Source-linked AI summary
When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration
AI Collaboration, Jiaying Wu, Caleb Ziems, Raymond Chan, Nancy F. Chen, Corlyss Chua, Gerard Chung, Jungpil Hahn, Wee Sun Lee, Zhengyuan Liu, Jamie Ng, Desmond C. Ong, Jeryl Ong, Da Ren Soon, Tianqi Song, Zhi-Xuan Tan, Sixing Tao, Emily Yang, Yajing Yang, Stella Xin Yin, Min-Yen Kan, Diyi Yang
TL;DR
Workplace AI measures often track adoption and aggregate outcomes rather than how work changes across entire workflows. This paper proposes a six-condition framework for assessing augmentation and applies it to AI-mediated social surveys and Singapore labour-market evidence. It concludes that augmentation requires durable value, meaningful human control, accountability and recovery, and sustained pathways for learning, progression, and purpose.
Problem
Workplace AI measures often track adoption and aggregate outcomes, leaving changes in workflows and the nature of work less directly assessed.
Source: §3 · p. 2Method
Results
The paper concludes that genuine augmentation combines durable net value, meaningful human control, accountability and recovery, and sustained pathways for learning, progression, and purpose.
Source: §8 · p. 7Takeaways & Limitations
A workflow-level framework gives organisations a basis for deployment decisions and gives researchers, educators, and public authorities a common object of analysis.
Source: §8 · p. 7A common workflow record can support comparison, but sector-specific thresholds remain necessary because workflows differ in quality, privacy, contestability, and risk requirements.
Source: §7 · p. 7
Abstract
from arXiv · showhide
We aim to characterise the value of artificial intelligence in the workplace. Current studies largely measure this value in terms of the current automation capabilities and public adoption of AI. However, such metrics ignore the greater impacts of human--agent collaboration in transforming the nature of work. To account for this, we must expand the scope of our analysis beyond atomised tasks of today, and instead focus on how AI can augment entire workflows of the future. To ground this analysis, we establish a precise definition of AI augmentation comprising six conditions, spanning durable net value, meaningful human control, accountability and recovery, and long-term human development through learning, career pathways, and job purpose. We elaborate on these conditions and apply the framework in a case study of AI-mediated social surveys. We conclude by outlining how organisations, researchers, and government leaders can use this framework to make sense of the future of work.
1 Introduction
The whitepaper draws on CIVIC-AI 2026 workshop discussions about redesigning work with AI agents and the governance needed for successful augmentation.
1 Introduction
- The whitepaper grew from discussions and findings at the CIVIC-AI 2026 workshop.Participants from academia, industry, and regulatory bodies discussed redesigning work with AI agents.Source: §1 · p. 2
- The discussions focused on how organisational and regulatory governance can support AI agents in augmenting future work.Source: §1 · p. 2
- The paper discusses recommendations for structuring human–AI interaction at work and conditions for successful AI augmentation.Source: §1 · p. 2
2 Governance Frameworks for Human–AI Collaboration
Existing frameworks emphasise human-centred governance but offer less guidance for evaluating how AI reshapes workflows and the human capacity to oversee them.
2 Governance Frameworks for Human–AI Collaboration
- Existing AI governance frameworks emphasise human-centred AI, oversight, accountability, and safeguards.Examples include OECD, WEF, U.S. Department of Labor, and Singapore guidance.Source: §2 · p. 2
- They provide less guidance on evaluating workflows as a unit, including changing task allocation, decision rights, verification work, recovery burdens, and human capability.Source: §2 · p. 2
- Disclosure, audit, and sign-off mechanisms can assume reviewers remain capable, while leaving open whether workflow redesign erodes that capacity.Source: §2 · p. 2
- The paper proposes a workflow-level unit of analysis and conditions distinguishing genuine from performative AI augmentation.Source: §2 · p. 2
3 A Workflow-Level Analysis of AI-Augmented Work
The paper argues that AI's effects should be assessed across whole workflows, where task allocation, human intervention, hidden work, and capability change together.
3 A Workflow-Level Analysis of AI-Augmented Work
- Workflow analysis captures changes in task allocation, hand-offs, and decision rights that adoption, investment, productivity, and headcount measures may miss.Aggregate outcomes can mask team reorganisations, shifting responsibilities, and how employees spend their time.Source: §3 · p. 2
- A workflow record should track what agents generate or execute, where humans review or intervene, and how unfamiliar cases and failures are handled.Source: §3 · p. 2
- Net value must include verification, exception-handling, and recovery effort, while longitudinal assessment should track whether workers retain the competence needed for meaningful oversight.The paper says organisations must provide substantive review work and exposure to AI failure to keep verification skills current.Source: §3 · p. 3
4 Six Conditions for Genuine Augmentation
The framework defines genuine augmentation through six conditions assessed at deployment and over time, with task boundaries and human capability central to meaningful control.
4 Six Conditions for Genuine Augmentation
- The six conditions have two assessment layers: deployment and periodic review, then longitudinal evidence about skills, progression, agency, and meaningful control.A workflow may meet the first layer at deployment but cease to augment work if it erodes the second over time.Source: §4 · p. 4
- Human–agent task boundaries depend on verifiability, reversibility, and the stakes of errors or delays.Delegation can increase when outputs are inspectable, errors recoverable, and stakes bounded; humans should retain stronger authority over goals and ambiguous or consequential judgments.Source: §4 · p. 4
- The conditions are interdependent: verification can consume saved time, approval is weak without failure-detection ability, and accountability needs decision-relevant information.Source: §4 · p. 4
- Delegating formative tasks can weaken the tacit judgment junior workers need for later oversight, so task allocation must preserve learning and career pathways.Meaningful human control depends on maintaining those learning opportunities over time.Source: §4 · p. 4
5 A Worked Example: AI-Mediated Social Surveys
The social-survey example assigns strategic design and consequential decisions to human researchers while AI handles bounded, reviewable execution. It proposes measuring both survey outcomes and researchers’ development, with a design rule for allocating work.
