Work redesign

How AI changes the way work is organized — which tasks get reallocated, recombined, or restructured inside a firm.

2025-hasan-model-design-organizational-design adds a complexity-allocation mechanism. Its conceptual GAS framework argues that a simple generalist-AI interface does not eliminate complexity; abstraction shifts it to infrastructure, data and compliance systems, specialized personnel, and workflow design. Redesign therefore has to decide who carries the work of making outputs reliable, not only which visible task moves from a worker to AI. The framework is a design hypothesis rather than evidence that a particular allocation improves organizational outcomes.

2026-kim-mapping-ai-into-production shows work redesign as an observed performance mechanism: when startups discovered more AI use cases, gains appeared mostly in internal tasks, suggesting that value came from reorganizing how work got done inside the firm.

2026-ermakov-designing-post-role-enterprise gives the stronger architecture claim. It argues that static job descriptions are the wrong unit of design for agentic work because AI systems can now plan and execute multi-step cognitive workflows across role boundaries. The proposed alternative is to organize around outcome-based “Fluid Skill Clusters,” assembled from human capabilities and digital labor capabilities, with explicit governance for where human judgment and accountability must remain present.

2026-lee-making-invisible-visible gives the worker-experience version of the same redesign problem. Its interview evidence shows that AI can be technically present in the workflow while badly positioned: workers reported AI taking over tasks they preferred or needed to own, leaving them with tedious backstage work, reducing autonomy and learning, and disrupting collaboration and mentoring. The design unit is therefore not “add AI to a task”; it is the whole sociotechnical workflow — task allocation, control rights, communication, feedback, and accountability.

2026-dellacqua-jagged-technological-frontier adds the capability-boundary version of work redesign. In the same consulting domain, GPT-4 assistance improved performance on some tasks and degraded it on another that appeared comparably complex. This makes redesign less like a binary automation decision and more like frontier mapping: identify which subtasks can be delegated, which require human verification, and where AI output should be blocked or constrained.

2025-dellacqua-super-mario-ai adds causal evidence that work redesign can fail when it treats automation as a plug-in replacement for one team member. In a laboratory coordination task, teams had already built routines over repeated rounds; replacing a human with a high-performing automated agent reduced output, increased bumps/coordination failures, and lowered trust and effort among the human teammate. Redesign therefore has to account for tacit routines, mutual predictability, and motivation, not only the replaced actor’s individual task performance.

2026-dellacqua-cybernetic-teammate adds a team-architecture version of work redesign. In P&G product-innovation tasks, individuals with AI matched non-AI teams on solution quality, while AI also helped R&D and Commercial professionals produce more balanced ideas across functional boundaries. Redesign therefore has to ask when a workflow needs a human team, when a solo worker with AI can reproduce some team benefits, and where human teaming remains necessary for coordination, commitment, accountability, or implementation.

2025-brynjolfsson-generative-ai-at-work shows a narrower but cleaner redesign pattern: AI assistance was inserted into live customer-support chats as recommendations that agents could accept or ignore. The work did not disappear; the agent role shifted toward using machine-supplied language and technical suggestions while retaining final authority over the customer interaction.

2022-dellacqua-falling-asleep-at-wheel adds a friction-design warning. In AI-supported HR screening, the nominal human role stayed intact, but AI quality changed how much effort recruiters invested. Redesign therefore has to specify not only who has final authority, but what interface, feedback, and escalation conditions keep the human sufficiently awake for that authority to matter.

2025-bastani-human-ai-contracting-paradox adds the incentive-design version of that warning. If human effort is unobservable and AI failures are rare, redesigning work around “AI plus human oversight” may require a contract that transfers too much value to the human inspector; the principal may then prefer AI-only operation or no AI use despite lower total system value. Work redesign therefore has to include compensation, monitoring, liability, and escalation economics, not only task decomposition and interface placement.

2026-tomei-what-jobs-can-ai-learn adds a measurement lens for redesign: tasks are more learnable by RL-style post-training when they have clear verification, simulable environments, observable state, dense feedback, accessible tools, and tangible outputs. That makes the redesign question more precise: which workflow components can be turned into learnable task environments, and which resist because the state, feedback, or judgment criteria are not clean enough.

