Three years of field experiments on AI at work have produced a small number of findings that keep coming back. They recur in consulting, call centres, professional writing, software and the intensive care unit.
Read one at a time, each looks like a productivity statistic. Read together, they describe how work is being rebuilt underneath the people doing it. Most companies are still asking whether AI makes people faster. The studies answer a more useful question. Which tasks gain, which people gain, and under what design.
This report sets out seven patterns. Each is stated as the evidence shows it. A separate reading follows, marked as such. The two are kept apart on purpose.
On method
This is a synthesis of roughly fifteen primary empirical studies published between 2022 and mid-2025 that could be traced to source, with foundational work from 2020 and 2021 and Bainbridge’s 1983 paper on automation. Figures are reported as the studies report them. Anything interpretive sits under the label Reading.
The first report in this series, What Is Actually Happening to Work, looked at the labour market. The second, What Companies Will Actually Do, reasoned forward to company decisions. This one goes inside the task.
Part I · The seven patterns
What the studies keep finding.
Pattern 1 · The jagged frontier
Tasks that look equally hard fall on opposite sides of the line.
Dell’Acqua and colleagues ran a field experiment with 758 Boston Consulting Group consultants. On creative product work inside AI’s capability, the AI group completed 12.2 percent more tasks, 25.1 percent faster, at more than 40 percent higher quality. On a strategy task that required combining qualitative and quantitative evidence, the same people with the same tool were 19 percentage points less likely to reach the correct answer.
Difficulty has more than one dimension. A task can be easy for a model on retrieval and out of reach on contextual judgment. A task that looks identical can reverse the profile.
Reading
Occupation-level measures of AI exposure average across the jagged edge, so they mislead. The unit of analysis is the task. The most valuable working knowledge right now is where the line runs in your own work. The model will not tell you. It sounds equally sure on both sides.
Pattern 2 · Skill compression
The least experienced gain the most.
In a study of 5,179 customer support agents, Brynjolfsson, Li and Raymond found that an AI assistant raised issues resolved per hour by 14 percent on average and by 34 percent for novice and low-skilled agents. The effect on the most experienced and skilled was minimal. In the BCG experiment, consultants below the average performance threshold improved 43 percent against their own baseline. Those above it improved 17 percent. Noy and Zhang found the same narrowing in professional writing.
The mechanism is tacit knowledge in transit. The models learned from the output of strong performers. When the support chats were analysed, newer agents began to write like veterans after exposure to the tool.
Reading
This runs against the skill-biased pattern of most earlier technology. It is good news for people early in their careers and a direct challenge to anyone whose edge was accumulated experience alone. It leaves a question the studies cannot yet answer. Do people who start with the shortcut ever build the judgment the shortcut was copied from?
Pattern 3 · Conditional complementarity
Human plus machine usually loses to the better of the two.
Vaccaro, Almaatouq and Malone’s meta-analysis covered 106 experiments and 370 effect sizes. On average, human-AI combinations performed worse than the human alone or the AI alone, whichever was stronger. They usually did beat a human working without AI.
Two conditions made the combination more likely to win. The human alone outperformed the AI alone. The task involved creating content rather than making a binary decision. Hemmer and colleagues add the structural requirement. The person must hold information the model lacks, or the two must make different kinds of error.
Reading
The useful design question is narrow. Is this task generation or judgment? Is the person in the loop actually better than the model at the part they are asked to check? Most studies in the meta-analysis predate GPT-4. Trust the direction of the finding more than its size.
Pattern 4 · Effort relocated
The thinking moved to checking.
Lee and colleagues at Microsoft Research surveyed 319 knowledge workers about 936 real uses of generative AI. The share of examples reported as needing less effort was 72 percent for knowledge, 79 percent for comprehension, 69 percent for application, 72 percent for analysis, 76 percent for synthesis and 55 percent for evaluation.
The effort that remained changed character. The researchers describe it shifting from gathering information toward verifying it, from solving problems toward integrating machine answers, and from executing tasks toward overseeing them. Bainbridge named the trap in 1983. Automation raises the need for skilled oversight and removes the practice that builds it.
Reading
That verification is now a larger share of the remaining effort is an inference from the pattern. The study reports the raw percentages. Drafting, calculating and analysing are the activities through which people used to become expert. The oversight role replacing them is the one that most needs expertise.
Pattern 5 · Trust calibration
Outcomes turn on how much people trust the tool.
In the same Microsoft study, higher confidence in AI was associated with less critical thinking (β = −0.69, p < 0.001). Higher confidence in one’s own ability was associated with more (β = 0.26, p = 0.026).
Calibration is hard from both directions. Models often state wrong answers with high confidence, so their confidence is a poor cue. People also arrive biased. Dietvorst, Simmons and Massey showed that people abandon an algorithm after seeing it err faster than they abandon a human forecaster who makes the same error.
Reading
The people who got most from the tool trusted their own judgment enough to keep using it. Trust is set task by task and drifts with familiarity. It needs managing the way any other operating risk is managed.
Pattern 6 · The drift
Augmentation slides toward automation without a decision.
Raisch and Krakowski describe an automation-augmentation paradox. Over time the two cannot be held apart. A task automated today creates new augmented work tomorrow, and a tool built to assist drifts toward deciding as people grow used to delegating.
