Work at the Frontier: task crossover across occupations

Key takeaways

1. ⭐⭐ The two headline figures, and why both must always be given. 16.8% of work-related messages AND 43.5% of occupation-specific messages. The funnel: 61.5% of work usage is generic (writing, summarizing, planning — « shared too broadly across occupations to be evidence of crossover ») and removed from the calculation; of the remaining 38.5%, 43.5% fall outside the occupation. → 43.5% is calculated on 38.5% of usage. Citing only the 43.5% figure nearly doubles the perceived scale. Both numbers are in the source; usually only one survives in coverage.
2. ⭐ The "inside or near" nuance. the complement of the 43.5% is not "within the occupation" but « inside or near a user's occupation ». The measured boundary is thus fuzzy by construction, which is honest and rarely passed on.
3. The methodological counter-move — the paper's real contribution. the literature starts from a fixed list of tasks per occupation and asks whether the model can perform them; here the question is who does what. « Our evidence suggests that AI is also changing who takes on which tasks. » → This measures an observed redistribution, not a theoretical substitutability. This is what distinguishes this work from exposure studies (cf. the 5-60% ranges discussed in [[dgtresor-ia-effets-emploi-2026-06-30]] and [[sfeir-ia-emploi-risque-decrochage-2026-07-23]]).
4. By occupation, outside-occupation share. (of occupation-specific messages): customer experience 77% · design 75% · HR 69% · legal 56% · marketing 53% · sales 40% · finance 40% · engineering 28%. « Outside occupation work is a majority in five of eight groups. »
5. ⭐⭐ The heatmap — the richest object, and absent from all coverage. (rows = user's occupation, columns = task's occupation of origin; diagonal in bold): | User ↓ / Task → | CX | Design | Eng. | Fin. | HR | Legal | Mktg | Sales | |---|---|---|---|---|---|---|---|---| | Customer experience | 11 | 5 | 20 | 14 | 6 | 6 | 26 | 11 | | Design | 6 | 12 | 28 | 10 | 4 | 4 | 28 | 8 | | Engineering | 4 | 4 | 53 | 9 | 4 | 4 | 20 | 2 | | Finance | 6 | 4 | 22 | 23 | 4 | 8 | 25 | 7 | | HR | 9 | 5 | 18 | 16 | 10 | 10 | 23 | 9 | | Legal | 5 | 4 | 17 | 12 | 7 | 31 | 18 | 6 | | Marketing | 10 | 6 | 17 | 11 | 4 | 4 | 36 | 12 | | Sales | 11 | 5 | 18 | 15 | 4 | 6 | 29 | 12 | Three readings the text does not make: 1. The diagonal is the real result. Engineering retains 53% of its own work; customer experience 11%, HR 10%, design 12%, sales 12%. → Some occupations, in their non-generic use of ChatGPT, have almost stopped doing their own job. 2. For four occupations, the marketing column exceeds the diagonal. A designer does more marketing tasks (28%) and engineering (28%) than design (12%); a customer-experience agent does more marketing (26%) than CX (11%); same for HR (23% vs 10%) and sales (29% vs 12%). 3. Two columns absorb almost everything: marketing (18 to 36% everywhere) and engineering (17 to 53%). The other six occupations account for crumbs. ⚠️ Reading caution: the low diagonal does not mean these employees no longer do their own job — it means their non-generic messages are classified elsewhere, on a base that already excludes 61.5% of usage. The tasks specific to these occupations may be precisely those classified as "generic".
6. The two directions of circulation, quantified. design imports 35.2% / exports 1.7%; engineering imports 18.5% / exports 7.4%; marketing imports 24.3% / exports 8.9% — « the highest outward share in the sample ». → Marketing is the only one to accumulate both directions.
7. The two universally borrowed tasks. financial calculation and technology troubleshooting appear in the top 3 of borrowings for the other seven groups. Creating marketing materials appears in five other groups, « especially prominent among design users ».
8. ⭐ The asymmetry of engineering — the most actionable figure for an IT services firm. engineering is both the least porous occupation (28%, imports 18.5%) and a major exporter (7.4%), and its diagonal of 53% is by far the highest. The celebrated porosity is thus largely one-way toward the technical side: other occupations do troubleshooting and technical systems work, the reverse is rare (design exports 1.7%). What this concretely means: AI does not dilute engineering into other occupations — it spreads engineering tasks everywhere else while leaving engineers on their core work.
9. ⚠️⚠️ The size effect is weaker than reported elsewhere — two caveats posed by OpenAI and often lost. 1. The shift from 18.9% to 16.3% holds "among average users" only. 2. « Among the heaviest users, we do not see the same monotonic pattern. » → The trend does not hold for heavy users. OpenAI offers two possible explanations (moderate users at small organizations turn to AI when they encounter work belonging to another function; heavy users have stabilized workflows that resemble each other across organizations, or use AI more intensively within their core occupation). 3. The conclusion is conditional: « AI may be especially useful as a generalist tool where specialist resources are scarce. » → Correction to [[sfeir-ia-frontieres-metiers-skill-based-organisation-2026-08-01]], which presented these 2.6 points as a clean-cut result and the generalist conclusion as an assertion. The source is more cautious than the commentary on it, and the 2.6-point gap I had flagged as the weak link is explicitly bounded by the authors themselves.
10. The claimed epistemic status — early signal. this data reveals recombinations « before firms rewrite job descriptions or create new job titles », and « may provide an early signal of occupational change that conventional labor-market statistics will capture only later ». → A leading indicator, not an employment measure. It says what people attempt, not what organizations ratify.
11. Doctrinal affiliation. the paper draws on the same team's AI Jobs Transition Framework, whose thesis is that many occupations will reorganize — « jobs whose day-to-day tasks could change substantially » — rather than disappear. Crossover is the usage-level evidence for it.
12. ⚠️ Substantive caveats, to be raised systematically.
13. OpenAI measures OpenAI. A single assistant, its own users only, presented as an advantage (« our unique window »). What is measured is ChatGPT usage, not AI usage.
14. U.S. users only. No international extrapolation is possible, and the report does not attempt one.
15. Non-representative population. of the workforce: these are ChatGPT users, self-selected.
16. The occupation classification is done by a model. , on messages, with no external validation described in the post. The categories (eight groups) and the "generic / specific" boundary are design choices that mechanically determine the results — this is where the robustness of the 43.5% figure is actually decided.
17. Stated aim. « to guide policy and practice ». An actor that produces the data on its own impact and aims to inform public policy. The same bias flagged regarding Bercy in [[sfeir-ia-emploi-risque-decrochage-2026-07-23]].
18. Meta / cross-references. primary source for [[sfeir-ia-frontieres-metiers-skill-based-organisation-2026-08-01]], which it supplements (16.8%, marketing 24.3%, heatmap) and corrects (non-monotonic size effect among heavy users); to be read alongside [[dgtresor-ia-effets-emploi-2026-06-30]] and [[sfeir-ia-emploi-risque-decrochage-2026-07-23]] on the gap between theoretical exposure and observed redistribution; extends [[mollick-organizational-theory-agentic-ai-spans-control-2026-02]] and [[mollick-valence-ai-hr-playbook-leader-lab-crowd-2025-07-23]] on the organizational-theory side; HR context in [[bersin-chro-pivotal-role-ai-transformation-2025-10-10]].