The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index.
Generative AI at Work
Open accessAbstract
We study the staggered introduction of a generative AI-based conversational assistant using data from 5,179 customer support agents.Access to the tool increases productivity, as measured by issues resolved per hour, by 14 percent on average, with the greatest impact on novice and lowskilled workers, and minimal impact on experienced and highly skilled workers.We provide suggestive evidence that the AI model disseminates the potentially tacit knowledge of more able workers and helps newer workers move down the experience curve.In addition, we show that AI assistance improves customer sentiment, reduces requests for managerial intervention, and improves employee retention.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
17 Own result · 6 Own method · 1 No basis given
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Access to the tool increases productivity, as measured by issues resolved per hour, by 14% on average”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In addition, we find that AI assistance improves customer sentiment, increases employee retention, and may lead to worker learning.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We document within-agent changes in communication following AI deployment, with larger changes for lower-skill workers. Acrossperson, we show that these changes increase the similarity of communication patterns between lowand high-skill agents.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Using information on software outages in which AI assistance is temporarily unavailable, we provide evidence that exposure to AI leads to durable changes in worker skills. We find that workers exposed to AI recommendations continue to perform better during outages”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“we show that workers are selective about the recommendations they adopt, following the recommendations 35% on average.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In Panel A, we show that the productivity impact of AI assistance is most pronounced for workers in the lowest skill quintile (leftmost side), who see a .29 log point or 34% increase in resolutions per hour.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“These changes come alongside a substantial decrease in worker attrition, which is driven by the retention of newer workers.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Agents with less than 1 month of tenure improve their resolutions per hour by .38 log points or 46% improvement (relative to agents of the same tenure who do not have access to AI assistance). In contrast, we see no effect for agents with more than a year of tenure.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers.”
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“with agent and tenure fixed effects, we find that the deployment of AI increases RPH by 0.30 chats or 13.8%.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Generative AI models often require significant computational power and resources for training and inference such as GPUs or cloud computing services [8, 15].”
From Generative AI in healthcare: an implementation science informed translational path on application, integration and governance (2024) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“However, generative AI technologies like GANs and LLMs are complex to understand and implement [8].”
From Generative AI in healthcare: an implementation science informed translational path on application, integration and governance (2024) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“matching prior findings that generative AI assistance reduces the “skill gap” by providing the largest benefits for the weakest students (7–9).”
From Generative AI without guardrails can harm learning: Evidence from high school mathematics (2025) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“LLM can be used to summarize and synthesize vast amounts of information (eg, legislative analysis tools), manage customer inquiries (eg, chatbot customer service applications), support organizational decision-making (eg, resume screening and job matching platforms) and generate technical and creative content (eg, content ideation tools) (17–21).”
From Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force (2026) · cites this paper as [17]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“In all three instances, AI tools supplemented expertise rather than displaced experts.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
Experimental — accuracy not yet measured
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“AI tools helped novice workers attain the capabilities of experienced agents in three months rather than 10.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Buffered by the AI tool, support workers experienced significantly less hostility both from their clients and also toward their clients.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“As an unexpected side-effect, quit rates among new agents also fell substantially, likely due to less customer rage directed at them in chat windows.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“However, the field of AI is evolving rapidly, particularly in areas such as Generative AI (Brynjolfsson et al. 2023) and Large Language Models (Min et al. 2023).”
From Artificial intelligence for literature reviews: opportunities and challenges (2024) · cites this paper as Brynjolfsson et al. 2023
The cited source says what the claim attributes to it.
In the citing paper
“That generative AI particularly benefited less able writers is paralleled in recent studies focusing on other domains in which generative AI has been shown to help less productive workers (4, 5).”
From Generative artificial intelligence enhances creativity but reduces the diversity of novel content (2023) · cites this paper as [5]
The cited source says what the claim attributes to it.
In the citing paper
“Previous studies have found that LLMs increase productivity the most for low-performing workers (Brynjolfsson et al. 2023, Noy and Zhang 2023, Dell'Acqua et al. 2023).”
From AI Meets the Classroom: When Do Large Language Models Harm Learning? (2024) · cites this paper as Brynjolfsson et al. 2023
The cited source says what the claim attributes to it.
In the citing paper
“Generative AI requires large amounts of data to train models effectively [8].”
From Generative AI in healthcare: an implementation science informed translational path on application, integration and governance (2024) · cites this paper as [8]
The cited source says what the claim attributes to it.
