Task Allocation

Bottom Line

AI is shifting tasks within jobs more than eliminating roles. Across causal studies, well-scoped subtasks are automated with human oversight; incentives and governance steer delegation; and job content and hiring are being redesigned, especially in knowledge work Mahinpei (2026); Jo (2026); Wang (2026). Main risks are reliability and coordination: agents struggle on long, multi-step, and negotiated work; users over-rely on outputs; and early deployments are narrow, so outcomes depend on authority, incentives, and oversight Li (2026); Yao (2026); Marusich (2026); Bonney.

New since the cutoff: field and lab experiments show incentives and human-in-the-loop controls steer delegation and quality, and new agent benchmarks document persistent coordination failures, reinforcing a centaur approach (humans plus AI with explicit governance) Wang (2026); Zhu (2026); Li (2026).

What This Means in Practice

• For customer support, set early human takeovers on emotional escalations and add checkpoints. This avoids rating drops and moves staff to AI-ineligible cases Wang (2026).
• If you need original work, pay for it. Tying pay to novelty cuts copy-paste from AI and encourages selective use. In tools, show model uncertainty and log step-by-step actions so people can catch errors and share control Jo (2026); Sbeyti (2026); Sabouri (2026).
• Route work by measured capability and risk, not model brand. Use top models for long-horizon planning and constraint tracking; smaller models for short, structured steps. Do not allocate tasks by models’ self-reported cost or success, or run auctions on those self-reports Karmakar (2026); Fradkin (2026); Bai (2026).
• Stage rollouts and aim for augmentation first, starting with writing and search. In highly digital firms, decentralize some decisions to capture local complements Bonney; Chen.
• Prevent skew and over-reliance. Add an "unknown" demographic bucket in ad budgets to avoid under-delivery, and avoid narrative explanations that raise reliance whether right or wrong Corpus (2026); Marusich (2026).

What the Research Finds

AI reallocates tasks within jobs and across roles, more than it eliminates jobs

• Hiring changes explain 52% of the decline in generative AI exposure in job ads; task redesign explains 39.5%. Senior roles adjusted earlier via hiring; junior roles changed both hiring and tasks Wang (2026).
• Across 35 European countries, a shift-share study (using regional industry mix to proxy exposure) finds no change in worker-reported tech-related task restructuring, even as adoption concentrates in high-exposure jobs Henseke (2026).
• A natural experiment on adoption is associated with employment moving away from routine cognitive tasks toward complex problem-solving and interpersonal work, with wage gains in the top quintile and declines in the middle quintile A. T. D..
• Inside U.S. firms, use is narrow: most apply AI to writing, document analysis, and search; 66% report augmentation over replacement, and 57% use AI in three or fewer functions Bonney.
• More digital firms decentralize decision authority to subsidiaries, especially when diversified and in uncertain environments Chen.

Incentives, governance, and human-in-the-loop design steer delegation and quality

• Paying for originality reduced verbatim copying and pushed selective co-creation (brainstorming, proofreading, targeted edits) vs paying for quality alone Jo (2026).
• In an RCT with teaching assistants, editable AI drafts raised the share leaving feedback by 10.8 percentage points without adding time per character; drafts worked as scaffolds, not substitutes Mahinpei (2026).
• In customer service, supervising an agentic AI shortened handling time and shifted worker effort to AI-ineligible chats, but ratings fell on AI-eligible emotional escalations unless humans intervened early and invested effort Wang (2026).
• A human-in-the-loop research workflow that blocked large language models (LLMs) from core data steps and added human gates cut AI-assisted research failures from 72% to 16% in tests Zhu (2026).
• Showing model uncertainty reallocated human effort: visualizing localization uncertainty improved annotation quality and reduced time; action-auditable spreadsheet agents improved error detection and shared ownership Sbeyti (2026); Sabouri (2026).
• Adding an "unknown" user group in ad budget splits reduced demographic under-delivery without exclusions Corpus (2026).

Agentic AI expands scope but remains coordination-limited

• Across live, evolving, and industrial benchmarks, leading agents complete about two-thirds of tasks, with systematic failures in long-horizon execution, proactive clarification, and tool orchestration (about 23% incorrect sequencing) Li (2026); Das (2026); Wang (2026); Li (2026).
• In multi-turn negotiations, paired agents often miss Pareto-optimal deals (improvements for at least one side without hurting the other) even when each has the needed knowledge alone, due to grounding failures such as anchoring, shallow fairness appeals, and dropped commitments Yao (2026).
• Agents’ self-reported success and cost are miscalibrated; markets that allocate tasks using those self-reports diverge from full-information benchmarks and only modestly improve with context cues Fradkin (2026).
• Mixed model stacks perform well: smaller open models handle short, structured steps cost-effectively, while frontier models keep an edge on long-horizon planning and constraint tracking Karmakar (2026).
• Production logs show autonomous agents that decompose and execute tasks change the scope and depth of work attempted, shifting follow-ups toward verification and extension Yang (2026).

Distributional and sectoral heterogeneity matter for who does what

• In Indonesia, women exposed to routine-task displacement moved into nonroutine interpersonal roles, temporarily narrowing the gender wage gap before later valuation shifts reversed gains Jamil (2026).
• In the U.S., WIOA retraining rarely moved participants into less automation-exposed jobs; many returned to prior fields. Apprenticeships performed best Jacobs (2026).
• A reinforcement learning (trial-and-error) "learnability" index for O*NET tasks flags high-risk jobs that exposure scores miss (for example, power plant operators), implying different future substitution patterns Tomei (2026).
• At home, a natural experiment suggests ChatGPT adoption raised leisure browsing’s share by about 30 percentage points while leaving productive online time unchanged; use clustered around productive contexts, with declines concentrated in search and news Blank.

Inside organizations, AI shifts role boundaries and control, not just tasks

• Developers prefer bounded delegation: AI for peripheral assembly work with provenance, uncertainty signals, scoped authority, and least privilege. Developers already spend roughly a tenth of their day writing code Choudhuri (2026).
• In a large tech firm, AI blurred role boundaries and increased peer collaboration, but disrupted informal mentoring and feedback channels needed for growth Rosenthal (2026).
• Evidence shows both over- and under-reliance in daily work, pointing to the need for calibrated supervision Ozdemir (2026); Ferino (2026).
• Heavy AI use can create cognitive overload. Narrative explanations increase reliance whether right or wrong, raising supervision risks in high-stakes work Westover; Marusich (2026).

What We Still Don't Know

• How task allocation settles over years. We lack comparative, longitudinal RCTs that test human-only vs centaur vs agentic automation across full workflows. Early national evidence shows mixed or null short-run task-change signals Henseke (2026); Luengo Vera (2026).
• Which governance patterns generalize across domains. Current positive results for early human intervention and auditable actions come from customer service, spreadsheets, and labeling, not heavily regulated or safety-critical settings Wang (2026); Sabouri (2026); Sbeyti (2026).
• The net distributional impact of AI-driven task reallocation by gender, race, age, and region as adoption scales. Most evidence is sector- or country-specific and often observational Jamil (2026); A. T. D.; Bonney.
• How agent teams can reliably coordinate in open-ended, negotiated, or multi-actor settings. Failures in grounding, calibration, and orchestration are documented; scalable fixes beyond narrow protocols are not Yao (2026); Fradkin (2026); Li (2026).
• Whether public workforce programs can be redesigned to move people into less automatable task bundles at scale. Current U.S. evidence suggests limited success with business-as-usual retraining Jacobs (2026).