Remote-Capable Knowledge Work Should Default to AI-Enabled Flexibility

1. Introduction

Return-to-office debates are often framed as clashes of preference, but the deeper variable is coordination technology. As outlined in Figure 1, for remote-capable knowledge work, the central question is whether the workflow-integrated foundation-model stack now lowers the search, memory, handoff, and context-reconstruction costs that once made daily co-presence the safe organizational default. Remote-capable knowledge work should default to AI-enabled flexibility, not by abolishing offices, but by making blanket attendance requirements increasingly hard to justify outside tasks with high tacitness, high coupling, or high relational stakes.

This is a position paper: it synthesizes evidence from labor economics, organizational research, and deployed AI systems rather than presenting a new causal field experiment.

Offices became central not merely because managers valued visibility, but because co-presence historically reduced the transaction and coordination costs of finding expertise, aligning ambiguous tasks, transferring norms, and evaluating progress [1,2,3,4,5]. In this paper, knowledge work means work whose primary inputs and outputs are symbols, judgments, analysis, writing, coding, design, advising, and decisions rather than repeated physical transformation of materials [6,7]. We focus on remote-capable knowledge work: knowledge work whose core tasks can, for substantial periods, be executed through digital artifacts without continuous on-site physical presence [8]. The pandemic demonstrated that such work can be physically decoupled from the office, but it also revealed the frictions of distributed work: informal mentoring degrades, intra-organizational networks silo, and exploratory creative search can suffer under purely virtual communication [9,10,11,12]. At the same time, rigid attendance is increasingly costly in a workforce shaped by caregiving burdens, disability, and periodic disruption.

What changes the analysis is not “AI” in the abstract, but a specific technical configuration. By foundation-model stack, we mean workflow-integrated foundation-model systems that combine natural-language interaction, multimodal capture, long context, retrieval, transcription, translation, and increasingly bounded tool use [13,14,15,16,17]. Earlier digital tools typically improved one coordination stage at a time—search, storage, messaging, or automation. The foundation-model stack compresses production, comprehension, memory, and routing within the same workflow, which makes distributed coordination materially cheaper. Table 1 fixes the paper’s terminology and scope.

Here, “default” means a rebuttable presumption rather than a universal mandate: for remote-capable, artifact-rich tasks, blanket attendance should require task-specific justification or evidence that learning, mental health, reliability, or worker agency would otherwise be harmed. This is no longer a speculative argument from laboratory capability alone. By late 2024, 45% of Americans ages 18–64 had used generative AI, 27% of employed respondents used it for work in the previous week, and 10% used it every workday [18]. Follow-on 2025–2026 worker and firm surveys show large cross-country gaps and make clear that adoption depends not only on task exposure but also on management practices and whether firms actively encourage use [19,20]. The relevant question is therefore not whether these systems exist, but which coordination frictions they now make cheap enough to reorganize.

This paper advances four linked claims. First, flexible work has already stabilized as a structural equilibrium rather than a temporary pandemic artifact. Second, the foundation-model stack matters mainly because it creates and upgrades artifact capital, thereby reducing coordination costs rather than merely automating isolated tasks. Third, the deeper organizational effect is on latency, modularity, and control over the memory layer of the firm: these systems change how quickly work moves, how cleanly tasks decompose, and who can inspect or contest the resulting artifacts. Fourth, the implication is not uniform remote work but selective co-presence: offices remain valuable, but for narrower and more intentional purposes such as onboarding, apprenticeship, trust formation, conflict repair, design sprints, and tightly coupled creative work. The appendix provides falsifiable predictions, adoption notes, workflow cases, climate notes, guardrail design, a short related-work positioning section, and an extended literature map to make the position concrete rather than rhetorical. Appendix L situates the argument relative to adjacent work on agent harnesses, governed artifacts, and human–AI productivity measurement.

2. Flexible Work Has Stabilized Rather Than Vanished

A useful baseline is not universal remote work but selective flexibility in remote-capable knowledge work. Many jobs cannot be done remotely at all, and even remote-capable jobs include tasks that benefit from periodic co-presence [8]. Still, three facts now make it difficult to treat attendance as the neutral default. Workers place real monetary value on flexibility [21]. Post-pandemic work patterns have stabilized above 2019 levels [22,23,24,25,26]. And hybrid is now the modal arrangement for many remote-capable jobs rather than a short-lived transition state [27,28]. Figure 2 summarizes that persistence.

