Invest1 publisher2 min readPublished
McKinsey survey finds just 6% of companies report significant impact from AI so far
A Fortune essay by the authors of the forthcoming When Machines Act argues that the pacing fight misreads how AI reaches the economy. The survey lines it cites put the clock in enterprise data.
The Investor · Invest desk
What happened
- A Fortune essay argues that both sides of the AI pacing fight share what it calls the Compute-to-GDP Fallacy, treating each leap in model performance as immediate macroeconomic output.
- McKinsey's survey of the business community found 6 percent of companies reporting a significant impact from AI, with modest earnings attribution.
- Among high-performing companies, more than two-thirds still name data as the primary barrier to implementing AI, a share that has held even as model capability advanced.
- McKinsey senior partner Asutosh Padhi, speaking on air with Fareed Zakaria, emphasized that technical availability is not economic transformation.
- The authors say more than 100 conversations with CEOs, policy leaders and AI scientists for their coming book, When Machines Act, convinced them the debate has lost first-principles thinking.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint If enterprises still need years to absorb what has already shipped, throttling the frontier costs the labs option value on future contracts and leaves the revenue they are billing now where it is.
- decision Anyone sizing a lab's revenue has to choose a leading indicator, and on the essay's figures deployed workflows and renewals lead, with benchmark position trailing.
- exposure The essay puts the advocates' own books in play by asking whether pacing is a marketing gambit from frontier labs and cybersecurity firms tidying financials before IPOs.
- contradiction The claim that pacing would cost the economy little sits beside an open question about whether Beijing would pace too, and the second leaves the policy case unresolved.
The diffusion series the essay leans on is a strange piece of evidence. Electricity took 75 years to lift productivity economy-wide, computers 50, the internet and mobile devices 25 [5]. Each step down is exactly 25 years, and extending that line gives the next general-purpose technology zero [1]. Read as ratios it behaves better: two thirds, then a half, which applied to 25 years puts AI's economy-wide productivity lag somewhere around 12 to 17 years [2]. The authors commit only to "an accelerated pace" [21].
Between now and then, the labs are selling into workflows that do not need the frontier. Companies are picking the high-reward, low-risk automation that models one or two generations old already handle [12]. Budget shocks from what the essay calls "tokenmaxxing" pushed daily enterprise work toward simpler models. Few tasks at the average Fortune 500 company require a frontier system at all, the authors argue [14]. Accelerators cast aside in the scramble for cutting-edge silicon are finding a second life on that work [15].
The readiness self-reports are the numbers I would put in front of anyone underwriting a lab's growth. Only 7% of companies call their data "completely ready" for AI, and 93% do not [9][3]. Fewer than a quarter have a data strategy at all [10]. And 63% either lack AI-suitable data management or are unsure whether they have it [11]. The completely-ready share sits one point above McKinsey's significant-impact share [4]. Those are self-assessments from separate surveys, and the essay does not say who collected the readiness figures.
So a frontier pause, on this evidence, costs the labs very little of what they currently bill. Fragmented data silos, legacy ERPs and compliance regimes govern the pace of absorption [20]. An unnamed former Wall Street CEO told the authors the systems will run in parallel with legacy stacks for years, to confirm they work and that no regulatory risk has been absorbed unknowingly [13]. In my view adoption speed sets enterprise AI revenue for the next several years, and benchmark position mostly sets the press cycle. The counter-argument I take seriously is that integration is itself a model task. A system that maps a legacy ERP, reconciles the silos and writes its own migration would convert capability straight into absorption, and the 7% readiness number would stop binding within a year. The cheaper way to find out is in the revenue line. If enterprise growth at the labs accelerates on release dates while deployed seats and workflows stay flat, capability is converting and the readiness surveys are measuring the wrong thing.
What to watch
- The next McKinsey read on the 6% significant-impact share: a multi-point move inside a year would compress the diffusion analogy the essay rests on.
- Whether the completely-ready data share climbs into double digits, which would loosen the absorption constraint the argument depends on.
- Whether When Machines Act puts its 100-plus CEO and policy interviews on the record, so the CEO claims can be checked.