AI Can Do 94% of the Job. Only 33% Is Happening. Here's What HR Leaders Must Track Before the Gap Closes.
AI Can Do 94% of the Job. Only 33% Is Happening. Here's What HR Leaders Must Track Before the Gap Closes.
Anthropic's researchers have produced a number that HR leaders should print out and put on their desks: 61.
That is the gap, in percentage points, between what AI can theoretically do across computer and math occupations — 94% — and what workers in those roles are actually using AI for today — 33%. It is the largest documented discrepancy between AI capability and AI deployment in the workforce, and it represents the window of time organizations have to prepare before disruption arrives all at once.
The finding comes from a March 2026 Anthropic research paper, "Labor market impacts of AI: A new measure and early evidence," which introduces a novel methodology for measuring not just what AI could do but what it is doing. For workforce planners, that distinction matters more than any capability benchmark.
Why "Observed Exposure" Is Different — and Why It Matters
Most prior AI workforce risk studies estimated exposure using task taxonomy: if a job's tasks could theoretically be performed by AI, the job was flagged as "at risk." The Anthropic approach goes further. By combining actual Claude usage data with O*NET occupational data and foundational work by Eloundou et al. (2023), the research team produced an "observed exposure" score: a measure of AI capability actually being deployed in real work contexts.
The result is a two-dimensional picture — theoretical capability on one axis, actual deployment on the other — and the space between them is the gap HR needs to be watching. [Source: Anthropic Research, March 5, 2026]
Office and administrative roles show a similar pattern: roughly 90% theoretical capability exposure, but a fraction of that in observed usage. [Source: Fortune, March 6, 2026] At the same time, 30% of the workforce has near-zero AI exposure — workers in physical-presence roles where current AI tools have little traction. [Source: Yahoo Finance, March 6, 2026]
Who Is Most Exposed
The research identifies which workers face the highest observed AI deployment today:
- Computer Programmers are the single most exposed role, with 75% observed coverage — meaning AI tools are already actively used for the majority of their core tasks. [Source: Anthropic Research, March 5, 2026]
- Other highly exposed roles include lawyers, financial analysts, software developers, data entry clerks, and customer service representatives. [Source: Fortune, March 6, 2026]
- The most exposed demographic skews female, higher-earning (+47% above median), and highly educated — graduate degree holders are four times more likely to be in heavily AI-exposed roles. [Source: Yahoo Finance, March 6, 2026]
The Michigan Journal of Economics notes that white-collar workers are most susceptible to the repetitive digital tasks AI handles best, and that AI adoption may widen the income gap between knowledge workers who augment with AI and those who do not. [Source: Michigan Journal of Economics, March 13, 2026]
Early Labor Signals — Read Carefully
The same Anthropic research team tracked early employment effects among workers aged 22–25 in AI-exposed fields. Since the launch of ChatGPT in late 2022:
- Job-finding rates in AI-exposed fields for this cohort have declined by approximately 14%. The researchers note this figure is "just barely statistically significant" — important context that should temper strong causal claims. [Source: Anthropic Research, March 5, 2026]
- Employment among AI-exposed workers aged 22–25 has fallen by approximately 16%. [Source: Fortune, March 6, 2026]
These signals are concentrated in entry-level roles — a point Anthropic CEO Dario Amodei reinforced in January 2026 when he predicted AI could displace half of entry-level white-collar jobs within one to five years. [Source: The Register, March 7, 2026]
Two caveats must travel with these numbers: first, alternative explanations for the decline in young worker employment exist beyond AI adoption alone; second, and critically, the overall unemployment effect remains indistinguishable from zero at the macro level. The February 2026 jobs report showed 92,000 jobs shed and unemployment at 4.4% — not the catastrophic signal some feared. [Source: Fortune, March 6, 2026]
The HR Action Framework: Measuring Your Exposure Gap
The implication of the Anthropic methodology for HR leaders is direct: your organization has its own exposure gap, and it is currently unmeasured.
Here is how to start mapping it:
Audit by role, not by department. Use O*NET task data or a comparable framework to identify which roles in your organization have high theoretical AI exposure. Then survey or instrument actual AI tool usage in those roles. The delta is your organizational exposure gap.
Prioritize early-career monitoring. The employment effects are most visible in workers under 30 in AI-exposed roles. If you are hiring entry-level talent in software, finance, legal, or administrative functions, your planning assumptions from 2023 are likely outdated.
Track the gap as a leading indicator. The 61-point gap for computer/math workers will not stay at 61 points. As deployment accelerates — driven by tool improvement, organizational adoption, and cost pressure — the gap will close. When it does, disruption will compress into months, not years. Organizations that have mapped their exposure will have runway; those that haven't will be reactive.
Build role-level metrics, not just headcount. Traditional workforce planning tracks headcount and attrition. Add AI exposure score and AI tool utilization rate as tracked metrics for high-exposure job families. Review quarterly.
Scenario-plan for the compression event. The Anthropic "Great Recession for white-collar workers" framing is a model, not a forecast — the researchers are explicit about this. But scenario planning for a rapid-close outcome is exactly what the data supports, and building that scenario into succession planning and skills investment decisions is prudent risk management.
The Bottom Line
Anthropic's observed-exposure framework is the most rigorous tool yet for understanding where AI deployment lags AI capability. The 61-point gap in computer and math occupations is not a sign that disruption won't happen — it is a sign that the timeline is still partially open. HR leaders who begin measuring their own exposure gaps now, before that gap closes, will have options. Those who wait for macro employment data to confirm the shift may find the window has already passed.
The signal is in the gap. Start measuring it.
Sources
- Anthropic Research — "Labor market impacts of AI: A new measure and early evidence" (March 5, 2026): https://www.anthropic.com/research/labor-market-impacts
- Fortune (March 6, 2026): https://fortune.com/2026/03/06/ai-job-losses-report-anthropic-research-great-recession-for-white-collar-workers/
- Yahoo Finance (March 6, 2026): https://finance.yahoo.com/news/anthropic-just-mapped-jobs-ai-163800195.html
- Michigan Journal of Economics (March 13, 2026): https://sites.lsa.umich.edu/mje/2026/03/13/ai-on-the-job-industry-how-blue-collar-and-white-collar-workers-are-impacted/
- The Register (March 7, 2026): https://www.theregister.com/2026/03/07/anthropic_bods_rework_ai_damage