level: CIOs need to prove ROI, managers push their teams, employees guess at what counts. So what does a rising adoption curve prove? Sometimes only that people learned what to say.
Employers Are Paying for AI Skills While Judgment Erodes
An HR Dive piece on a Payscale study found that 58% of companies are paying premiums for AI skills or plan to. Sixty-one percent are rewriting job descriptions because of AI, but fewer than half say their salary structures have kept up. Payscale projected what comes next: new hires commanding premiums of 20% to 40% while current employees who upskilled get nothing, a “retention time bomb.” Employers are pricing AI skill into new hires and leaving employees who upskilled without a raise. (Should workers be paid for AI skills?)
Three in four employees told IBM that critical thinking and other skills, human judgment among them, were starting to erode because of AI. Eighty percent of CHROs said AI creates invisible work: validating recommendations, fixing errors. The skill that makes AI output trustworthy is the one employees said they were losing, and the validation labor was landing on them anyway.
Where workflows built judgment in, 62% of CHROs reported rising confidence in AI-enabled decisions. Where they didn’t, 57% reported it falling. That’s a design choice, and the data showed it paying off, but only slightly more. (IBM report warns AI could erode human skills deemed vital by CHROs)
Jack Clark made the human-side case in Fortune back in April: the degrees gaining value are the ones built on synthesis across subjects and knowing which questions to ask, and Anthropic’s own researchers had estimated AI could theoretically handle 94% of computer and math tasks. (Anthropic billionaire cofounder Jack Clark studied literature, not code—and says liberal arts degrees will win in the AI age | Fortune) Pay is following AI tool fluency, while the judgment Clark and IBM both pointed to isn’t showing up in Payscale’s numbers.
The Bottleneck Is the Workflow, Not the Model
More than two-thirds of the 154 executives AWS interviewed across 27 countries named legacy workflows, not technology limits, as the biggest barrier to AI value, a CIO Dive piece reported. AWS’s Tom Godden called it a business problem before an IT one and said treating it as an IT problem yields nominal gains at best. The executives who got value rebuilt end-to-end processes and reinvested the time saved in areas they’d planned in advance.
Godden’s explanation for where the gains go was simple: speeding up coding does nothing for governance, funding, approvals or coordination across teams. Making one step faster doesn’t make the whole chain faster, and that’s where the savings disappear. BMW’s CIO told AWS the biggest surprise was how much alignment, process reengineering and change management mattered compared with the technology. Sysco is targeting $500 million in AI savings over three years and Chewy projected $50 million a year (targets, not results). Gartner research cited in the piece put the share of enterprises that had scaled AI across multiple operations below 25%. (AI gains call for organizational overhauls: AWS)
Only 7% of enterprises describe their data as completely ready for AI, a Fortune commentary piece reported, and fewer than a quarter have a data strategy at all. Sixty-three percent either lack AI-suitable data management or aren’t sure they have it. McKinsey’s survey, the piece noted, found just 6% of companies reporting significant impact. Data readiness, not model quality, is where enterprise AI stalls. (The backwards AI pacing debate and how far business is from the front…)
Microsoft’s own AI write-up, published the week before, had already made the workflow argument from the inside. Access and usage didn’t change how the work got done until the sales team rebuilt around a business outcome: a tool licensed to more than 200,000 people, the post said, plateaued until the team mapped how account managers spent their week. The supply-chain team simplified its processes first, then built a single source of truth so agents reasoned from the same data, and cut cycle time by up to 75% in selected workflows. (What we’ve learned from Microsoft’s own AI transformation) The footnote on the 20% close-rate gain compares 687 sellers who used Copilot heavily with those who used it lightly, January through June 2024.
The Frontier Debate Keeps Skipping What’s Already Deployed
Bill Gates told NBC News that “No one thinks self-regulation is enough,” and called for required safeguards and monitoring that he described as overhead for the industry, not a dramatic slowdown. Mark Zuckerberg told the same network the opposite: no industrywide coordination needed. Microsoft’s Brad Smith asked a UN session whether an airplane maker worried about crashes would say it needs to win the race. California, Maryland and New York each moved on AI rules. The pacing argument was still about who sets the pace: Congress, the labs together, or each lab alone. (Bill Gates says AI companies self-regulating isn’t enough and governments should be involved in monitoring)
A class action filed the Friday before argued that even the slowdown talk breaks the law. The plaintiffs, paid subscribers to ChatGPT, Claude, Grok and Gemini, alleged that the labs’ September 12 responses to Dario Amodei’s pacing essay amounted to an illegal agreement that reduces what consumers get. Amodei’s essay had itself flagged antitrust exposure and asked the government for a narrow waiver, and Sen. Josh Hawley said he wouldn’t agree to any exemption. The plaintiffs argued that agreeing to slow down together is itself the violation, even if each lab slowing on its own wouldn’t be. (Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown)
An Ethan Mollick post argued the whole debate underrates what’s already here. GPT-6 Astra and Fable 5.1, he wrote, can reliably do weeks of human work when properly guided, and the capability overhang between what those models can do and what most people do with them is massive. If every lab stopped training tomorrow, the models already deployed would still change how large parts of the economy work. His four human advantages (deep knowledge, wide knowledge, taste and agency) sound a lot like the judgment IBM and Clark were pointing at. (The Overhang)
A Fortune piece argued pacing would cost the economy little, because enterprises need time to absorb what’s already on the table. Mollick argued a slowdown wouldn’t change much. Both put the unabsorbed capability in organizations, not labs. So who owns absorption? Gates wants Congress to own safety, Zuckerberg wants each lab to own it, and nobody in that argument owned the harder job: getting organizations to absorb what’s already shipped.
Inflated usage numbers, pay structures rewarding tool fluency, workflows nobody redesigned, data that only 7% of enterprises call ready... they all trace to one choice: leaders bought the capability and left the organization around it as it was. The pacing fight will run for months, and it’s a fight over models. Are executives running an AI strategy when nobody owns the job of absorbing what’s already deployed? In my opinion, the answer is clearly no.
That’s it for this week’s BeAIReady brief!
If you appreciate the depth of reporting and how I connect the dots, please like, share this post, and subscribe (or share the Brief with a friend!). Thanks!
~erick