2025-mazeika-remote-labor-index adds a deliverable-level constraint on redesign. Its RLI benchmark tests whole freelance projects rather than isolated subtasks and finds current agents near the floor, with the best automation rate at 2.5%. For work redesign, the implication is that delegation boundaries cannot be drawn from tool capability demos alone; the redesigned workflow has to cover file handling, artifact quality, integration across project components, and acceptance by a reasonable client.

2025-shukla-ai-assisted-design-ironies adds the automation-history version of work redesign. In UX practice, designers expected GenAI to remove repetitive design meta-work and support ideation, but the paper warns that automation can shift work into monitoring, validation, troubleshooting, and defending outputs. The redesign question is therefore not just which design tasks AI can perform, but how AI changes designers’ visibility into rationale, skill practice, accountability, and the flow of creative judgment.

2025-handa-economic-tasks-ai gives a task-demand map for redesign. By mapping Claude.ai conversations to O*NET tasks, it shows where workers are already pulling AI into task execution: heavily in software development and writing, with broader but uneven spread across occupations. For redesign, this is not proof that those jobs should be automated; it is evidence about where task-level pressure and experimentation are already concentrated.

2025-marguerit-augmenting-automating-labor gives redesign a labor-market outcome frame: augmentation AI is associated with new work creation and, for higher-skilled occupations, higher wages, while automation AI is associated with weaker low-skilled outcomes. That suggests redesign is not only about reallocating existing tasks inside a role; it also changes whether new job titles emerge, which skill groups capture the upside, and where organizations may redesign toward substitution rather than capability expansion.

2025-agarwal-what-work-ai-doing adds a more granular task-characteristics map. Observed Claude use was highest for tasks with high creativity, cognitive demand, complexity, and low routineness, and clustered especially in a “Dynamic Problem Solving” archetype. For redesign, this suggests conversational GenAI is being pulled toward high-friction cognitive starts and synthesis work, not only toward routine task removal.

2026-reich-work-design-multidimensional-ai-threat brings classic work-design variables directly into AI adoption. Skill variety and autonomy predicted adoption, while task significance predicted shorter AI-use duration. Redesign therefore has to preserve the job resources that let workers experiment and learn, while recognizing that highly consequential work may invite more selective or bounded AI use.

2025-klonek-ai-work-stress adds an AI-specific work-design distinction: AI control and human–AI team processes should be designed separately rather than collapsed into an automation–augmentation spectrum. Higher AI control was associated with higher stress, but human–AI transition and action processes were associated with lower stress and some team processes buffered the AI-control/stress relationship. For redesign, the choice is not simply “more automation” or “more augmentation”; it is how much decision control the AI receives and what planning, action, interpersonal, and feedback resources remain available to the worker.

2026-ehsan-future-workers makes preservation of expertise a redesign requirement. Its sociotechnical-immunity frame says AI workflows need early-warning signals, containment actions, and recovery routines so workers can detect and reverse skill erosion rather than only deliver faster outputs.

2025-shao-future-work-ai-agents makes worker-desired agency a redesign input rather than an afterthought. Its Human Agency Scale separates full automation from equal partnership and high-touch human involvement; 45.2% of occupations had H3 equal partnership as the dominant worker-desired level. That makes redesign a matching problem among task capability, worker desire, control rights, and the level of human involvement the work should preserve.

2025-yun-generative-ai-knowledge-work gives a concrete redesign pattern for information-heavy product work. The source shows AI being inserted not as a single assistant but as a workbench for Q&A, pain-point tracking, user-insight visualization, source citation, export, and reuse across stakeholders. That makes redesign less about replacing a product-manager task and more about re-architecting the flow from raw user evidence to clustered interpretation, team communication, and accountable prioritization.

2026-doshi-human-ai-task-tensor adds an upstream classification layer for redesign. Its Human–AI Task Tensor says a redesign decision should specify task definition, AI contribution, interaction modality, audit requirement, output definition, decision-making authority, AI structure, and human persona before claiming that a task is augmented, automated, or collaboratively performed. That makes the design unit more precise than “add GenAI”: the workflow has to define authority, auditability, output criteria, and the human role.

2025-kellogg-novice-risk-work adds a risk-mitigation design layer. Junior consultants often recommended human-routine or project-level fixes for GenAI risks, while expert recommendations included system design, deployer controls, and ecosystem interventions. For redesign, that means local training and checking routines may be necessary but insufficient when the risk source sits in model behavior, interface design, deployment governance, or data/ecosystem conditions.

