OpenAI Economic Research published the working paper Work at the Frontier: How AI Is Expanding What People Do at Work in July 2026, authored by Caroline Chin and Alex Martin Richmond. Drawing on a sample of more than 800,000 work-related messages from U.S. ChatGPT users, the paper introduces a concept that deserves serious attention from enterprise-service and labor-market researchers alike: task crossover. The report's central proposition is not that "AI is replacing someone's job," but rather that "AI is changing who performs which tasks." Written from the vantage point of a corporate research institute, this article combines the task-based labor economics framework, OpenAI's productized measurement methodology, and real-world use cases to unpack the report's propositions, data, and limitations layer by layer, and extends the discussion to implications for organizational design and the evolution of enterprise-service products.
From "Occupation" to "Task Portfolio": A Paradigm Shift
The report's most important theoretical contribution lies in re-anchoring the debate on AI and employment—from whether occupations will survive to how tasks are redistributed across occupations. This reorientation is not original to the report, yet it gains a new empirical footing from its object of measurement: it observes directly the actual behavior of workers assisted by AI, rather than inferring model capabilities from a fixed checklist of tasks.
The report explicitly invokes the academic lineage of the task-based framework. Autor, Levy, and Murnane (2003) analyzed the impact of computer technology at the level of tasks rather than occupations. Acemoglu and Restrepo (2019) further argued that technology reshapes job demand by reallocating existing tasks and creating new ones in which labor holds a comparative advantage. Gans (2026) identified the framework's inherent limitation—its default assumption that the set of tasks each worker performs is fixed. OpenAI's contribution is to use real usage data to show that this fixedness assumption is being broken: occupational boundaries can be softened.
For enterprises, this means job redesign should not stop at "bolting AI tools onto legacy responsibilities," but should proactively re-examine the boundaries of the task portfolio. When employees in sales, design, and human resources routinely draw on engineering and marketing tasks, the lag of the job description (JD) itself becomes an organizational cost—and organizations must recalibrate "who owns what" far more frequently.
The Scale of Task Crossover and Its Methodological Foundations
The core figures the report presents are robust enough to support the claim that task crossover is a pervasive phenomenon rather than a marginal behavior. Yet its measurement scope must be understood precisely, or it is easily misread as "nearly half of all work has crossed boundaries."
Among all 800,000 work-related messages, 16.8% were classified as cross-occupation, 21.8% as within occupation, and 61.5% as generic (e.g., writing emails or scheduling meetings). The crucial point is that once generic work is excluded, the cross-occupation share among "occupation-specific messages" rises to 43.5%—meaning roughly one in every two non-generic tasks falls outside the user's historical occupational boundary. Across the eight occupation groups, five exceeded the halfway mark for cross-occupation share: customer experience (77%), design (75%), human resources (69%), legal (56%), and marketing (53%).
Measurement scope | Cross-occupation share | Notes |
All work-related messages | 16.8% | Includes 61.5% generic messages |
Occupation-specific messages (generic excluded) | 43.5% | Mean across eight groups; range 11%–30% |
Customer experience / Design / Human resources | 77% / 75% / 69% | Three groups with the highest cross-occupation share |
Legal / Marketing | 56% / 53% | Also above half |
The implications of this data for enterprise services are clear: the value proposition of AI assistants is expanding from "boosting efficiency in one's own role" to "filling capability gaps in adjacent functions." For product builders, the breadth of task coverage may explain adoption in small and mid-sized teams better than point-solution depth—and this is the important premise for the organizational-scale effects developed in later sections.
Which Tasks Travel Across Occupations
Task crossover does not diffuse uniformly; rather, it exhibits a clear emit–absorb asymmetry. Understanding this structure carries more managerial and product-oriented value than speaking loosely about "work being broadened."
The report decomposes the flows into two directions. The first is "borrowed-in tasks": 35.2% of messages from designers involve work of other occupations, yet designers' own tasks account for only 1.7% of messages from other occupations—an extreme absorption profile. By contrast, only 18.5% of engineering messages cross boundaries, but engineering tasks appear in 7.4% of messages from other occupations, making engineering a steady task exporter. The second is marketing—the only other group active in both directions: 24.3% of marketers' messages cross over, while marketing tasks constitute 8.9% of messages from other occupations, the highest in the entire sample. Specific high-frequency cross-occupation tasks include calculating financial data (Finance → 7/7 groups; 14.3% among sales), troubleshooting software (Engineering → 7/7 groups; 7.6% among customer experience), creating marketing materials (Marketing → 5/7 groups; 13.6% among design), communicating product information to customers (Customer Experience → 7/7 groups; 70.8% among marketing), and liaising with government agencies (Legal → 7/7 groups; 21.4% among customer experience).
To organizational designers, this flow map functions as a capability heatmap. Engineering and marketing are the de facto sources of tasks, whose output is diluted across the entire company through AI; design, customer experience, and human resources are capability sinks that depend most heavily on borrowing from outside. Resource allocation, training design, and the build priorities of internal enablement platforms should be differentiated accordingly, rather than spreading an equal AI-capability budget across every function.