In the citing paper
“Generative AI models learn from data [8], so the healthcare organisation needs to collect and prepare relevant data for training the models.”
From Generative AI in healthcare: an implementation science informed translational path on application, integration and governance (2024) · cites this paper as [8]
The cited source says what the claim attributes to it.
In the citing paper
“Broadly speaking, these studies have focused on productivity, finding that workers can perform knowledge-intensive tasks significantly more efficiently when given access to generative AI.”
From Generative AI without guardrails can harm learning: Evidence from high school mathematics (2025) · cites this paper as [8]
The cited source says what the claim attributes to it.
In the citing paper
“These results are consistent with the existing literature on the productivity effects of generative AI (7, 8).”
From Generative AI without guardrails can harm learning: Evidence from high school mathematics (2025) · cites this paper as [8]
The cited source says what the claim attributes to it.
In the citing paper
“Unlike in workplaces (Brynjolfsson et al. 2023, Dell'Acqua et al. 2023), LLMs appear to increase inequality in education.”
From AI Meets the Classroom: When Do Large Language Models Harm Learning? (2024) · cites this paper as Brynjolfsson et al. 2023
The cited source says what the claim attributes to it.
In the citing paper
“Many of the reported productivity gains of AI have occurred in cases in which the technology complemented human work (Fügener et al. 2022, Brynjolfsson et al. 2023, Dell'Acqua et al. 2023, Keppler et al. 2024).”
From AI Meets the Classroom: When Do Large Language Models Harm Learning? (2024) · cites this paper as Brynjolfsson et al. 2023
The cited source says what the claim attributes to it.
In the citing paper
“They also estimated a significant, roughly 14% improvement in productivity, and akin to Noy and Zhang's study, these gains were most pronounced among novice workers.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
The cited source says what the claim attributes to it.
In the citing paper
“As an unexpected side-effect, quit rates among new agents also fell substantially, likely due to less customer rage directed at them in chat windows.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
The cited source says what the claim attributes to it.
In the citing paper
“They also estimated a significant, roughly 14% improvement in productivity, and akin to Noy and Zhang's study, these gains were most pronounced among novice workers.”
From Applying AI to Rebuild Middle Class Jobs (2024) · cites this paper as Brynjolfsson et al., 2023
In the citing paper
“Brynjolfsson et al. (2023) report a 14% productivity increase using Generative AI-based conversational assistants for customer service representatives [9].”
From Sentiment Analysis in the Age of Generative AI (2024) · cites this paper as [9]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ NBER WORKING PAPER SERIES GENERATIVE AI AT WORK > 3.2 Summary Statistics
“Most agents in our sample, 83 percent, are located outside the United States, primarily in the Philippines.”
§ NBER WORKING PAPER SERIES GENERATIVE AI AT WORK > 5 Conclusion
“| Share US Agents | .11 | .15 | .081 | .072 |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ NBER WORKING PAPER SERIES GENERATIVE AI AT WORK > 4.6.3 Vertical and Horizontal Workflow
“In Column 1 of Table 5, we find that AI assistance generates an almost 25 percent decline in customer requests to speak to a manager.”
§ NBER WORKING PAPER SERIES GENERATIVE AI AT WORK > 5 Conclusion
“| | (1) | (2) | (3) | | --- | --- | --- | --- | | VARIABLES | Leaves this Month Request Manager Transfer | | | | Post AI X Ever Treated | -0.0868** | -0.00875*** | 0.00792 |”
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
Jakubův, J., Chvalovský, K., Goertzel, Z., et al. (2023) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.itp.2023.19 · Cited by 76,648
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
Neural Machine Translation by Jointly Learning to Align and Translate
Bahdanau, D., Cho, K., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1409.0473 · Cited by 16,213 · cited 1×
“Together, this approach enables models to capture long-range semantic relationships within an input text, even when that text is broken up into smaller segments and processed in parallel (Vaswani et al., 2017; Bahdanau et al., 2015).”
Scaling Laws for Neural Language Models
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020) · arXiv · doi:10.48550/arxiv.2001.08361 · Cited by 9,550
Difference-in-Differences with multiple time periods
Callaway, B., & Sant’Anna, P. H. C. (2021) · Journal of Econometrics · doi:10.1016/j.jeconom.2020.12.001 · Cited by 9,126
Difference-in-differences with variation in treatment timing
Goodman-Bacon, A. (2021) · Journal of Econometrics · doi:10.1016/j.jeconom.2021.03.014 · Cited by 8,046 · cited 1×
“…assumptions about the homogeneity of treatment effects, and may be biased when treatment effects vary over time or by adoption cohort (Cengiz et al., 2019; de Chaisemartin and D’Haultfœuille, 2020; Sun and Abraham, 2021; Goodman-Bacon, 2021; Callaway and Sant’Anna, 2021; Borusyak et al., 2022).”