Persistence reflects both revealed firm behavior and worker demand. Workers place real monetary value on flexibility [21], and international survey evidence suggests that preferences for remote work remain widespread rather than uniquely American or pandemic [25,26]. The speed and shape of this equilibrium will vary across countries, sectors, and firm types because digital infrastructure, labor regulation, and management practice vary materially [19,20,26]. Once AI lowers the coordination penalty of distributed work, flexibility becomes part of the effective employment contract. Firms that insist on high-frequency attendance are then not merely choosing a culture; they are implicitly choosing a narrower talent pool, higher compensating differentials, or higher attrition risk.

The performance evidence is likewise more nuanced than either side of the public debate usually admits. Meta-analytic and review work already showed that telecommuting effects are heterogeneous, with benefits depending on autonomy, communication quality, and the fit between task design and work arrangement [29,30]. Since 2020, field studies have strengthened the case. In the Chinese call-center experiment, working from home increased measured performance and reduced attrition, although with career frictions for home workers [31]. Hybrid work at Trip.com improved satisfaction and reduced quits without harming performance [32]. Geographic flexibility raised output among patent examiners [33], and well-designed “smart working” increased both productivity and well-being [34]. Even studies that find costs typically point to mentoring, collaboration, or knowledge spillovers rather than to any generic inability to work at distance [9,35].

That evidence does not say that all jobs should become fully remote. It does say that rigid attendance is no longer the neutral baseline against which flexibility must plead for exceptions. Flexible work is already durable enough that the strategic question has changed from whether it is viable to which frictions still prevent it from scaling. AI enters at precisely that margin.

3. Why the Foundation-Model Stack Matters: Cheaper Coordination

The distinctive contribution of the foundation-model stack is not that it suddenly makes flexible work possible. Cloud software, search, version control, and messaging already did that. The discontinuity is that a workflow-integrated stack now lowers the cost of producing, understanding, storing, and routing artifacts inside the same routine. Natural-language interfaces lower self-service costs; multimodality turns speech, screenshots, slides, and mixed-format documents into machine-readable inputs; long context and retrieval make prior work more queryable; and bounded agents begin to move workflow state across tools [13,14,15,16,17]. By late 2024, work use of generative AI had already reached 27% of employed U.S. adults in the previous week, with estimated time savings equivalent to about 1.4% of total work hours [18]. Follow-on 2025–2026 surveys show large cross-country adoption gaps and highlight management encouragement as a meaningful complement to technical capability [19,20].

The recent wave has unfolded in layers rather than in one leap. The first layer was conversational assistance: workers could ask for a draft, explanation, rewrite, or code snippet in natural language, which made AI a practical general-purpose assistant rather than a specialist tool and sharply lowered the cost of self-service problem solving [18,36]. The second layer was workflow embedding and multimodality: systems increasingly appeared inside meetings, documents, productivity tools, and code workflows, while also parsing speech, screenshots, slides, tables, and mixed-format context [13,14,37]. The third layer was long context plus retrieval, which made decisions, tickets, repositories, and policy documents far more queryable across time [14,15]. The fourth layer is bounded agents: systems increasingly gather evidence, draft updates, and move workflow state across tools, so the gain is no longer only faster writing but lower follow-through latency [16,17]. Evidence on worker demand suggests that these coordination components are among the tasks workers most want AI agents to support [38]. Each layer removed a different reason people once needed the same room every day. Related systems work on harness engineering makes the same point from the agent side: tool mediation, runtime control, verification, and trace retention belong to the effective capability layer that organizations deploy, not merely to prompting style [39].

Crucially, the paper’s core claim does not depend on frontier-level agent reliability. The strongest evidence today is for transcription, summarization, retrieval, translation, first-pass drafting, and code assistance [36,37,40,41]. Bounded agents amplify the argument by lowering follow-through latency, but the repricing of co-presence begins before agents become fully robust. What matters for AI-enabled flexibility is therefore not text generation in the abstract. It is the cumulative reduction of distributed-work frictions through artifact capital. Table 2 summarizes where the new stack substitutes for historical office functions and where co-presence retains an advantage.