2026-zhu-choose-your-agent adds an interdependence constraint to redesign. In the bargaining experiment, Delegate access changed not only the user’s action but the trading environment faced by others, creating possible spillovers for non-users. In redesigned work, agentic delegation should therefore be treated as a mechanism design problem: who can delegate, what the agent can execute, who bears accountability, and how one person’s delegation changes everyone else’s workflow.

2009-sanchez-levine-competency-modeling-job-analysis is a pre-AI reminder that redesign is not only task description. Its distinction between a job as an object to be described and a role to be enacted matters for AI-era work: new workflows need task evidence, but they also need signals about how people should interpret discretion, strategy, accountability, and contextual performance inside redesigned roles.

2026-wang-agent-development-real-world-work adds a benchmark-validity constraint to redesign. If agent evaluations are concentrated in programming-heavy, easily verified domains, then benchmark progress should not be treated as evidence that agents are ready for management, legal, interpersonal, or other work where objectives, verification, and accountability are less clean. Redesign has to ask whether the evaluation sample resembles the work node being delegated.

2025-wang-ai-agents-human-work adds direct workflow-comparison evidence that similar task sequences can conceal different execution architectures. Across 16 computer-use tasks, human and agent workflows shared 83.0% of steps and preserved aligned-step order 99.8% of the time, yet agents used programming tools in 93.8% of workflow steps while humans relied more on interfaces and perceptual judgment. Redesign therefore has to specify not only which steps remain, but how each step is executed, what representations and tools it depends on, and where the resulting strategy creates new verification or translation work.

2026-vuong-genai-job-performance adds a bottom-up redesign mechanism: employees may use GenAI as a job resource for job crafting before formal role architecture changes. In Vietnamese SMEs, GenAI usage was associated with job crafting and work engagement, which partially mediated job performance. For redesign, the important signal is that local worker-led restructuring can be part of the value pathway, but it also means redesigned work may be happening informally before governance, measurement, or role expectations catch up.

2007-serenko-organizational-size-knowledge-flow adds a pre-AI unit-design constraint: knowledge flow can degrade as organizational units grow because formal structure, weaker trust, reduced connective efficacy, and less frequent communication make it harder for knowledge to reach the right people. For AI-era redesign, that baseline matters because AI knowledge tools may change the effective scale of a unit or the cost of cross-unit exchange, but they do not automatically solve the trust, routing, overload, and communication problems that made size consequential in the first place.

2026-buechsenschuss-genai-organizational-networks adds a network-architecture mechanism. In a randomized field experiment, a grounded GenAI assistant increased collaboration and knowledge-sharing centrality, with specialists gaining more knowledge in-degree and generalists gaining more project output. Redesign therefore has to look beyond task allocation inside a role: AI may change who needs whom, who becomes easier to work with, and where knowledge requests concentrate in the organization.

2025-randazzo-cyborgs-centaurs-self-automators adds a workflow-mode layer to redesign. In the same nominal GenAI-enabled consulting task, professionals enacted different architectures: Fused/Cyborg users integrated GenAI across the workflow, Directed/Centaur users selectively preserved human control over analytical logic, and Abdicated/Self-Automators handed over much of the problem-solving and evaluative work. Redesign therefore has to specify not only which tasks AI supports, but which co-creation mode the organization wants to cultivate for accuracy, productivity, AI fluency, domain upskilling, or expertise preservation.

2026-johnston-shift-agentic-ai-codex gives early behavioral evidence for redesign around delegated agent work. The paper’s useful redesign signal is not simply that Codex adoption grew; it is that intensive users organize work through parallel agents, long-running turns, reusable skills/plugins, and larger task requests. That pattern makes agentic AI closer to workflow re-architecture than tool substitution: workers may shift from executing task components to assigning work, monitoring progress, reviewing outputs, maintaining reusable procedural context, and integrating artifacts across multiple simultaneous workstreams.

2026-hitzig-agentic-coding-expertise complements the Codex evidence by showing the internal work split in Claude Code sessions: humans typically decide what to do, while Claude decides how to execute it. The redesign implication is that agentic coding may weaken the boundary around who can produce technical artifacts, but it does not eliminate the need for domain expertise; work may shift toward problem framing, acceptance criteria, verification, and recovery when the agent runs into trouble.

