Covid Was the Rehearsal. AI Is the Show.

During lockdown, companies sorted their workforce into two piles: jobs that could be done from a laptop in a spare bedroom, and jobs that couldn't. Accountants went home. Plumbers didn't. Data analysts kept working from the couch. Physical therapists came back to the clinic as soon as they were allowed.

That sorting revealed something about the nature of work that nobody was looking for at the time. The jobs that moved home easily shared a quality: the work could be packaged up and handed off without losing much in translation. Clear inputs, structured outputs, minimal need for physical presence or real-time human contact. I call this quality Delegation Resistance: how hard it is to delegate a task to someone (or something) that isn't in the room.

Three years later, AI is following the same fault line. I built a formal measure of Delegation Resistance from US Department of Labor data on 894 occupations and validated it against 4.4 million Indeed job postings. The sectors where remote work took hold most deeply are the same sectors where job postings have fallen furthest since 2022. If you know how remote-friendly a sector was, you can predict about half of the variation in its AI-era hiring decline.

Most analyses stop there. But when you look at the task level, the exposure is much narrower than the correlation implies. Only 29 out of 894 occupations have no task at all that anchors them to a human. That's 1.6% of the US workforce. Everyone else has at least one thing in their job that AI can't absorb (yet). Search for yours:

The rest of this post explains how that score works, what the evidence looks like, and where it breaks down. If you just came to check your job, you're done. If you want to understand why, keep reading.

What Is Delegation Resistance?

Think about what made some jobs easy to do remotely. An accountant's inputs are numbers in spreadsheets and tax codes in documents. The work can be fully described, transmitted, and verified without anyone being in the same room. A physical therapist's inputs are a patient's body, their pain response, the way they flinch when you push a joint past its comfortable range. That work can't be packaged up and sent over Zoom.

I call this quality Delegation Resistance (DR): how difficult is it to package up a task and hand it to someone (or something) that isn't in the room? Low DR means the work travels easily. High DR means it resists delegation. The pandemic revealed DR before anyone named it.

AI exploits the same seam. The binding constraint is not how smart ChatGPT is but how easily the work can be communicated to it. A tax return has clear inputs and structured outputs. A therapy session requires reading body language and responding to half-formed thoughts as they surface. Both require intelligence, but only one fits inside a prompt.

DR is built from O*NET survey data across 894 detailed occupations. It captures face-to-face interaction requirements, physical presence needs, real-time coordination, and how much the work involves hands-on manipulation of the physical world. A data analyst scores low because the work is mostly text and numbers that move fine over a wire. A physical therapist scores high because you need to be in the room, hands on the patient, adjusting as you go.

The same pattern showed up in offshoring. In the early 2000s, certain kinds of work could be described well enough to survive the trip across a fiber optic cable to Bangalore. Managing a sales team couldn't, even though both jobs happen at a desk. Remote work, offshoring, AI: each time, the jobs that move are the ones where the work can be cleanly separated from the worker.

Labor economists will notice that DR overlaps with the routine/non-routine distinction in the Autor-Levy-Murnane framework. The overlap is real but not complete. DR is specifically about delegability: can the work be packaged and handed off? A routine task that requires physical presence (running a machine on a factory floor) scores high on DR even though it's highly routine. DR and routine-task share are correlated, but they're measuring different seams.

The Rehearsal: Remote Work

To test whether DR predicts anything, I used Indeed's Hiring Lab data: 4.4 million daily job posting observations across 44 sectors from 2020 through early 2026. I built a crosswalk mapping Indeed's sector taxonomy to O*NET occupations through keyword matching, giving each sector a DR score based on the average of its constituent occupations.

First, the rehearsal. DR predicts which sectors adopted remote work most aggressively during and after the pandemic. The correlation was r = −0.70, meaning sectors with low delegation resistance went remote and sectors with high friction didn't. Legal, Mathematics, and Banking & Finance all have 25–32% remote postings. Hospitality, Veterinary, and Pharmacy have 3–4%. That's the rehearsal.

The Show: AI-Era Job Postings

Now the show. I compared each sector's job posting level before ChatGPT's launch (January through October 2022) to its most recent level. The same measure that predicts remote work adoption also predicts which sectors are losing hiring demand:

Sectors where work can be easily packaged up and handed off saw the steepest declines in job postings since 2022. Data & Analytics fell 68%. Marketing dropped 53%. Scientific R&D declined 59%. Meanwhile, high-DR sectors held up. Therapy grew 11%. Physicians & Surgeons grew 4%. Sports grew 5%.

One caveat: posting declines are not the same as job losses. A company might post fewer roles because it's doing more with existing staff, because it's hiring through other channels, or because a hiring freeze has nothing to do with AI. The correlation between DR and posting change is real, but what it measures is hiring intent, not employment. The real test comes when BLS releases occupation-level employment counts for 2025 and 2026.

Ruling Out COVID Normalization

Maybe the sectors that declined in job postings were simply normalizing from a post-COVID hiring boom. Maybe DR correlates with whatever other forces are at play, and AI has nothing to do with it.

