The Human-AI Partnership: How Artificial Intelligence Changes Learning, Decision-Making, and Workplace Performance
A Protergates research review of how generative AI changes knowledge work: where it speeds things up, where it fails, and why skill retention, calibration and unaided performance matter as much as output.
In short: generative AI speeds up work on tasks that match its strengths, but the benefits are uneven and it can make some tasks worse. Human oversight works best when reviewers know the subject, have time to verify and can see the evidence. Judge AI adoption not just by output today, but by whether people still learn and can work unaided.
Research review. This is case study 04 of the Protergates research portfolio Technology, Society and Human Capability (November 2026). It synthesises published research and public regulatory guidance; it does not report new data.
Abstract
This review examines empirical findings on generative AI assistance in knowledge work, contrasting immediate speed with long term skill retention, error calibration and the ability to transfer skills to tasks done without AI.
Complementarity, overreliance and learning transfer
AI assistance improves throughput where model strengths match task requirements, but the gains are not uniform. A Harvard Business School field experiment with 758 consultants described a jagged frontier: AI improved speed and quality on tasks inside its capabilities, while performance fell on a task just outside them. Studies in Science and the Quarterly Journal of Economics found productivity gains in writing tasks and customer support.
Human oversight is most effective when reviewers have domain knowledge, time to verify claims and access to the underlying evidence. Uncritical acceptance can spread errors and weaken independent judgement.
| Measure | Why it matters |
|---|---|
| Output quality | Is the work correct right now? |
| Speed | How much time is saved? |
| Calibration | Does confidence match actual accuracy? |
| Retention | Does the person remember and understand the work later? |
| Unaided transfer | Can they do similar work without AI? |
Sustainable adoption looks at all five, not just the first two.
Practical recommendations
- Train people to verify: check claims, explain reasoning and recognise model uncertainty.
- Assess independent mastery as well as AI assisted output quality.
- Keep humans in the loop for high stakes work, and evaluate models for each specific task.
- Keep an audit trail of tool versions, prompts and evaluation criteria.
What it means for you
Treat AI output as a draft to check, not an answer to accept. Ask for sources and verify them, and keep doing important tasks yourself sometimes so the skill stays yours. Our tips on what your AI chatbot remembers and whether your chats train the AI cover the privacy side of everyday AI use.
Portfolio synthesis
Across all four reviews in this portfolio, technology delivers its strongest benefits when capability, autonomy, oversight and accessibility are designed together. Digital literacy enables informed adoption, quality of life assessment tests whether benefits reach people, compliance provides institutional safeguards, and human and AI evaluation checks whether assistance strengthens judgement rather than replacing it.
Frequently asked questions
Does generative AI make people more productive?
On tasks that match its strengths, published experiments found faster and better work. On tasks just outside its capabilities, performance can fall, so the benefits are uneven.
When does human oversight of AI work best?
When reviewers know the subject, have time to verify claims and can see the underlying evidence. Uncritical acceptance spreads errors.
How should organisations measure AI adoption?
By output quality and speed, and also by calibration, skill retention and whether people can still do the work without AI.
Is this based on new data?
No. It is a research review of published experiments in Science, the Quarterly Journal of Economics and a Harvard Business School field study.
Sources
- Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381, 187 to 192. doi:10.1126/science.adh2586
- Brynjolfsson, E., Li, D., and Raymond, L. R. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889 to 942
- Dell'Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI. HBS Working Paper 24-013