On August 27, 2026, the Bureau of Labor Statistics published its first AI exposure classification: four categories โ Low, Moderate, High, Very high โ assigned to all 831 detailed occupations in the 2025โ35 employment projections. On the same page, BLS states that the classification "is not a forecast of employment growth or decline," "is not an estimate of the probability of AI adoption," and "is not a worker replacement estimate" (U.S. Bureau of Labor Statistics, 2026).
Within days, the trade press had converted it into at-risk and safe job lists.
That is the one reading the publisher explicitly rules out โ in writing, on the page the lists were built from. And it matters more than a pedantic correction, because the BLS AI exposure table is about to arrive in mid-market 2027 workforce plans as the most authoritative-looking input in the deck. A federal statistical agency, every occupation, four clean tiers. It will be treated as a cut list. It cannot carry that decision.
What the BLS AI Exposure Classification Actually Says
The classification is a relative ranking, produced by a clustering algorithm over five external exposure measures. BLS is specific about the limits: the categories are "relative to other occupations," they do not "imply an absolute 'high' or 'low' level of exposure," and they "do not distinguish between AI impacts from automation versus augmentation" (U.S. Bureau of Labor Statistics, 2026).
Sit with that last one. Automation and augmentation are the same category in this table. "Very high exposure" is equally consistent with the work disappearing and with the work being done by the same person at three times the throughput with a review step attached. The table does not tell you which โ and the direction is the entire content of a headcount decision.
The surrounding projections are the context the exposure ranking is usually stripped out of. BLS projects the U.S. economy adding 5.9 million jobs between 2025 and 2035, employment rising from 170.3 million to 176.2 million โ growth of 3.5 percent, which the agency notes is "slower than the 10.9 percent growth recorded over the 2015โ25 decade" (U.S. Bureau of Labor Statistics, 2026). Slower growth, not contraction. The AI-related adjustments concentrate in computer and mathematical occupations, arts and media, business administration, and legal services.
The Construction Detail That Sets the Ceiling
The five inputs are three theoretical exposure measures and two observed-usage datasets โ the observed side drawn from Anthropic's Claude usage data and Microsoft's Copilot data. BLS takes the median percentile rank within each dimension, then clusters. Across 4,155 possible occupation-source combinations, 3,944 were observed and 211 imputed, affecting 75 occupations (U.S. Bureau of Labor Statistics, 2026).
Then the line that should govern how far you let this table travel: "The theoretical sources conceptualize AI capabilities as those available no later than mid-2023."
Three of five inputs describe a model generation that predates every agentic product your team has piloted. BLS also notes that all five sources treat occupations as bundles of independent tasks and "do not consider bottlenecks, complementarities, or other non-independent features of occupations."
That is not a flaw in the BLS work โ it is a disclosed property of the available literature, and disclosing it is what a statistical agency is for. It is a flaw in the use. A ten-year staffing plan built on this ranking embeds a three-year-old capability snapshot, and does so inside the one document your CFO will not think to question.
What the Ranking Does Map: Verification Load
Here is the defensible read, and it inverts the popular one.
If an occupation clusters as highly exposed, what the five underlying sources are collectively detecting is that a large share of its constituent tasks are the kind of work current models will attempt. Whether the model is right is a separate question the exposure score does not touch โ because, as BLS notes, the inputs treat tasks as independent and ignore complementarities.
So high exposure predicts high draft volume. Draft volume creates review demand. And review demand is where the cost actually sits.
McKinsey QuantumBlack priced this directly for banking workflows: token costs run 20 to 25 percent of an AI agent's variable run cost, while human oversight by functional and risk experts accounts for 70 to 75 percent (McKinsey QuantumBlack, 2026). The expensive line is the reviewing, and the reviewer has to be senior enough to overrule the output.
The MIT Technology Review Insights and Microsoft study of 300 technology executives and practitioners, which ranked 101 agentic tasks on a 0โ100 trust scale, points the same way: confidence tracked task verifiability and business-context completeness rather than model capability, and 59 percent already plan permanent human oversight (MIT Technology Review Insights, 2026).
Read the BLS table through that lens and it becomes genuinely useful. Your Very high and High occupations are where AI-generated output will concentrate, which means they are where verification hours will concentrate, which means they are where you need people with enough domain standing to reject bad output. Cutting there first removes the reviewers from the functions with the most to review.
Why a Job Title Can't Carry the Decision Anyway
There is a second problem, independent of the mid-2023 ceiling, and it is structural: exposure is measured per occupation, while adoption happens per person.
Bick, Blandin, Deming and Schumacher, linking a nationally representative survey to detailed O*NET tasks, find that genAI exposure measures explain only about half of the variation in adoption across workers โ individuals doing similar work adopt systematically differently (NBER Working Paper 35677, 2026). They also find adoption broad but shallow: at least one in five workers uses genAI in 80 percent of occupations and 40 percent of job tasks, while in most cases adoption stays below 50 percent.
Half the variance sits inside the job title, not between job titles. Which means a plan that assigns capacity changes by SOC code is, at best, half-informed โ and the half it misses is the half you could actually manage, because person-level adoption responds to enablement, tooling and role design.
Use the BLS AI exposure categories the way a national statistic should be used: as a prior about where to go look. Then do the looking inside your own company, against your own task inventory. Two occupations in the same category, in your building, will not behave the same way, and the residual is not noise โ it is the part of the problem you have leverage over.
The Fair Counter: This Is Still the Best Public Baseline
I want to be even-handed about what BLS did, because "don't use it as a cut list" is not "don't use it."
It is the first classification to cover every projected occupation on a consistent method, it blends theoretical and observed-usage evidence rather than picking one camp, and it documents its own limits more candidly than most vendor research does. For benchmarking your workforce mix against the national distribution, for sequencing which functions get an enablement budget first, for arguing a case to a board that wants a citable source โ it is the best public instrument available, and nothing in the mid-market will produce a better one internally.
The failure is not the table. It is the compression: four tiers, one column, a category label that reads like a verdict. That formatting invites a headcount conclusion the underlying method cannot support, and BLS pre-emptively said so on the same page. The discipline required is small โ carry the caveats with the number every time it appears in a deck.
What to Decide This Quarter
Four changes to how the BLS AI exposure ranking enters your 2027 planning cycle.
- Re-label the column. Wherever the exposure category appears in a planning document, the header is "expected review load," not "risk." Same data, correct verb. The label does most of the work in how a table gets read.
- Pair every exposure score with an internal adoption measure. Exposure explains about half the variance; your own usage data explains the rest. If you cannot measure adoption per worker, that instrumentation is the prerequisite to the plan, not an output of it.
- Protect reviewer depth in your highest-exposure functions. Those functions will generate the most machine output and need the most senior judgment to check it. Model the reviewer hours before you model the savings.
- Date-stamp the capability assumption. Write "theoretical inputs reflect AI capabilities no later than mid-2023" directly beneath the table. It ages the claim in front of whoever reads it next โ which is the only defence against a 2026 snapshot silently governing a 2035 projection.
The BLS gave you a map of where the work will change. It said, five separate times and in its own words, that it did not give you a map of where the workers go.
If a 2027 headcount number in your plan traces back to a Very high exposure label, you have not made a workforce decision. You have quoted a ranking that ruled out the use you just made of it โ and the agency that published it will not be the one defending that number when it is wrong.
