Mustafa Suleyman, the CEO of Microsoft AI, told the Financial Times that human-level performance on most professional tasks is twelve to eighteen months away, and he named the targets directly: accounting, legal, marketing, project management, anything that happens sitting at a computer. He pointed at the curve of compute as the reason it is close to inevitable, and he is one of several leaders saying a version of this. Dario Amodei put a number on entry-level white-collar jobs last year, Ford's CEO said much the same about half of US white-collar roles, and the drumbeat has been loud enough that most leaders have already picked a reaction to it, either bracing for it or rolling their eyes.
The reporting that follows that forecast, in the same article, complicates it.
What the same article quietly admits
A controlled study this year found AI made experienced software developers roughly 19% slower at their own tasks. Thomson Reuters looked at lawyers, accountants and auditors actually using the tools and found targeted, marginal gains, well short of the displacement the headlines describe. Apollo's chief economist pointed out that profit margins climbed for big tech through late 2025 while the broader market barely moved, because investors do not expect AI to lift earnings outside the technology sector itself. The capability is real and improving fast, and the economic result for most companies outside that sector is still close to nothing.
Sixty-one percent is the figure a leadership team should sit with, because it held stable while the models underneath it improved dramatically. Better models kept arriving while the result held flat, which is the clearest signal available that the constraint sits somewhere other than the technology.
Why the deadline keeps slipping
Suleyman gave away the actual mechanism in one line of the same interview. He said companies will retrofit the technology to perform any required job function, and retrofit is the entire problem compressed into a single word. It means taking a process that was designed around people doing the work by hand and fitting AI into the gaps that process happens to leave. What you get is a faster version of a shape that was built for a different kind of worker, which is exactly why the measured gain stays small and sometimes turns negative. The clock keeps resetting twelve to eighteen months out because every model release makes that retrofit a little better, and a better retrofit still tops out far short of the promise. The deadline keeps being set by people measuring capability when the binding constraint is the operating model that capability has to run inside.
What the 39% did differently
The thirty-nine percent of organizations that do show measurable impact are running the same widely available tools as everyone else. What they changed is the work those tools plug into. They looked hard at how a function actually runs, decided what should exist at all once AI is in the room, and rebuilt around that answer before the technology went in. The AI then compounds on a shape that was designed for it, and that single decision is most of the distance between the 39% and the 61%. This is slower to start and it does not make for a good keynote slide, which is part of why the loud forecasts skip it, and it is also the only place the measured returns have reliably shown up.
The question worth sitting with
At some point the capability stops being arguable. The models keep improving, and the eighteen-month clock, wrong as a deadline, is right about the direction. When that moment lands, the technology will be good enough for everyone at roughly the same time, so it stops being the thing that separates one company from the next. What separates them is whether the way the business runs can absorb it, or whether it quietly converts the most capable technology in a generation into one more entry in the sixty-one percent. No model release decides that for you, and the answer gets more expensive the longer it goes unasked. Your margins register it well before your strategy deck does.