AI Task Crossover and the Skill-Based Organization

Key takeaways

1. Key thesis. AI does not merely accelerate work, it redistributes who does what. The blurring of role boundaries is an organizational fact that is already measurable, not a projection. Design consequence: the job description can no longer be the unit of work organization.
2. The robustness argument (the real contribution of the pairing). two opposing methods converge — a preregistered field experiment (causality, small N, controlled context) and a massive usage trace (scale, correlational, no context). Neither alone would suffice; together they cover each other's blind spots. It is this pairing that matters, more than any single figure taken in isolation.
3. P&G — the mechanism (figures).
4. 791 experienced professionals. (R&D and sales), a full day on real product-innovation challenges drawn from their business units, with real stakes (the best proposals presented to executives). 2×2 protocol: alone / cross-functional pair × with / without AI.
5. Individual + AI = pair without AI. +0.37 σ vs. individuals alone, versus +0.24 σ for teams without AI. AI reproduces part of the benefit of human collaboration.
6. Team + AI. roughly ×3 the probability of a solution ranked in the top 10% — the effect plays out in the tail of the distribution, not the average.
7. Silos erased. without AI, sales staff produce commercial content, R&D staff produce technical content. With AI, the distinction disappears — solutions balanced across the full technical-commercial spectrum, with no loss of quality. The authors name the mechanism: boundary-spanning. Employees far removed from product development, alone with AI, reach the level of teams that include a domain expert.
8. Emotional dimension. more enthusiasm and energy, less anxiety and frustration — AI restores part of the social benefit of teamwork.
9. The caveat that holds. AI raises the average quality of ideas, but human pairs without AI remain better at identifying their best idea (~50% correct selections versus 37% with AI). → Human evaluative judgment keeps its place in the loop. Worth comparing with the "generate / verify" split in [[sfeir-code-review-anneau-contraintes-2026-07-30]]: same asymmetry, different domain.
10. OpenAI — the scale (figures).
11. 43.5%. of role-specific messages fall outside the user's own role, once generic tasks (writing, summarizing, planning) are set aside. ⚠️ The reprocessing matters: the figure is not "43.5% of usage" but "43.5% of role-specific usage."
12. By function. customer experience 77%, design 75%, HR 69%, legal 56%, marketing 53% — sales and finance 40%, engineering 28%.
13. Import / export. design imports massively (35.2%) and exports almost nothing (1.7%); engineering does the opposite (imports 18.5%, exports heavily — troubleshooting and technical systems = 7.4% of other functions' messages); marketing does both, with the highest export rate (8.9%). Two tasks appear in the top 3 borrowed tasks across all groups: financial calculation and technology troubleshooting.
14. Size effect. share of out-of-role tasks 18.9% (workspaces of 2-5 roles) → 16.3% (>100 roles). "Where specialized resources are lacking, AI plays the role of the generalist."
15. Claimed epistemic status. an early signal — this data makes the recomposition visible before job descriptions, job titles, and labor-market statistics.
16. ⚠️ The most interesting and least exploited figure (for an IT services firm). engineering is the least porous and most exporting role. The porosity described is thus largely one-way toward the technical side — other roles do troubleshooting and work with technical systems, the reverse is rare (1.7% export for design). For a firm whose core is engineering, this is the most actionable result — and the article draws nothing from it. An open angle for thought leadership.
17. ⚠️ The weak link in the argument — since confirmed against the primary source. the size effect rests on a 2.6-point gap (18.9% → 16.3%), and these values are not commensurable with the 43.5% figure (different calculation base). The claim that "AI plays the role of the generalist" is a generous interpretation of a small gap. → Checked against the OpenAI report ([[openai-work-at-the-frontier-task-crossover-2026-07-27]]): the gap only holds "among average users", and "among the heaviest users, we do not see the same monotonic pattern." OpenAI concludes in the conditional ("AI may be especially useful as a generalist tool") where SFEIR states it as fact. Not to be cited as a strong result.
18. Mollick — the agenda in three steps. (co-author of the P&G study, he himself links the two publications; the cross-functional usage measured by OpenAI is more pronounced than the experiment alone suggested): (1) organizational boundaries are becoming porous, the walls are thinning; (2) companies will have to rethink the division of labor, and "things are getting chaotic right now" — refusing to deal with the change will not make it disappear; (3) properly orchestrated, the recomposition pays off handsomely, both in employee satisfaction and in company performance. Extends [[mollick-organizational-theory-agentic-ai-spans-control-2026-02]] (spans of control, boundary objects) and the "HR is R&D now" line from [[mollick-valence-ai-hr-playbook-leader-lab-crowd-2025-07-23]].
