Customer behavior and adoption · Preregistered randomized controlled laboratory-in-the-field experiment, with realistic consulting tasks and a no-AI, GPT-4, and GPT-4-plus-prompt-engineering condition.
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
Dell'Acqua, F., III, E. M., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., & Candelon, F. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science.
Plain-English take
In this study, Within the AI capability frontier, AI users completed 12.2% more tasks, 25.1% faster, with higher quality; on an out-of-frontier managerial task they were 19% less likely to be correct. Founder use: Design AI product analytics around task-level success and verification, not aggregate usage; expose where the model helps and require review where it can degrade decisions.
How much weight to give this
4.5/5 Strong experimental or quasi-experimental evidence for its stated setting
The 1–5 evidence-check score is a practical reading aid, not a formal peer review or a probability that a conclusion is true. It combines design strength (2 points), sample reporting (1), comparison/control clarity (1), and how much of the paper could be checked (1). Access affects how confidently this directory can assess a study, not the study's underlying quality. Blinding is reported separately because it is not relevant to every design.
Design description was checked against the directly fetched article source; conclusions remain specific to the studied setting.
Sample size or the source's lack of a reported count is stated explicitly.
The comparison or identification strategy is described from the article record.
A directly retrieved full paper is available.
Double-blind? Blinding is not stated in the source reviewed; that does not establish whether blinding was or was not used.
Research question
What the authors investigated
When does GPT-4 improve versus degrade the productivity and quality of complex knowledge work?
Design
Preregistered randomized controlled laboratory-in-the-field experiment, with realistic consulting tasks and a no-AI, GPT-4, and GPT-4-plus-prompt-engineering condition.
Sample and setting
758 knowledge workers performing 18 realistic management-consulting tasks developed with Boston Consulting Group, plus a separate out-of-frontier managerial task.
Reported result
What the study found
Within the AI capability frontier, AI users completed 12.2% more tasks, 25.1% faster, with higher quality; on an out-of-frontier managerial task they were 19% less likely to be correct.
This is a paraphrase of the source. Check the original paper for definitions, estimates, statistical uncertainty and analysis details.
Scope and transfer
Limitations to keep in view
Consulting tasks and GPT-4-era capabilities may not represent other work or later models; task performance is not a direct measure of product adoption, revenue, or organizational learning.
Possible founder application
Design AI product analytics around task-level success and verification, not aggregate usage; expose where the model helps and require review where it can degrade decisions.
Evidence access
Open full text captured directly
Publisher page directly retrieved with Firecrawl; scrape reference 01a1227f-1ac8-74ec-9bba-2a65012e56ec. Claims are limited to the reported article record. Correction notices were not separately checked.
If only an abstract or metadata was available, the profile is limited to claims supported there. A paper's paywall does not change its evidentiary status, but it limits what we can substantiate from accessible text.