We refer to the resulting stock as artifact capital: durable, searchable, reusable representations of work state such as summaries, decisions, design rationales, code comments, exemplars, translations, retrieval layers, and structured handoffs that reduce the cost of future coordination. The idea is consistent with classic accounts of the firm as a response to transaction and knowledge problems [1,2,5], with theories of tacit knowledge and knowledge creation [42,43,44,45], and with coordination research emphasizing expertise location, common ground, and shared memory [4,46,47,48].

A compact way to express the mechanism is given in Equation 1:

$$V\left(m\right)=B_{\mathrm{trust}}\left(m\right)+B_{\mathrm{learning}}\left(m\right)+B_{\mathrm{creative}}\left(m\right)-\sum _{k\in K}{C}_k\left(m\right),$$

(1)

where $K=\left\{\mathrm{search},\mathrm{handoff},\mathrm{memory},\mathrm{monitor},\mathrm{space}\right\}$. In this equation, m represents a specific work mode (e.g., asynchronous remote work, hybrid work, or dense co-presence), and $V\left(m\right)$ is the net organizational value of that mode. The B terms represent the benefits that historically favored the office: $B_{\mathrm{trust}}\left(m\right)$ is the value of interpersonal trust formation, $B_{\mathrm{learning}}\left(m\right)$ captures apprenticeship and tacit skill transfer, and $B_{\mathrm{creative}}\left(m\right)$ denotes the output of tightly coupled creative search. The final term subtracts the sum of coordination costs, $C_k\left(m\right)$, across the set of organizational frictions K. AI mainly lowers these $C_k$ costs for artifact-rich tasks by reducing the expense of finding prior decisions, reconstructing context, preparing handoffs, and producing inspectable outputs. It does much less to erase the B benefits of trust, apprenticeship, and tightly coupled search. The result is not “remote wins,” but a repricing of where the office still earns a large premium.

The deeper organizational effect is often on latency rather than on average task speed. Offices historically earned a premium because they collapsed waiting times: the missing answer, quick clarification, or informal escalation that prevented a task from stalling for a day. AI copilots and retrieval layers attack that tail latency directly by producing first-pass summaries, surfacing precedents, drafting follow-ups, and routing the next action before a colleague is free. Even modest task-level savings can therefore propagate into larger coordination gains when they shorten queues and reduce idle waiting.

Four channels matter most. First, Context Portability: AI makes context portable by turning ephemeral discussions into searchable summaries and action items, enabling enterprise retrieval over past specs, tickets, code reviews, and policies, and extracting structure from multimodal inputs like slides and forms—addressing coordination failures that stem less from missing talent than from missing retrieval and shared context [49,50,51]. Second, Communication Compression: AI compresses communication overhead by rewriting, translating, summarizing, and drafting, reducing both the blank-page and reconstruction taxes; evidence from the Microsoft 365 Copilot field experiment shows workers spending about three fewer hours (roughly 25%) less on email weekly without significant changes in meeting time, consistent with AI first lowering individually adjustable coordination burdens [37]. Third, Expertise Diffusion: AI diffuses expertise by delivering disproportionate gains to less experienced workers or those facing high setup costs, as seen in customer support (+15% productivity, especially for lower-skilled workers), software development (+26.08% completed tasks), and broader experiments where AI substantially narrows education-based productivity gaps—indicating that workers can access patterns and guidance without needing immediate proximity to experts [40,41,52,53]. Fourth, Coordination Automation: AI begins to automate coordination preparation and follow-through through agentic systems that collect evidence, prefill approvals, update systems, and draft plans, while humans retain control over judgment and exceptions; however, as a system technology, AI’s value depends on complementary organizational redesign rather than standalone deployment [20,38,54,55,56].

The practical effect is cultural before it is fully macroeconomic. Conversational assistants lower the threshold for self-service problem solving; workflow-integrated multimodal copilots turn meetings and drafts into reusable artifacts; long-context retrieval reduces the gatekeeping value of sitting near the person who “knows where things are”; and bounded agents reduce follow-through costs across tools. The net result is a shift from “ask the person” to “query the work,” from meetings as memory to meetings as escalation, and from commute rhythms to selective co-presence organized around high-coupling tasks. In the Microsoft 365 Copilot field experiment, treated workers reduced time spent on email and also spent less time working outside regular hours [37]. Survey evidence also suggests that the durable endpoint is more likely hybrid than permanently all-remote: workers continue to value flexibility, yet fully remote workers report worse well-being than hybrid workers [26,57,58].