Organizational Scale as a Moderator under Resource Constraints
The intensity of task crossover is inversely related to enterprise size, yet this effect depends heavily on users' intensity of use—suggesting what we observe may be a resource-substitution mechanism rather than a universally valid law.
The report focuses on "typical users" in the middle 50% by message volume and finds that the cross-occupation share declines as workspace size grows: 18.9% for workspaces with 2–5 seats, falling to 16.3% for those with 101 or more seats—a relative drop of about 13% (roughly 2.5 percentage points). The report's interpretation is that small organizations lack readily available internal specialists to delegate to, so occasional or moderate users turn to AI to handle tasks that would otherwise require someone else—a small business uses AI to draft marketing copy, troubleshoot software, review contracts, or perform basic analysis. Notably, among the top 25% "heavy users" by message volume, this gradient disappears; they may have already formed stable workflows that are highly similar across organizations, or they may use the additional volume to iteratively refine core tasks (e.g., coding, editing, analysis) rather than continually expanding into new functions.
For enterprise-service providers, this means the retention logic of small and mid-sized customers may be fundamentally different from that of large customers: the former buy "capability substitution" (replacing missing specialist functions), while the latter buy "capacity augmentation" (embedding into existing workflows). A product's onboarding, value narrative, and success metrics should be designed in tiers accordingly, rather than applying one enterprise template to every customer segment.
From Lagging Indicator to Leading Signal
The report's most policy- and strategy-relevant insight is that it reveals AI usage data can serve as a leading signal of shifts in occupational structure—something official statistical systems represented by O*NET and the BLS struggle to provide.
In its conclusion, OpenAI states plainly that government datasets offer a valuable historical snapshot of "how work is divided across jobs," but if AI changes "who performs specific tasks," metrics based solely on existing job descriptions will gradually diverge from how work is actually organized. Prior studies by Atalay et al. (2020) and Autor et al. (2024) observed task changes only after job titles, job postings, or employment patterns had already shifted; by contrast, generative-AI usage data is an earlier-stage window—workers can begin experimenting with and recomposing tasks before firms rewrite JDs or invent new titles. Research by Yang et al. (2026) based on Perplexity queries corroborates this boundary-crossing tendency, indicating the phenomenon is not unique to the ChatGPT platform.
For labor policy, career development, and organizational governance, the key question is no longer "which jobs will be eliminated," but "which new human–machine task combinations become viable, and who is permitted—and paid—to perform them." This requires enterprises to build their own task-flow monitoring mechanisms, rather than passively awaiting annual job inventories or third-party reports to reveal changes in their capability structure; first movers will gain a temporal advantage in organizational orchestration.
Governance, Accountability, and Role Redesign
The report repeatedly emphasizes that its findings are descriptive rather than causal, and cautions that "tasks becoming easier for non-experts to attempt" does not mean "experts are no longer needed." This restraint precisely points to the real difficulty of enterprise adoption: governance and accountability after capability spillover.
Methodologically, the report built a reproducible hierarchical classification pipeline: messages are first assigned to an Intermediate Work Activity (IWA), then mapped via IWA title embeddings to a Detailed Work Activity (DWA), with O*NET's historical task distribution serving as the occupational-boundary baseline and a cosine similarity of 0.80 used to determine boundary crossing; finally, each message is sorted into one of three categories—generic, within-occupation, or cross-occupation. This design ensures transparency of classification, but it also exposes clear limitations: the unit of analysis is "a message," not "a completed task" or "a job"; it cannot observe whether the output was adopted, how good it was, or whether it was reviewed by an expert, nor does it estimate employment or productivity effects. The sample draws on self-reported occupations from ChatGPT Business and cannot be generalized to Enterprise users, let alone to the national workforce.
For product teams, these very limitations constitute a roadmap: in enterprise scenarios, the most valuable next-generation capability is not "enabling more tasks to be done by AI," but "quality review and accountability for boundary-crossing output"—for example, when a sales-generated financial estimate enters the decision chain, the system should indicate who must review it and what audit trail to retain. For organizations, the report's open question—whether repeated boundary-crossing use will harden into responsibility changes—requires HR to design retraining and accountability processes in advance rather than remediate after the fact; otherwise capability spillover will metastasize into compliance and quality risk.
Conclusion
The value of OpenAI's report lies not in predicting the demise of any occupation, but in using rigorous descriptive evidence to turn the abstract claim that "work is being recomposed by AI" into a measurable, observable phenomenon.
From the 43.5% crossover rate among occupation-specific messages, to the asymmetric flows of design absorbing while engineering and marketing export, to the stronger boundary crossing in small organizations—three evidentiary threads converge on one conclusion: occupational boundaries are being softened by AI, while official measurement systems lag systematically behind. This judgment is especially critical for traditional enterprises that rely on hierarchical job ladders.
For the enterprise-service industry, the real window of opportunity lies not in "replacing experts" but in "letting non-experts safely expand their capabilities while experts focus on judgment and review." Whoever can transform "task crossover" from spontaneous individual behavior into a governable, measurable, and accountable organizational capability will establish a structural advantage in the next phase of enterprise AI competition.