The Skill Content of Recent Technological Change: An Empirical Exploration
Autor, D. H., Levy, F., & Murnane, R. J. (2003) · The Quarterly Journal of Economics · doi:10.1162/003355303322552801 · Cited by 6,205
Estimating dynamic treatment effects in event studies with heterogeneous treatment effects
Sun, L., & Abraham, S. (2021) · Journal of Econometrics · doi:10.1016/j.jeconom.2020.09.006 · Cited by 5,518
Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects
de Chaisemartin, C., & D’Haultfœuille, X. (2020) · American Economic Review · doi:10.1257/aer.20181169 · Cited by 5,031
Research works printed but not linked: 4
- [2]2. and Pascual Restrepo, "Low-Skill and High-Skill Automation," Journal of Human Capital, June 2018, 12 (2), 204-232.
- [4]4. Philippe Aghion, Claire Lelarge, John Van Reenen, and Fabrizio Zilibotti, "Tech- nology, Information, and the Decentralization of the Firm*," The Quarterly Journal of Eco- nomics, November 2007, 122 (4), 1759-1799. _eprint: https://academic.oup.com/qje/article- pdf/122/4/1759/5234557/122-4-1759.pdf.
- [57]57. Polanyi, Michael, The Tacit Dimension, Chicago, IL: University of Chicago Press, May 1966.
- [61]61. Rosen, Sherwin, "The Economics of Superstars," The American Economic Review, 1981, 71 (5), 845-858.
Cited by
Sorted by citations
Experimental evidence on the productivity effects of generative artificial intelligence
Noy, S., & Zhang, W. (2023) · Science · doi:10.1126/science.adh2586 · Cited by 2,150
“These results are consistent with other studies showing productivity-enhancing and equalizing effects of recent AI technologies (8, 15, 16, 18).”
ResultBrynjolfsson, E., Li, D., & Raymond, L. (2025) · The Quarterly Journal of Economics · doi:10.1093/qje/qjae044 · Cited by 1,083
Self-citationGenerative artificial intelligence enhances creativity but reduces the diversity of novel content
Doshi, A., & Hauser, O. (2023) · doi:10.2139/ssrn.4535536 · Cited by 1,350
“…( 3 ), increase quality and efficiency of production of typical white-collar work ( 4 ), promote productivity in customer support relations ( 5 , 6 ), speed up programming tasks ( 7 ), and enhance persuasion messaging ( 8 ), little is known about generative AI’s potential impact on a…”
Bankins, S., Ocampo, A. C., Marrone, M., Restubog, S. L. D., & Woo, S. E. (2024) · Journal of Organizational Behavior · doi:10.1002/job.2735 · Cited by 799
Generative artificial intelligence
Banh, L., & Strobel, G. (2023) · Electronic Markets · doi:10.1007/s12525-023-00680-1 · Cited by 884
“…GET3D (Gao et al., 2022), and Point-E (Nichol et al., 2022), which generate realistic and complex 3D models that facilitate a range of applications from product design and architecture to virtual reality and game development with their customers (Brynjolfsson et al., 2023; Mondal et al., 2023).”
Background
Metadata
- Authors
- Brynjolfsson, E., Li, D., & Raymond, L.
- Affiliations
- National Bureau of Economic Research (US)
- Venue
- National Bureau of Economic Research
- Publisher
- National Bureau of Economic Research
- Published
- 2023-04
- OpenAlex
- W4366817968
- Cited by
- 995
- References
- 61
- Citations by year
- 2026: 287 · 2025: 301 · 2024: 352 · 2023: 51
- Type
- report
- Language
- en
- Open access
- green
- doi.org
- Locations
- National Bureau of Economic Research (publishedVersion)
- Topics
- Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Labor market dynamics and wage inequality
- Social Sciences › Business, Management and Accounting › Organizational Behavior and Human Resource Management › AI and HR Technologies
- Social Sciences › Business, Management and Accounting › Marketing › AI in Service Interactions
- Keywords
- generative AI, chatbots, productivity, employee experience, tacit knowledge, knowledge transfer, employee retention, customer sentiment, management interventions, artificial intelligence adoption
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-09-25 (prompt extract-v1).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