Current scientific evidence also suggests that this wave changes execution before it fully changes occupations. In software development, AI already writes an estimated 29% of Python functions in the United States and is associated with a 3.6% increase in quarterly online code contributions, yet the measured gains accrue mainly to experienced, senior-level developers while early-career developers show no significant benefits [59]. That is exactly what one would expect if current systems are strongest at accelerating artifact production, review, and exploration rather than replacing judgment about architecture, integration, or exceptions.

4. Deeper Implications: Latency, Modularity, Control Rights, and Culture

Four structural implications follow from this shift toward artifact-based coordination. First, Tail Risk Compression: AI fundamentally alters the economics of tail risk in coordination by surfacing precedents, reconstructing context, and drafting follow-ups, thereby reducing reliance on physical co-presence and enabling compounding team-level productivity gains even when individual tasks save only minutes. Second, Modularity Premium: AI disproportionately rewards decomposable, legible work, meaning teams with clear ownership, stable interfaces, and retrieval-ready documentation benefit far more than those reliant on undocumented interactions—reinforcing that AI, as a complementary system technology, delivers the greatest value when workflows are redesigned around durable artifacts rather than merely adopting models [20,54,55,56]. Third, Memory Layer as Strategic Asset: as machine-readable artifacts become the substrate of coordination, governance shifts from physical arrangements (“who sits where”) to data control (“who can write, inspect, and retain artifacts”), enabling scalable flexibility when properly permissioned but risking software-enforced presenteeism if misused [60,61,62,63]. Fourth, Cultural Encoding Shift: organizational culture, once transmitted through informal in-person interactions, is increasingly encoded in what is summarized, cited, and made searchable, enabling calmer asynchronous work when designed well but potentially creating pressure for constant algorithmic visibility when mismanaged—reframing the debate from office versus remote culture to high-trust artifact cultures versus low-trust digital presenteeism [60,61,62,63].

Ultimately, the foundation-model stack represents a two-stage reduction in coordination friction. Generative copilots first lowered the cost of producing and comprehending artifacts. Now, bounded agents are lowering the cost of moving workflow state across tools. The cumulative effect is profound: a critical mass of knowledge work can now be governed by shared memory and supervised workflows, undermining the historical necessity of habitual, daily co-location.

5. A Task-Based Theory of Selective Co-Presence

The strongest version of this position is not “remote work wins.” It is that co-presence becomes a scarce input allocated where tacitness, coupling, and relational stakes are highest. A useful mental model is a $2\times 2$ over tacitness and coupling, consistent with classic ideas about task equivocality and media fit [64], as depicted in Figure 3. AI helps most when the missing ingredient is explanation, retrieval, or first-pass structure, and least when the missing ingredient is trust, taste, conflict repair, or political judgment. A practical decision rule follows: remote-capable, artifact-rich tasks should default to AI-enabled flexibility unless tacitness, coupling, or relational stakes are high enough to justify selective co-presence.

Low tacitness, low coupling. These tasks are the clearest candidates for asynchronous flexibility: drafting, routine analysis, well-specified coding, triage, document revision, and common customer interactions. They are already artifact-mediated, and AI further lowers their search, handoff, and setup costs. Low tacitness, high coupling. These tasks still benefit from synchronization, but not necessarily from daily co-presence. Hybrid blocks, coordinated overlap windows, and AI-supported handoffs are often sufficient. High tacitness, low coupling. These tasks need judgment transfer more than constant interaction: coaching, calibration, some forms of review, and identity-forming early-career work. The right response is usually periodic and intentional co-presence rather than blanket attendance. High tacitness, high coupling. These are the strongest residual cases for dense co-presence: early-stage team formation, crisis response, sensitive negotiation, conflict repair, onboarding waves, and creative synthesis under ambiguous objectives.

This framework explains why negative findings about remote work do not refute the thesis. If remote work reduces spontaneous cross-team ties or creativity [9,10,12], the implication is not that work should be colocated. It is that firms should purchase co-presence when they need bridging ties, joint search, or repair. Likewise, if proximity helps training and long-run skill formation [11], firms should design richer in-person windows for novices rather than impose uniform daily attendance on everyone. Flexible work becomes durable only if evaluation systems stop penalizing it through visibility bias or flexibility stigma [65,66,67]. That is the sense in which “default” is used throughout the paper: a rebuttable presumption, not a mandate.