5 A Worked Example: AI-Mediated Social Surveys
- Human researchers set survey purpose, seed questions, weighting, interpretation, and final coding; an AI interviewer such as SparkMe [1] can revise wording, ask follow-ups, collect data, and initially review responses.Table 2 describes the agent’s execution as bounded and reviewable, with humans retaining design authority, oversight, and responsibility for consequential decisions.Source: §5 · p. 4, Table 2 · p. 5
- Researchers should design protocols and perform initial coding to counter anchoring and vacuous verification while giving junior researchers practice that builds mastery for strategic decisions.That mastery supports decisions about study meaning and validity, operationalised concepts, coding conflicts, and interpretation.Source: §5 · p. 5
- An AI interviewer becomes part of the measurement instrument: adaptive questioning may reduce burden or social-desirability pressure, but can add hidden assumptions through reformulation, follow-ups, coding, or aggregation.The passage calls for measuring and assessing these effects as first-class variables.Source: §5 · p. 5
- A pilot could compare fixed-form, human-interviewed, and adaptive-AI surveys using response, respondent-experience, researcher-time, validity, consistency, and coding measures.Suggested measures include completion and response rates, burden and response depth, comfort and trust, review time, construct validity, and coding accuracy and disagreement.Source: §5 · p. 5
- Researchers should retain interview paths to compare agent-presented questions with the IRB-approved instrument, then inspect outputs, revise the instrument, challenge coding, and explain conclusions.Source: §5 · p. 5
- Longitudinal evaluation should track researchers’ coding mastery, junior and senior talent development, job satisfaction, and purpose.Source: §5 · p. 5
- Assign agents work where gains are scalable, actions reversible, and errors inspectable; retain human authority over goals, validity, interpretation, and consequential use, while measuring effects on human efficacy over time.Source: §5 · pp. 5–6
6 Operationalising the Augmentation Framework: Singapore as an Application Context
Singapore’s existing labour-market evidence offers a baseline for assessing augmentation, but it does not yet establish whether workflows meet the framework’s conditions. The evidence is especially limited on longer-term worker development and unequal ability to exercise control.
6 Operationalising the Augmentation Framework: Singapore as an Application Context
- 28.5% of firms reported adopting AI, and 70.7% of adopters reported worker-productivity improvements; role redesign was reported more often than headcount reduction, 18.9% versus 6.2%.These statistics offer preliminary evidence on durable net value but omit verification, exception handling, rework, recovery, and unofficial AI use.Source: §6 · p. 6
- Entry-level PMET openings rose from 32,500 in December 2025 to 32,800 in March 2026, while fresh-graduate employment remained broadly resilient.These snapshots neither indicate broad-based deterioration in entry-level opportunities nor establish AI’s longer-term effects on learning, career pathways, agency, or job purpose.Source: §6 · p. 6
- Among young workers, 38% of those with secondary qualifications used new technology at work, versus 74% of degree holders.Human control should be assessed as a capability workers can exercise, not merely a formal role; existing measures leave open whether particular workflow redesigns meet the conditions.Source: §6 · p. 6
7 Moving Forward: Workflow Records
The paper proposes shared workflow records, complemented by employee-level and other evidence, to make augmentation claims auditable and comparable without assuming one operating model fits every sector. It recommends incorporating workflow modules into pilots and labour-market research.
7 Moving Forward: Workflow Records
- Each workflow record should document the baseline and objective, agent tasks and authority, human review and escalation, exception handling and recovery, significant errors, total effort, and effects on human capability and progression.The record also includes unintended agent actions, verification and repair, AI literacy, learning opportunities, agency, and satisfaction.Source: §7 · p. 6
- Complement workflow records with employee-level evidence and triangulate employer and employee surveys, administrative records, job postings, expert input, industry interviews, and qualitative evidence.These sources capture different aspects of adoption and can reveal effects that aggregate measures obscure.Source: §7 · pp. 6–7
- A common workflow record can support comparisons without imposing a uniform operating model, while sector-specific thresholds reflect differences in quality, privacy, contestability, and risk requirements.Source: §7 · p. 7
- Organisations could use workflow modules to expand, revise, or discontinue deployments; researchers could compare workload, capability, and career-stage effects over time; public authorities could link measures with workforce and training data.Source: §7 · p. 7
- Workflow-level measures complement existing sources by showing how adoption and workforce outcomes arise from work reorganisation, making gains, oversight, recovery, and learning effects testable.Source: §7 · p. 7
8 Conclusion
The conclusion presents workflow design as the locus of AI-driven change and offers a framework for assessing when that change constitutes genuine augmentation. Its central criterion combines net value, human control, accountability and recovery, and sustained human development.
8 Conclusion
- Whether AI-driven workflow change constitutes augmentation depends on how human and AI capabilities are combined, including task and decision-authority allocation and error handling.Source: §8 · p. 7
- The workflow-level framework gives organisations a basis for deployment decisions and gives researchers, educators, and public authorities a common object of analysis.Source: §8 · p. 7
- AI genuinely augments work when a redesigned workflow creates durable net value, preserves meaningful human control, provides accountability and recovery, and sustains learning, progression, and purpose.Source: §8 · p. 7