The timeseries helps here. If this were just post-COVID normalization, you'd expect low-DR and high-DR sectors to converge back toward their pre-COVID baselines at roughly similar rates. They don't. Through most of 2021 and 2022, the two groups track each other closely. The divergence begins in early 2023 and widens steadily through 2026.

I initially hoped ChatGPT's launch date (November 30, 2022) would provide a clean natural experiment (DR-posting correlation near zero before, strong after). It didn't hold up. The 2022 tech sector cooling, driven by interest rate hikes and the end of cheap money, hit low-DR desk-work sectors hardest. That pattern looks a lot like the AI-era pattern and was already underway when ChatGPT launched. I can't cleanly separate the two forces with this data.^†

The divergence between low-DR and high-DR sectors has persisted and widened for over three years now, well past the point where a one-off interest rate adjustment would explain it. The correlation between DR and posting decline (r = 0.65) holds across the full AI era, not a narrow post-launch window.

Both groups track each other through the COVID recovery and most of 2022. Starting in early 2023, the low-DR sectors pull away. By early 2026, the gap is roughly 25 index points, and still widening.

Tasks Are Not Occupations

DR predicts which sectors are losing job postings. You could rank all 894 occupations by DR score and declare the bottom third "at risk." A lot of AI-and-jobs analyses do exactly this.

The problem is that O*NET measures occupations, not tasks. It gives you a single DR score for "Veterinarian" by averaging across everything a vet does. A vet's day includes reading radiology images (low DR, AI can already do this), reviewing bloodwork (low DR), and performing surgery on a squirming Labrador (very high DR). The surgery anchors the job. As long as someone needs to be in the room with the animal, the occupation survives even if AI absorbs a large share of the diagnostic work.

How many occupations are actually low-DR all the way through, with no anchoring task at all? I went back to the O*NET item-level data and checked. For each of the 894 occupations, I looked at all 11 individual items that feed into Delegation Resistance and asked: is there even one that scores above average?

Only 29 occupations out of 894 have no anchor whatsoever. That's 3.2%. Those 29 occupations employ about 2.3 million workers, or 1.6% of the US workforce. The remaining 88% of occupations have at least one task dimension that scores solidly high on delegation resistance, even if their average DR is low.

The 29 anchorless occupations are a specific type: search marketing strategists (861,000 workers), billing and posting clerks (418,000), cost estimators (220,000), data entry keyers (135,000), financial quantitative analysts (127,000). Further down the list: economists, physicists, mathematicians, business intelligence analysts, web developers. Every one of these jobs lives entirely on a screen. No task in the bundle requires physical presence or real-time human contact.

For the other 96.8% of occupations, including many that get called "AI-vulnerable" in the press, there's at least one task that keeps a human anchored to the role. AI will reshape what those workers spend their time on, and will probably reduce headcount in some cases. But you can't fully automate a job when part of it requires being in the room with the patient, or reading the client's face across a table.

Not all anchors are equal. To stress-test this piece I asked an AI to simulate critiques from four leading labor economists and got plausible, detailed responses in about 60 seconds.^‡ That says something about the framework: the physical anchors (surgery, plumbing, standing in a classroom) are far more durable than the cognitive ones (reading a room, exercising judgment, synthesizing ambiguous information). As AI capabilities expand, the cognitive anchors erode first.

The anchor score at the top of this post makes this visible. It runs from 0 (fully exposed) to 1 (well anchored), summing the strength of each occupation's above-average delegation resistance items on a normalized scale. Your AI exposure is simply 1 minus your anchor score.

What About Offshoring?

If Delegation Resistance predicts which work can be delegated, it should also predict which jobs were vulnerable to offshoring in the 2000s. I tested this using 21 years of Bureau of Labor Statistics employment data (2004–2024).

It didn't work. Low-DR occupations actually grew more during the offshoring era (2004–2010), not less. The relationship was statistically significant in the wrong direction (t=2.12, p=0.036).

Offshoring was a much smaller economic force than AI, and it happened alongside a massive expansion of the financial and professional services sectors that employ lots of low-DR workers. The same jobs that were theoretically offshorable (finance, consulting, analysis) were also the ones riding the pre-2008 boom. DR may be measuring the right thing but getting swamped by a bigger signal.

Or offshoring and AI-driven disruption work differently in ways DR doesn't capture. Offshoring required setting up operations in another country and managing across time zones. AI requires a subscription. The friction of adopting AI is so much lower that delegation resistance becomes the binding constraint in a way it wasn't for offshoring.

I don't want to explain away a result that doesn't fit the story. The backward validation failed. The comparison also isn't quite apples to apples: the Indeed validation uses job postings, while the offshoring test uses BLS employment counts. Postings and employment measure different things, so the failure of the backward test doesn't directly contradict the Indeed results. But it does mean the historical evidence I was hoping for isn't there.

Data Sources

Occupation data: O*NET 28.1 Work Context survey (894 occupations, 11 items). Job postings: Indeed Hiring Lab (44 sectors, 2020–2026). Employment and wages: BLS OEWS May 2024.