Reporters and researchers find mixed evidence on whether AI increases worker productivity. Vox reports clinicians using AI-drafted patient messages often spend extra time reviewing and editing them, with Philip Barrison quoted saying clinicians must decide if the draft "is actually something that they think they would say" and then "edit it to the point where they think it's acceptable" (Vox). The National Bureau of Economic Research paper "Generative AI at Work" found access to a conversational assistant raised productivity by 14% on average and by 34% for novice agents (NBER, 2023). Anthropic's March 5, 2026 research introduces an "observed exposure" measure and finds no systematic rise in unemployment to date, though it links higher exposure to slower projected occupational growth and slower hiring of younger workers (Anthropic). Editorial analysis: practitioners should expect heterogeneous outcomes by task complexity, worker skill, and measurement choices.
What happened
Vox reports clinicians who tried an AI tool for drafting patient-portal messages often faced extra work because they had to review and edit drafts, a task Philip Barrison, an MD-PhD student, describes as judging whether the draft "is actually something that they think they would say" and then "edit it to the point where they think it's acceptable" (Vox, May 26, 2026). Vox also reports companies including Meta set internal targets to "go 5X faster by eliminating the frictions that slow us down" and that managers asked employees to prove they "cannot get what they want done using AI" before approving new hires, per the article (Vox).
The National Bureau of Economic Research (NBER) working paper "Generative AI at Work" (Working Paper 31161, 2023) reports that access to a generative conversational assistant increased productivity by 14% on average and by 34% for novice and low-skilled customer support agents (NBER, 2023). The paper also reports improvements in customer sentiment and employee retention in the tested setting.
Anthropic's March 5, 2026 report introduces a new metric, "observed exposure," combining model capability and real-world usage, and reports that actual AI coverage remains far below theoretical capability; Anthropic finds no systematic increase in unemployment for highly exposed workers since late 2022 but reports suggestive evidence of slower hiring of younger workers in exposed occupations (Anthropic, 2026).
McKinsey research and commentary cited in public reporting place generative AI's economic potential in the trillions, while the Harvard Business Review argues that AI frequently "intensifies" work rather than reducing overall burden (McKinsey; HBR, Feb 9, 2026).
Editorial analysis - technical context
Observed outcomes across studies are consistent with a core technical pattern: generative AI tends to automate substeps of a workflow while introducing oversight and verification tasks. Companies and researchers typically measure productivity as speed or throughput, but AI assistance often shifts time from content generation to review, quality control, and exception handling. Studies that find large gains, like the NBER experiment, generally measure narrowly scoped, repeatable tasks-customer support conversations-where the assistant can apply best-practice text reliably. Studies and reporting that document frictions, like the Vox clinician example, capture contexts where outputs require high judgment or customization.
Industry context
Editorial analysis: practitioners should expect large heterogeneity in AI's effects. Occupation-level measures such as Anthropic's "observed exposure" show that adoption and measurable impact vary by task automability, workplace processes, and worker skill. Organizational choices-how tools are integrated, whether verification is required, and which performance metrics are used-drive whether time saved on one subtask translates into net productivity gains.
What to watch
- •Adoption metrics that separate automated, low-judgment uses from augmentative, high-judgment uses (Anthropic's framework).
- •Quality- and cognitive-load measures beyond raw throughput, including error rates, time spent on review, and worker-reported mental fatigue (Vox; HBR).
- •Heterogeneity by experience level: replication of the NBER finding that novices gain more than experienced workers in new settings.
For practitioners
Editorial analysis: when evaluating AI pilots, measure both speed and downstream verification overhead, track customer outcomes and retention, and collect worker-reported measures of cognitive load. Narrow-scope, repeatable tasks are where current generative assistants tend to show the clearest productivity gains, while high-judgment tasks risk adding review work. Over time, observed exposure metrics and rigorous A/B testing at the task level will be the most useful evidence for whether AI produces net gains in a given role.
Scoring Rationale #
The story compiles experimental and field evidence showing meaningful productivity gains in specific settings alongside real-world frictions. This is important for practitioners designing and measuring AI deployments, so it is a notable, actionable development but not a paradigm shift.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.