2026-edwards-developer-productivity adds an interview-based developer view of that reallocation. Senior engineers described moving from implementation toward architecture, planning, and system integrity as AI handled more low-level work, while redefining productivity around durable impact rather than code volume. The source sharpens a redesign requirement: if implementation is delegated, roles, career progression, and performance criteria need to recognize architectural judgment, metacognitive specification, verification, and long-run product quality rather than simply ratcheting up output expectations.

2026-rodriguez-when-not-to-automate makes preservation itself a proposed redesign outcome. Its PHP-AIO protocol treats the choice among automate, augment, hybrid, and preserve as a role-level decision made before implementation, using risks to tacit knowledge, resilience, future regulatory constraint, and institutional trust. The implication is not that every consequential role should remain manually executed; it is that redesign should preserve enough human capability, redundancy, authority, and recovery capacity for the workflow to remain governable when automation fails or conditions change. The framework is conceptual and its financial-services parameters are not yet validated outside stylized cases.

2026-tomita-philosophy-cognitive-assay adds a design-with-AI distinction that capability mapping alone misses. It argues that AI can compress the externalization cost of turning expertise into formal output without inheriting the legitimacy to decide what the output should optimize, exclude, or count as good. For redesign, the practical question is therefore not only which steps are technically delegable, but where a workflow still needs a contestable human source for its evaluative criteria. The paper’s proposed Sein/Sollen assay is a conceptual way to make that distinction visible; its more durable contribution here is the design requirement to protect the socialization, practice, and deliberation through which those criteria are renewed when codification work accelerates.

2026-sherson-hybrid-intelligence-manifesto moves this from a task-level design question to a multi-level coordination claim. It frames redesign around a prediction-judgment frontier and argues that interfaces, organizational incentives, market valuation, and education have to align if human judgment is to remain economically central rather than becoming symbolic oversight. The useful design implication is not that every workflow should preserve human involvement; it is that an automate-and-elevate redesign must visibly reinvest capacity freed by routine automation in accountable, context-rich work and make that human contribution part of the system’s value logic. The manifesto is a public, consensus- and foresight-based preprint, not comparative evidence that this architecture outperforms automation-first alternatives.

2026-chin-work-at-frontier supplies an early behavioral signal that redesign pressure can cross occupational boundaries before an organization formally redraws roles. In a large sample of work-related ChatGPT messages linked to self-reported role information, 16.8% of all messages and 43.5% of occupation-specific non-generic messages concerned tasks historically associated with another occupation. That is not evidence that firms have successfully redesigned roles; it shows that AI can make workers attempt task bundles outside their historical occupational boundary, creating a governance and capability question about who is permitted, prepared, and accountable to do the newly reachable work.

2026-bakker-artificial-intelligence-job-design gives work redesign a JD-R architecture rather than a simple allocation vocabulary. Its central distinction is two-dimensional: who initiates adoption (bottom-up discretion versus top-down mandate or embedded system) is analytically separate from what the system does (augmentation versus automation). The review argues that the configuration, including whether top-down systems enable reciprocal interaction or impose unilateral algorithmic control, shapes whether redesign expands cognitive support and feedback or erodes autonomy, task completeness, learning opportunity, and social resources. It is an integrative theory review, not a causal test of the 2 × 2 framework.

The review also names job demand substitution as a design risk: AI can remove routine work but refill the saved time with harder escalations, review, accountability, or emotionally demanding work. Its cited examples include senior developers spending 6.5% more time reviewing code after Copilot adoption while junior developers coded faster, and sales agents shifting from repetitive tasks toward creative persuasion. Redesign should therefore measure the whole demand distribution across workers and roles, not call a workflow better because one visible task becomes faster.