6. Alternative Views

The claim that remote-capable knowledge work should default to AI-enabled flexibility is an empirically falsifiable position, not an ideological absolute. While the current trajectory favors this transition, the equilibrium could fail under specific socio-technical conditions (detailed fully in Appendix C). Below, we address the most significant counterarguments to this framework.

• View 1: The Frontier Task Objection.

Some argue that AI will strengthen the rationale for the office because the hardest, most valuable frontier tasks remain highly interactive and ambiguous. If AI automates the routine execution, what is left is high-stakes, tightly coupled human brainstorming, which benefits immensely from physical proximity and high-bandwidth, in-room collaboration.

Response: This is a serious objection and likely true for specialized, frontier-research teams. However, it is insufficient as a general equilibrium for the broader knowledge economy. For most roles, the average day does not consist entirely of frontier brainstorming; it consists of triage, drafting, search, status updates, code review, and asynchronous handoffs. Those are precisely the tasks whose coordination tax is falling most quickly. The paper does not claim the office vanishes for frontier work; it claims the office is no longer the default for the massive volume of routine production.

• View 2: The Digital Taylorism Objection.

A highly critical view posits that firms will use AI not to enable geographic flexibility, but to intensify monitoring and algorithmic management. If work becomes perfectly legible to software via “artifact capital,” managers may use these systems to recreate the office in software through continuous surveillance, keystroke tracking, and automated performance ranking, negating any gains in worker autonomy.

Response: This is less a rebuttal of the paper’s mechanism and more a competing—and highly concerning—implementation path. The capability to perfectly summarize and retrieve work state can absolutely be co-opted for surveillance. Our claim is therefore conditional: AI-enabled flexibility scales sustainably only when organizational governance tilts toward memory, support, and output-based accountability rather than toward continuous behavioral ranking. This makes privacy-preserving AI design a core labor issue.

• View 3: The Skill Atrophy and Apprenticeship Objection.

Recent empirical evidence suggests that while AI drives short-run output gains, it can coexist with worse long-term learning. By outsourcing search, debugging, and drafting to AI, junior workers bypass the struggles through which deep expertise is built. Therefore, placing junior workers in remote environments with AI copilots risks generating a “lost generation” of talent that lacks foundational tacit knowledge.

Response: This concern is scientifically robust [68]. However, the solution is not a mandatory five-day return to the office for all employees. Instead, firms must distinguish execution support from apprenticeship. The framework of selective co-presence explicitly reserves intentional, in-person time for mentoring and judgment calibration. Furthermore, the ML community must build “scaffolded” AI that prioritizes explanation, Socratic dialogue, and calibrated uncertainty for novices, rather than just output generation.

• View 4: The Mental Health and Social Isolation Objection.

A growing body of psychological and sociological research argues that remote work strips away the social fabric of the workplace, leading to increased rates of depression, loneliness, and anxiety. The office serves a latent function as a primary source of adult social connection; removing it degrades long-term mental health, well-being, and ultimately, organizational loyalty.

Response: This is a crucial limitation of the “all-remote” ideology, which is exactly why this paper argues for a hybrid flexible default rather than total virtuality. Fully remote workers do report higher rates of loneliness than their hybrid counterparts [57]. But forcing daily commutes for quiet, headphone-bound work does not solve loneliness; it repurposes time poorly. Selective co-presence is better aimed at social bonding, trust formation, and team building than either strict isolation or forced, performative attendance.

• View 5: The Bounded-Agent Brittleness Objection.

Critics point out that AI agents hallucinate, lose context, and fail in enterprise environments. Because bounded agents cannot reliably move workflow state without errors, the coordination tax is not actually falling; rather, humans must remain tightly coupled and available to catch cascading AI mistakes and manage exceptions.

Response: This is true of early bounded-agent systems. But AI does not need perfect reliability to shift the equilibrium; it only needs to reduce follow-through costs relative to manual coordination, with humans reviewing exceptions and supervising the resulting artifacts.

• View 6: The Macroeconomic Lag Objection.

Linked survey and administrative evidence frequently finds that rapid technology adoption coexists with null effects on aggregate earnings and recorded hours for years after deployment. Therefore, assuming that AI will immediately disrupt the deeply entrenched, macroeconomic real-estate and labor structures of the office is premature.

