A Survey Series on AI for Value
AI4
where X ∈ { Productivity, Science, AI }
From model capability to real-world value. A unified inquiry into how AI becomes practical, trustworthy, and accountable when it enters the value chains of work, science, and intelligence itself.
Meaning is defined by use.
For a long time, AI progress was measured by closed benchmarks and leaderboards. Yet its most meaningful leaps arrive when it enters the real world, becomes embedded in everyday processes, and delivers sustained, measurable value.
The central question is no longer only “Can the model answer accurately?”, but “Can it execute reliably?” — can it turn parametric knowledge into trustworthy actions within complex human systems, and drive tangible value? Once AI joins a value chain, it is judged by throughput, quality, reliability, risk, accountability, and return on investment. The frontier is moving from model capability to real-world system performance.
Three settings, one inquiry.
We develop a single inquiry into AI for value across three closely related settings — each a higher order of transformation than the last.
-
I First-Order · Efficiency
AI for Productivity
How AI is integrated into everyday work to generate economic and organizational value — accelerating routine tasks, coordinating multi-step workflows, supporting decisions, and enabling work that was previously too costly, slow, or complex. The central concern is utility.
-
II Second-Order · Discovery
AI for Science
Moving from labor to discovery — how AI supports the scientific cycle: hypothesis generation, experiment design, simulation, data analysis, interpretation, and literature synthesis. Here value is epistemic: knowledge that is rigorous, traceable, reproducible, and verifiable.
Coming soon -
III Third-Order · Intelligence
AI for AI
Turning inward to the engine of AI progress itself — agent-driven data generation, automated evaluation, synthetic supervision, model refinement, and iterative self-improvement. Value lies in bootstrapping capability under data, compute, and evaluation constraints, in a way that is controlled and sustainable.
Coming soon
Part I — What we found.
Key insights from AI for Productivity in the Agentic Era. Parts II and III of the series are forthcoming.
AI for Productivity is the use of Agentic AI in real-world workflows to deliver economically valuable work with less human effort.
Efficiency
Reduces the time and labor required for tasks that are already being done.
Expansion
Makes previously infeasible work economically viable — new tasks, not just faster ones.
- L1Conversational Assistant
- L2Reactive Operator
- L3Adaptive Coordinator
- L4Self-Governing Production System
Manufacturing · Agriculture · Real Estate · Government
Trade · Media · Education · Legal Services
Information Technology · Finance · Healthcare
- Safety
- Long-horizon reliability
- Academia–industry gap
- Adoption barriers
- Governance
- Labor-market uncertainty
Citation
@article{xi2026ai4productivity,
title = {{AI for Productivity in the Age of Agentic AI}},
author = {Xi, Zhiheng and Zhou, Enyu and Fang, Xinyu and Sun, Zhe and Huang, Baodai and Sun, Jiajun and Deng, Bicheng and Zhang, Zhihao and Chen, Wenxiang and Zhang, Jiazheng and Guo, Xin and Liu, Shichun and Lei, Zhikai and Wang, Junke and Jin, Senjie and Nan, Yang and Yang, Yajie and Zheng, Rui and Yan, Hang and Yang, Mengyue and He, Yinghui and Tian, Yuchen and Qian, Cheng and Zhao, Xuandong and Yin, Zhenfei and Yang, Ling and Zhuge, Mingchen and Wu, Fang and Ying, Kejun and Wu, Yingcheng and Wang, Jun and Torr, Philip and Wu, Zuxuan and Qiu, Xipeng and Zhang, Qi and Huang, Xuanjing and Gui, Tao and Jiang, Yu-Gang},
journal = {Preprints},
year = {2026},
doi = {10.20944/preprints202608.1248.v1},
url = {https://doi.org/10.20944/preprints202608.1248.v1}
}