Cui, Tan and Shi documented this at a logistics company. Algorithmic recommendations first supported managers. As they became institutionalised, they were followed by default and challenged less.
Reading
Nobody chose this. Comfort compounds. Too much augmentation leaves inefficient work in place. Too much automation erodes skill and blurs accountability. The balance needs resetting continuously, which means someone has to own it.
Pattern 7 · Agency by design
Human judgment survives only where it is designed in.
When outputs are AI-generated and human-certified, accountability diffuses. Tsamados, Floridi and Taddeo observe that most deployed systems sit at the supervision end of the spectrum, whatever the language of teaming says.
Bienefeld, Keller and Grote compared preferences across six core ICU tasks, using task analysis, a Delphi survey of 19 data science experts and interviews with 61 clinicians. Both groups favoured augmentation for most tasks, on condition of transparency, predictability and clinician control over outcomes. Both agreed responsibility must stay with the clinician. Schmitt sets out the design side: interfaces that keep the user in control, feedback that builds skill, and decision points where the person’s judgment changes the result.
Reading
Each of the first six patterns wears agency down a little unless something pushes back. That something is design. It does not arrive by default.
Part II · How the patterns connect
One system underneath.
The seven patterns act on the same unit: one person, one system, one task. The underlying paper organises them into a diagnostic frame, the Complementarity-Agency-Stewardship model. It is offered as one lens on the evidence. It has not been independently tested.
Complementarity · Patterns 1 to 3
Does the task let the combination win?
It needs an asymmetry between person and model, a task that lets the asymmetry be used, and an interface that lets the person fold the output into their own judgment.
Stewardship · Patterns 4 and 5
Can the person still check the work?
As the role moves from producing to overseeing, the practice that builds oversight thins. Better models need better stewards, and better models reduce the practice that produces them.
Agency · Patterns 6 and 7
Does the person still decide?
Real decision points, feedback that develops skill, and a clear owner for each outcome. Without them, drift and diffusion do the rest.
The frame also predicts a tension it cannot dissolve. Designs that weave suggestions tightly into the workflow raise complementarity and tend to lower agency. Designs that demand approval at every step protect agency and add friction that lowers complementarity. This trade-off is a prediction. Nobody has tested it yet.
Part III · Reading, derived from the patterns
If you design the work.
First. Diagnose by task.
Before buying a tool, map the specific tasks against the frontier. An occupation-level exposure score will not show where your people get 40 percent better and where they get 19 points worse.
Second. Count verification as work.
Checking is now a large part of the job. Budget time for it, train for it and measure it. A team that is faster and checks less has moved risk somewhere it will surface later.
Third. Protect the practice that builds judgment.
Juniors who never draft will not become the seniors who can review drafts. Keep some production work deliberately human, for development rather than output.
Fourth. Name an owner for the line.
Someone has to decide, and keep deciding, what the system does and what people must keep doing. If that role does not exist, drift is setting the line.
What this does not show
Most of the evidence comes from text and image systems, observed over months. The meta-analysis largely predates GPT-4. The effort figures are self-reported. Evidence from manufacturing is thin, mostly reviews and small pilots.
Credible readings run against this one. The frontier may smooth as models improve. Skill compression may prove a passing onboarding effect. Stewardship may turn out to be an occupational upgrade. Each is falsifiable. Longitudinal data will settle them, and it has not yet.
The tools will keep improving. The patterns describe the human side of the system, and that side does not improve by itself. The organisations that come through well will decide, task by task, where the line runs and who stays able to check it.
The evidence this rests on
- Dell’Acqua, F., et al. (2023). Navigating the jagged technological frontier. Harvard Business School Working Paper. SSRN 4573321
- Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2). NBER w31161
- Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654). DOI
- Vaccaro, M., Almaatouq, A., and Malone, T. (2024). When combinations of humans and AI are useful. Nature Human Behaviour, 8. Article
- Hemmer, P., et al. (2025). Complementarity in human-AI collaboration: Concept, sources, and evidence. European Journal of Information Systems, 34(6).
- Lee, H.-P., et al. (2025). The impact of generative AI on critical thinking. Proceedings of CHI 2025. DOI
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6). DOI
- Dietvorst, B., Simmons, J., and Massey, C. (2015). Algorithm aversion. Journal of Experimental Psychology: General, 144(1). DOI
- Raisch, S., and Krakowski, S. (2021). Artificial intelligence and management: The automation-augmentation paradox. Academy of Management Review, 46(1). DOI
- Cui, T., Tan, B., and Shi, Y. (2024). Fostering humanistic algorithmic management. Journal of Strategic Information Systems, 33. DOI
- Tsamados, A., Floridi, L., and Taddeo, M. (2025). Human control of AI systems: From supervision to teaming. AI and Ethics, 5(2). PubMed
- Bienefeld, N., Keller, E., and Grote, G. (2024). Human-AI teaming in critical care. Journal of Medical Internet Research, 26, e50130. Article
- Schmitt, A. (2024). Ensuring human agency: A design pathway to human-AI interaction. ICIS 2024 Proceedings.
- Gonzalez, C., and Heidari, H. (2025). A cognitive approach to human-AI complementarity in dynamic decision-making. Nature Reviews Psychology, 4(12).