Response: This caution disciplines the timeline, not the mechanism. Communication norms, meeting loads, and hybrid schedules can shift faster than aggregate wages or real-estate totals.

7. A Socio-Technical Research Agenda

If the argument advanced in this paper holds, AI-enabled flexibility is a concrete systems-and-evaluation problem for the machine learning community rather than a peripheral enterprise use case. The field needs benchmarks for handoff fidelity, context reconstruction, memory provenance, escalation quality, rollback after bad bounded-agent actions, after-hours spillover, and carbon per resolved task; Appendix Table A5 and Table A10 summarize the capability classes and evaluation targets. We therefore propose the following socio-technical agenda so the shift enhances autonomy, learning, and sustainability rather than hardening into software-mediated presenteeism.

• 1. Team-Level and Longitudinal Evaluation Paradigms.

Current ML benchmarks overwhelmingly evaluate static, single-agent, or single-turn task completion. Flexible work, however, succeeds or fails through handoffs, shared memory, escalation, and coordination across time. The community should develop dynamic benchmarks that measure context reconstruction time, handoff fidelity, asynchronous onboarding speed, decision-provenance coverage, and retrieval accuracy over organizational history. We also need longitudinal field studies to track interruption load, meeting displacement, and whether AI adoption reduces coordination burden or merely displaces it into after-hours spillovers [37,57]. Related work on human–AI productivity measurement argues that deployment-relevant gains should be reported as time-to-acceptance under explicit acceptance tests, which is closely aligned with our emphasis on verification, rework, and governed handoffs rather than raw draft speed [69].

• 2. Scaffolded AI for Apprenticeship and Skill Formation.

AI-enabled flexibility changes who can access labor markets, potentially widening geographic matching and disability inclusion [33,70]. Yet the gains are distributionally uneven. If AI primarily accelerates experienced workers while automating the intermediate steps necessary for junior workers to learn [59,68], distributed work risks exacerbating a “novice penalty” and driving long-term skill atrophy. A critical research direction is designing scaffolded copilots—systems that prioritize calibrated explanations, Socratic dialogue, and reliable escalation to human experts rather than only final-output generation.

• 3. Privacy-Preserving and Interoperable Memory Layers.

If artifact capital becomes the substrate of coordination, the strategic asset is the organizational memory layer. The governance risk is that this layer devolves into a proprietary, panoptic surveillance apparatus [62,63]. We should build enterprise RAG and memory systems where provenance, access control, reversible summaries, and retention limits are formalized design constraints rather than compliance afterthoughts. Open schemas for interoperable artifacts are also essential to prevent cognitive and infrastructural lock-in within singular vendor ecosystems.

• 4. Carbon-Aware Routing and Systems-Level ESG Accounting.

The environmental case for flexible work is doubly material: it affects corporate real estate resilience and societal Scope 3 emissions [71,72,73,74]. However, the climate dividend is easily wiped out if commuting emissions are simply replaced by highly compute-intensive generative models and inefficient residential conditioning [75,76]. The Green AI community should prioritize carbon-aware model routing—sending routine summarization or retrieval tasks to smaller, lower-carbon models or off-peak compute regions [77]. Systems-level research should pair ML adoption with occupancy-responsive building tools so organizations consolidate physical footprints.

• 5. Pro-Agency Guardrails for Bounded-Agent Workflows.

As AI shifts from conversational interfaces to bounded agents that autonomously move workflow state across tools [16,17], the risk of unreviewable algorithmic management intensifies. Anti-surveillance and pro-agency design should therefore be treated as primary evaluation targets. The field should develop standards for inspectable agent actions, previewable multi-step plans, rollback mechanisms, and meaningful human override [78]. Evaluating these systems must answer not only whether they improve throughput, but whether they preserve worker discretion and trust.

8. Conclusion

The foundation-model stack does not make the office obsolete, but it does make blanket attendance a weak default for remote-capable knowledge work. By converting ephemeral interactions into queryable artifact capital, it reduces the historic frictions of distributed coordination and makes a large share of remote-capable knowledge work more asynchronous and portable across time and location. The resulting equilibrium is selective co-presence: intentional in-person time for apprenticeship, conflict repair, trust formation, and other high-tacitness or high-coupling tasks, paired with AI-enabled flexibility for routine execution. The central socio-technical challenge is to ensure that this transition strengthens worker autonomy, learning, and sustainability rather than hardening into software-mediated presenteeism.