Here's what I keep hearing in every leadership conversation about AI right now:
Will it replace my people?
Wrong question. It is the wrong question because it turns a complex problem into a binary one. It frames headcount as the thing to protect. It invites a false choice, keep the humans or swap in the AI, and it locks you into a threat model that does not match how this actually works.
The right question is harder and more specific: Will AI make my people's expertise worth more, or worth less?
These are not the same question. The answer changes everything about how you should be running your organization right now.
Last fall I sat in on a planning session with the COO of a mid-size insurance carrier. They were about to roll out a claims-triage model across their first-notice-of-loss team, and the whole business case was built on cutting that team roughly in half.
I asked one question. Which part of the job is the model actually taking. It turned out the model was good at the routine claims: clear liability, small dollar amounts, clean documentation. The adjusters spent maybe a third of their day on those. The other two-thirds was the messy work, the contested claims, the conflicting statements, the judgment call on when a file smelled like fraud.
So the automation was removing the easy third and leaving the hard two-thirds. By the framework in this paper, that should make the surviving adjusters more valuable, not less. But the plan treated every head as interchangeable and cut for headcount. Six months later they were paying retention bonuses to keep the senior adjusters they had nearly let go, because nobody else in the building could work a contested file.
MIT's Initiative on the Digital Economy published a synthesis of new research last week. David Autor, the labor economist who has spent a career tracking how technology reshapes what humans do for money, collaborated with MIT's Neil Thompson on one of the papers. They call it Expertise. The core finding, backed by 40 years of labor market data, is this:
Automation doesn't uniformly destroy expertise. Sometimes it destroys it. Sometimes it concentrates it. The outcome depends entirely on which tasks get automated.
When automation removes the less expert tasks from a job, the remaining work becomes harder and more specialized. Wages tend to rise. Employment tends to fall. You have fewer people in the role, but the ones who remain are better and more expensive to replace.
When automation removes the most expert tasks from a job, the remaining work gets simpler. Wages fall. Employment grows. You end up with more people doing what's left, but they are cheaper and easier to churn through.
The same wave of automation can do both things simultaneously in different occupations. Autor and Thompson put it plainly: automation "simultaneously replaces experts and augments expertise."
The accounting clerk versus inventory clerk comparison runs through 40 years of data, but it is worth slowing down on.
Both jobs saw significant automation in the 1980s and 1990s. Computers took over the routine work: data entry, basic reconciliation, simple calculations for accounting; weighing, counting, tracking for inventory.
For accounting clerks, what got automated was the low-expertise work. What remained required real judgment: interpretation, problem-solving, decision-making under uncertainty. The role became more expert. Wages rose. There are fewer accounting clerks today than there were in 1980, but the ones who exist are paid well and genuinely difficult to replace.
For inventory clerks, the opposite happened. The tasks that got automated were actually the most expertise-heavy ones: the pattern recognition, the supply chain intuition, knowing where things actually were versus where the system said they were. What remained was physical and procedural. The role became less expert. Wages fell. Employment grew, then eventually shrank through different means.
The GPS story is even cleaner. Navigation used to be a genuine skill. London cabbies still pass "The Knowledge," a multi-year examination requiring mastery of 25,000 streets. GPS didn't replace taxi drivers. It automated away the expertise requirement for driving, which opened the door to Uber and Lyft. There are far more drivers today than before. They earn less. The expertise was automated away and the occupation was deskilled at scale.
Same underlying technology. Opposite outcomes for the humans involved.
So before any automation decision, one question needs an honest answer: Are we automating the less expert tasks, or the most expert ones?
The answer tells you whether you're building a stronger team or hollowing one out.
The second paper in the MIT IDE batch makes a separate point that deserves time too. The researchers examined over 3,000 text-based tasks across the U.S. labor market and found that AI is not arriving like a crashing wave, sudden, discontinuous, devastating. It's rising like a tide. Current AI performance across job categories runs between 47% and 73% success on text-based tasks. If current trends continue, AI could complete most text-based tasks at an 80–95% success rate by 2029.
This is not a reason to relax. But it does mean you have a runway. Not unlimited time, three years is not a long planning cycle, but enough time to look at every significant role in your organization and ask: if AI keeps improving, which of my people's skills become scarcer and more valuable, and which become commodities?
Most organizations are not asking this question. They are asking about productivity and cost. They are optimizing for efficiency, which is the right answer to a slightly wrong question.
I watched a head of product at a B2B SaaS company do this well, almost by accident. She had a roadmap that said automate the analysts, full stop. I asked her to split the analyst job into the part that was pulling and formatting data and the part that was deciding what the data meant.
Once she saw it on the whiteboard, the plan changed in about ten minutes. She automated the pulling and formatting, kept every analyst, and retitled the role around interpretation and recommendations. A year later she had the same number of people producing roughly three times the output, and a hiring bar that was genuinely hard to clear. The work she protected was the work that was getting scarcer.
The aggregate numbers can lull you into thinking the stakes are lower than they are. The researchers found that the total share of U.S. labor compensation attractive for computer vision automation, for instance, is under 4%. That sounds manageable. Four percent of the economy is not a crisis.
Except it is not a question about the economy. It is a question about your organization.
In the occupations and roles where AI automates the nonexpert tasks, the remaining experts are worth more. They are harder to hire, harder to retain, harder to replace when they leave. You want more of these people, and you want to know who they are before the market reprices them.
In the roles where AI automates the expert tasks, wages soften and turnover accelerates. That can look like a cost savings for several quarters. It tends to look like capability erosion by year three, when nobody on your team actually understands the work at a level that matters.
Autor and Thompson's framework should be in every AI adoption conversation your leadership team is having. Not as a way to manage headcount. As a way to manage expertise: where it is, whether you're protecting it or selling it off, and what you will be left with when the automation cycle completes.
The question is not whether AI will take your job.
The question is whether AI is making your expertise scarcer, or obsolete.
If you do one thing this week, pull the org chart for the team you are most eager to automate and write next to each role the single task you would hand the model first. Then ask whether that task is the easy part of the job or the expert part. If it is the expert part, stop the rollout there and rethink it, because you are about to deskill the role and pay for it in eighteen months. That is a one-afternoon exercise and it is the difference between sharpening a team and quietly hollowing it out.