19. SFEIR's response — the Skill Based Organisation. at the initiative of Rosalie Zandona (VP People & Culture). Principle: if tasks are moving around, a fixed job description can no longer be the unit of organization → genuinely operational skill becomes the base unit (up to 13 skills identified per role), deployed wherever the need is concrete, regardless of the job title of whoever holds it. The underlying shift: status-based identity ("I am a manager") → operational identity ("I know how to design complex architectures"), made objective and transparent for each employee. Claimed benefit: task crossover becomes visible, tooled, and valued instead of remaining "informal improvisation in the shadow of the org chart"; and "an employee who knows their strengths also knows what they can delegate to GenAI." Closing figure: according to Deloitte, skill-based organizations are 98% more likely to be perceived as excellent places to grow.
20. ⚠️ Three caveats to raise in pre-sales. 1. Reversed chronology — the SBO shift at SFEIR is documented in Zandona's article of February 19, 2026, i.e. five months before the two publications it is supposed to "answer point by point." The SBO was not adopted in response to this diagnosis; it is reread in light of it. The firm actually owns this ("we made the shift ourselves before recommending it"), but the article's argumentative order reverses the order of the facts. 2. The central logical leap is not addressed — both studies measure task overflow; the SBO responds with a change in the unit of organization. Nothing in the data says that a skill-based organization absorbs crossover better than a role-based one. This is a design hypothesis — plausible, but untested, and not the only possible response: the other common response is the merged role (cf. SFEIR's own "Product Engineer," April 2026), which goes in exactly the opposite direction (fewer, broader roles, rather than more, finer-grained skills). 3. The Deloitte figure measures perception, not performance — "perceived as excellent places to grow." It therefore only covers the satisfaction side of the dual benefit Mollick announces; the performance side of the SBO is quantified nowhere in the article.
21. ⚠️ Source caution.
22. P&G freshness. the result has been circulating since March 2025 (NBER w33641 / HBS WP, under the title "… A Field Experiment on Generative AI Reshaping Teamwork and Expertise"). The article dates it to its publication in Organization Science (2026) and presents it as having appeared "a few weeks earlier" — true for the peer-reviewed publication, not for the result. For a watch note, this is a consolidation, not a novelty.
23. Diverging N. the article states 791 professionals; the 2025 working paper states 776. The gap is unexplained (final sample of the reviewed version?) — to be checked against the Organization Science version before citing the figure.
24. Title variant. "… on Generative AI and Teamwork" (SFEIR) vs. "… on Generative AI Reshaping Teamwork and Expertise" (working paper).
25. OpenAI base. US ChatGPT users — neither a representative sample of the working population, nor multi-assistant. The generalization "close to half of professional usage" applies to this base only. And it is OpenAI measuring OpenAI's own effect — the same source bias flagged by SFEIR itself in [[sfeir-ia-emploi-risque-decrochage-2026-07-23]], not disclosed here.
26. ⚠️ A missing figure, and it changes the scale. the report gives two headline figures — 16.8% of work-related messages and 43.5% of role-specific messages, the latter calculated after removing 61.5% of generic usage. SFEIR keeps only the second, more spectacular one. See [[openai-work-at-the-frontier-task-crossover-2026-07-27]].
27. Meta / cross-references. this is the HR/organizational counterpart to the "decoupling" thesis in [[sfeir-ia-emploi-risque-decrochage-2026-07-23]] — same diagnosis (AI is recomposing work), different lever (organization rather than training). Complementary to the budget angle in [[girard-sfeir-ai4it-vs-ai4business-budgets-2027-2026-06-24]] and the CHRO angle in [[bersin-chro-pivotal-role-ai-transformation-2025-10-10]]. Useful counterpoint: [[shipper-every-after-automation-frame-framer-2026-05-21]] (the recomposition of knowledge work should not be read as a collapse). ⚠️ Disambiguation: "skill" in the SBO sense (an HR unit of organization) has nothing to do with the corpus's Agent Skills ([[agent-skills-anthropic-2025-10-16]]) — an FR/EN homonym that must not be allowed to merge in the graph.