How can AI improve your organization's workflow?
On the right tasks, AI can save time and raise quality. On the wrong ones it can make results worse. The gains in the studies below depended on the task and on who used the tool, so test AI one workflow at a time (Dell'Acqua et al., 2026) (Brynjolfsson et al., 2023).
What does the evidence show?
| Study | What it found | Limit |
|---|---|---|
| Professionals doing writing tasks (Noy & Zhang, 2023) | Among 453 college-educated professionals, those given ChatGPT took 40% less time on average and their output quality rose by 18%. | Short, incentivized writing tasks done online. |
| Customer support agents (Brynjolfsson et al., 2023) | Across 5,179 agents, an AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers, with minimal impact on experienced workers. | One company's support teams, in a working-paper version. Gains varied a lot between workers. |
| Consultants on realistic tasks (Dell'Acqua et al., 2026) | Among 758 consultants, on tasks inside AI's capabilities, those using AI completed 12.2% more tasks, 25.1% more quickly, with better quality. | One firm, one model, tasks built for the study. The frontier moves as tools change. |
Where does AI backfire?
The same consulting experiment included a complex managerial task outside the frontier of AI's capabilities. There, people using AI were 19% less likely to produce correct solutions than people without it (Dell'Acqua et al., 2026). Confidence matters too. In a survey of 319 knowledge workers, higher confidence in generative AI was associated with less critical thinking, while higher self-confidence was associated with more (Lee et al., 2025). That survey shows association, not cause.
The practical lesson is to check the output, especially on work you have not tested AI on before.
How do you try AI without a big bet?
This is our suggestion: small steps, measured, repeated. It is not a finding from the studies.
- Pick one repeated task that is slow or tedious.
- Write down its current steps, how long it takes, and how you judge quality.
- Try AI on one step only.
- Compare time and quality against your notes, and have a person check the output.
- Keep it, change it or drop it, then choose the next step.
What should stay with people?
Judgment calls, anything outside what you have tested, and anything involving confidential information. Read a tool's data terms before you put customer or company data into it.
How can we help?
We advise on organizational operations and on using AI deliberately inside them. Start with a free 30-minute call, or read about AI and wellness.
Sources
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science. Read the source
Brynjolfsson, E., Li, D., & Raymond, L. (2023). Generative AI at work (NBER Working Paper 31161, revised November 2023). Read the source
Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science. Read the source
Lee, H., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. C. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. CHI 2025. Read the source