What SHRM's 2026 survey actually measured — and why the "nontechnical barrier" finding is the interesting one

SHRM's 2026 Automation and AI Survey reports that about 20% of US employment, roughly 31 million jobs, is now at least half automated, up from 15% the year before.¹ The headline number is doing the rounds. The more useful finding, buried one layer down, is a correlation that wasn't there a year ago: the more automated an occupation is, the more likely its workers are to name a specific reason their job still can't be fully done by a machine. I want to walk through what that actually means, because it changes how you should read the 20% figure.

What the survey is, and how it's different

Most of the AI-and-jobs numbers you have seen over the last decade come from a particular kind of analysis. A researcher takes an occupation, breaks it into tasks using a structural dataset like O*NET, scores each task for how automatable it looks, and adds the scores back up. Frey and Osborne's famous 2013 paper worked this way. So did most of the McKinsey and OECD projections that followed. Call this the task-imputation method: the researcher looks at the job from the outside and infers what a machine could do.

SHRM's survey does something different. It asks the workers. Each respondent rates how much of their own role is currently automated, and identifies what, specifically, is stopping the rest from being automated. The unit of analysis is the occupation, but the data source is the person doing the work. Call this the worker-report method.

These two methods measure different things, and they can disagree sharply. Task-imputation tells you what is technically automatable in principle. Worker-report tells you what is actually being automated in practice, and what the people closest to the work think is in the way of more. Neither is the truth. They are two different lenses, and the SHRM survey is one of the few large datasets we have through the second lens.

The 20% number, in context

A five-percentage-point jump in one year sounds dramatic, and it might be. But two cautions before you anchor on it.

First, "at least 50% automated" is a self-rated threshold. A worker saying half their tasks are automated includes everything from spreadsheets and email filters to LLM-drafted reports. The category is broad. It is not the same as "half this job has been replaced by AI in the last year".

Second, one-year intervals are short. A survey wave can shift for reasons that have nothing to do with the underlying labour market — different sample composition, a slightly reworded question, a different moment in the news cycle that primes how workers think about their own roles. SHRM itself is careful about this. The 2025 baseline plus one follow-up is the start of a trend line, not a trend line yet.

So: the 20% is real, it is a meaningful jump, and it is also one data point on a young series. Hold it loosely.

The finding that wasn't there last year

Here is where it gets interesting. SHRM also asks workers to identify nontechnical barriers to fuller automation of their role. These are the obstacles that are not about whether the model is smart enough — they are about everything else. SHRM groups them into roughly five families:

• Trust and judgment — decisions where someone needs to be accountable, or where the cost of a wrong call is high enough that a human signs off.
• Regulatory constraints — rules that require a licensed or certified human in the loop (medical diagnosis, legal advice, financial sign-off).
• Relational dependencies — work where the relationship with another person is part of the product (sales, therapy, teaching, management).
• Physical presence — work that requires a body in a place.
• Ethical accountability — work where someone has to be answerable, in a way an algorithm cannot be.

In the 2025 wave, there was no clear relationship between how automated an occupation was and how likely its workers were to name one of these barriers. In the 2026 wave, there is — a positive correlation. The more automated the occupation, the more its workers can name what is still in the way.

The natural reading is: workers who have used AI tools in their actual job have learned where those tools hit a ceiling. They've watched the model draft the report and then hand back the part that requires a judgment call. They've seen the chatbot handle the easy tickets and escalate the ones with a regulatory wrinkle. Exposure is teaching them where the moats are.

That is a different story from "AI is coming for your job". It is closer to: AI is arriving, it is doing real work, and the people closest to it are getting a clearer view of what it can't do — not in principle, but in the specific contexts they work in.

What this means for how to read displacement numbers

Two practical things follow.

The 20% headline and the barriers finding belong together. Reading either alone gives you a distorted picture. Twenty percent of jobs being half-automated, and the most-automated workers being the ones who can most clearly articulate what is left — that is a more accurate read than either fact in isolation.

The displacement risk is concentrated, not general. If you accept the worker-report lens, the occupations that should worry you most are not the most-automated ones overall. They are the subset with high automation share and no nontechnical barriers identified. That intersection is small, and it is targetable. It is a much more useful object for policy or workforce planning than a single 20% number.

There is a real caveat, which the survey doesn't resolve. Workers have an incentive to name reasons their job is hard to replace. OECD work on worker self-assessment of job security has documented this optimism bias for years. The barriers may be real, and partly inflated. We won't know which until the survey wave gets a wage-and-hours-worked anchor from datasets like Brookings and Hamilton Project, which so far haven't confirmed a matching displacement signal at the payroll level.²

The thing most worth remembering: the survey's most useful contribution isn't the 20% figure. It is the year-on-year evidence that workers who use AI are getting better at naming what AI doesn't do. That is a kind of distributed empirical knowledge about the technology's limits — generated, for free, by everyone using it. Whether you trust it or not, it is data we didn't have before.

Footnotes and links

Further reading

• SHRM 2026 Automation and AI Survey, full report: https://www.shrm.org/topics-tools/research/automation-generative-ai-and-job-displacement-risk-in-u-s--employment/2026-full-report
• Acemoglu and Restrepo, "Robots and Jobs: Evidence from US Labor Markets," Journal of Political Economy, 2020 — on the divergence between task-level and worker-report methodologies. https://www.journals.uchicago.edu/doi/10.1086/705716
• OECD Employment Outlook 2023, "Artificial Intelligence and the Labour Market" — on survey-wave comparability and worker self-assessment. https://www.oecd.org/employment/oecd-employment-outlook-2023-08785bba-en.htm

¶