Connections

• Connects to ai-adoption because adoption becomes a work-architecture problem, not just tool provisioning or training.
• Connects to ai-mediated-work-experience because redesign changes autonomy, skill use, feedback, task significance, workload, and identity threat.
• Connects to automation-and-substitution because the key design question is not simply which tasks AI replaces, but which workflow nodes require human judgment, care, trust, or visible accountability.
• Connects to human-ai-collaboration because control and verification are work-design choices, not merely interface features.
• Connects to automation-complacency because redesign can either preserve meaningful oversight or quietly convert human review into rubber-stamping.
• Connects to ai-induced-skill-erosion because redesigned workflows can either maintain deliberate practice and feedback or strip out the conditions that sustain expertise.
• Connects to ai-supported-knowledge-synthesis because knowledge-synthesis tools redesign how raw evidence, AI-generated summaries, stakeholder context, and final decisions move through the workflow.
• Connects to human-ai-task-taxonomy because redesign needs a task-classification language before tasks can be delegated, protected, audited, or recombined.
• Connects to novice-risk-work because novice users may misplace redesign responsibility at the individual or project level when system-level intervention is needed.
• Connects to competency-modeling because role-redesign claims often need a language for desired behavior, not only a decomposition of existing tasks.
• Connects to ai-agent-benchmark-validity because redesign decisions built from agent scores need evidence that the benchmark actually represents the targeted work.
• Connects to ai-enabled-job-crafting because worker-led redesign can reveal where GenAI is becoming a job resource, not only where leaders have formally reallocated tasks.
• Connects to ai-mediated-organizational-networks because redesigned work can emerge as changed collaboration and knowledge-flow patterns before formal role architecture catches up.
• Connects to human-ai-co-creation-modes because redesign decisions become enacted as recurring patterns of who steers, who executes, who evaluates, and who learns.
• Connects to agentic-delegation because agentic redesign turns delegation, review, concurrency, and reusable workflow infrastructure into core work-design variables rather than optional tool-use style.
• Connects to domain-expertise-in-agentic-work because redesign around agents changes which human capabilities matter: less direct execution in some cases, more problem framing, review, and contextual judgment.

Contradictions & open questions

• 2026-edwards-developer-productivity supports a shift from implementation toward architectural work, but its senior-user interviews cannot show whether organizations actually redesign roles, promotion systems, performance measures, or staffing around that shift—or whether it is a durable benefit rather than a rationalization for higher output demands.

• 2026-ermakov-designing-post-role-enterprise is a conceptual framework, not empirical evidence. It should not be treated as proof that fluid skill clusters outperform role-based structures.

• The source is strongly post-role; the wiki still needs counterevidence on where fixed roles, professional boundaries, or regulated accountability structures remain necessary.

• 2026-lee-making-invisible-visible complicates simple automation logic: workers were not opposed to AI taking work, but they cared which work. Some wanted AI to absorb tedious, repetitive, or lower-stakes work while preserving human control over expert, meaningful, accountable, or relationship-facing tasks.

• 2026-dellacqua-jagged-technological-frontier complicates redesign optimism: even skilled consultants with access to GPT-4 performed worse when the task sat outside the model’s capability frontier. Redesign needs guardrails for where not to use AI, not just patterns for where to insert it.

• 2025-brynjolfsson-generative-ai-at-work is evidence for augmentation inside a structured support workflow, not a general result for open-ended knowledge work. Its strongest design lesson may be that AI worked because suggestions were embedded, optional, and withheld when training data were thin.

• 2022-dellacqua-falling-asleep-at-wheel complicates the default “better tool plus human oversight” redesign story: making advice more reliable can also make humans less effortful. The design target is calibrated effort, not maximum automation or maximum friction.

• 2026-tomei-what-jobs-can-ai-learn may over-weight what can be represented as a learnable task environment. Work redesign still has to handle accountability, meaning, relationship work, and organizational context that are not captured by RL feasibility alone.

• 2025-mazeika-remote-labor-index is a benchmark of autonomous project completion, not an organizational redesign study. Its low automation rates argue for caution about full delegation, but they do not rule out large gains from human-guided decomposition, review, or tool-specific workflow changes.

• 2025-shukla-ai-assisted-design-ironies is practitioner-discourse evidence, not observed workflow performance. It is useful for surfacing likely redesign failure modes, but it does not show whether GenAI actually causes de-skilling or worse design outcomes over time.

• 2025-handa-economic-tasks-ai observes AI use, not organizational redesign. Conversation-level use can identify where work is being attempted with AI, but it cannot show whether task allocation, accountability, coordination, or job architecture changed around that use.

• 2025-marguerit-augmenting-automating-labor observes labor-market outcomes, not implemented redesign decisions. It supports the claim that automation and augmentation exposure have different consequences for new work, employment, and wages, but it cannot show which organizational design choices converted exposure into those outcomes.

• 2025-agarwal-what-work-ai-doing has the same observed-use limitation, plus measurement dependence on LLM-scored task characteristics. It is useful for identifying likely redesign pressure points, not for deciding which task clusters should be delegated without human review.

• 2026-reich-work-design-multidimensional-ai-threat is correlational, so it cannot establish whether good work design causes AI adoption or whether AI adopters are in roles they already experience as more autonomous and varied.

• 2025-klonek-ai-work-stress is cross-sectional social-media evidence, not an implemented redesign study. It sharpens the design variables organizations should measure — AI control, team-process resources, and stress — but it does not yet show which redesign interventions reduce stress in real workflows.

• 2025-shao-future-work-ai-agents is an audit of preferences and expert capability, not an implementation study. It helps identify where redesign friction is likely, but it does not prove that matching worker-desired agency levels produces better adoption, performance, or experience.

• 2025-yun-generative-ai-knowledge-work studies short-term prototype interaction in product-management tasks. Its redesign lessons are strong for interface and workflow requirements, but weak as evidence about durable organizational redesign or productivity gains.

• 2025-kellogg-novice-risk-work is interview-based and compares novice recommendations to expert literature, not implemented redesign interventions. It is strongest as a warning about misplaced mitigation levels, not as direct evidence that any specific deployer-level control improves outcomes.

• 2009-sanchez-levine-competency-modeling-job-analysis predates AI and should not be treated as evidence about agentic workflows. Its value is conceptual: it warns that redesign grounded only in current task description may miss the strategic and behavioral signals needed for role enactment.

• 2026-wang-agent-development-real-world-work is benchmark evidence, not implementation evidence. It can say where evaluations are misaligned with work, but not whether a redesigned workflow will produce better performance, motivation, learning, or accountability.

• 2025-wang-ai-agents-human-work reports that selective AI augmentation improved efficiency by 24.3% with limited workflow disruption while end-to-end automation slowed human work by 17.7% through verification and debugging. That contrast is useful but comes from 16 bounded computer-use tasks; it should not be generalized into a universal augmentation advantage or treated as field evidence of durable organizational redesign.

• 2026-vuong-genai-job-performance treats job crafting as a positive mediator, but informal AI-enabled crafting can also create unmanaged variance in process quality, compliance, and performance criteria if it remains invisible to the organization.

• 2026-buechsenschuss-genai-organizational-networks is strong field-experimental evidence that grounded GenAI can shift network centrality, but the self-reported network measures and single-firm design mean it should not yet be treated as proof that AI generally improves organizational coordination.

• 2025-dellacqua-super-mario-ai shows why redesign should not equate individual AI capability with system improvement. The automated agent was individually strong, but the redesigned human-AI team initially performed worse because coordination routines and motivation changed.

• 2025-randazzo-cyborgs-centaurs-self-automators is useful for redesign vocabulary but limited by its single-session consultant task and April 2023 GPT-4 context; organizations should not assume its mode/outcome pattern transfers unchanged to other jobs, tools, or capability frontiers.

• 2026-johnston-shift-agentic-ai-codex is strong process evidence for frontier agentic work practices, but it is not a field experiment showing that those practices improve productivity, quality, coordination, learning, or worker well-being. OpenAI-internal results are best read as a low-friction frontier case, not as the current baseline for ordinary organizations.

• 2026-hitzig-agentic-coding-expertise suggests coding occupation may matter less than task-specific expertise for agentic coding success, but it does not show how organizations should redesign roles, training, review rights, or accountability when non-coders can direct technical work.

• Preservation criterion vs. preservation instrument. 2026-rodriguez-when-not-to-automate usefully makes capability regeneration and organizational redundancy redesign criteria, but its five-gate protocol has not yet shown that scored preservation decisions avert later failure or outperform other participatory, qualitative, or regulatory governance approaches.

• Legitimacy boundary vs. usable design method. 2026-tomita-philosophy-cognitive-assay usefully separates technical delegation from the human warrant for value-laden criteria, but its scoring protocol remains a conceptual preprint: the wiki does not yet know whether teams can apply the distinction reliably, who legitimately represents contested values, or whether the assay improves redesign decisions rather than formalizing them after the fact.