Research SummaryMulti-country · 25 locations

Research summary · The Quarterly Journal of Economics

Real-World Evidence Shows AI Can Spread Expertise—Unevenly

A large field study of customer-support work found an average productivity increase from generative AI assistance, with larger gains for less experienced workers and more limited benefits for top performers.

Generative AI at Work4 February 2025

Read the original source ↗

Conceptual visual

Independent editorial analysis

This is FUURAA’s own editorial analysis of the cited public source, prepared independently from the cited institution. Source materials remain attributable to their authors and publishers; FUURAA is responsible for their selection, synthesis and interpretation. No cited institution has reviewed or endorsed this article unless expressly stated.

External evidence

What the field study found

The study followed 5,172 customer-support agents after a generative AI assistant was introduced. On average, workers resolved about 15% more issues per hour, with the largest improvements among less experienced and lower-skilled workers.

The workforce spanned 25 locations; 89% of agents were outside the United States, mainly in the Philippines. Benefits were uneven: the most experienced and highest-skilled workers saw small gains in speed and small declines in quality.

FUURAA editorial analysis

FUURAA editorial perspective

Evidence-led analysis in the public interest

The strongest near-term case for workplace AI may be capability transfer rather than simple labour substitution. Systems can make useful organisational knowledge available at the moment of work.

Evidence from one company and one occupation cannot be treated as a universal productivity number. Each workflow needs its own evaluation of quality, safety, learning and worker experience.

Key judgments

The field study provides meaningful evidence that generative AI assistance can improve performance in a specific customer-support setting, with larger gains among less experienced workers.
The result supports a capability-transfer interpretation, but it should not be converted into a universal productivity promise or a conclusion that every occupation benefits in the same way.
Responsible workplace adoption should measure service quality, learning, worker agency and the distribution of gains and burdens—not throughput alone.

The evidence is important because it comes from real work

The study followed 5,172 customer-support agents across 25 locations after a generative AI assistant was introduced; 89% of the workforce was outside the United States, mainly in the Philippines. It reported an average increase of about 15% in issues resolved per hour. Larger improvements appeared among less experienced and lower-skilled workers, while the most experienced and highest-skilled workers saw small gains in speed and small declines in quality. This is stronger evidence than a laboratory demonstration for the setting examined because it observes people performing consequential everyday tasks. It remains evidence from one company and one occupation, however, so the reported average should not be treated as a general number for all firms, countries or types of work.

AI may spread practices as well as automate tasks

One reasonable interpretation is that the assistant made useful patterns of stronger workers available at the moment less experienced colleagues needed them. That suggests a role for AI as a channel for organisational knowledge, not simply a substitute for labour. Yet a recommendation system can transmit weak practices as readily as strong ones if its underlying examples, feedback or objectives are poor. Organisations should therefore ask what knowledge is being encoded, whose practice is treated as exemplary and whether workers can question the guidance. The study offers evidence consistent with capability diffusion; the wider institutional design is an editorial inference requiring separate evaluation.

Productivity is a distribution, not only an average

An average improvement can conceal different experiences across workers, customers and tasks. Less experienced workers may benefit from timely support, while experts may find suggestions repetitive or constraining. Faster handling may improve access, but speed can also conflict with accuracy, empathy or the time needed for unusual cases. Managers should examine who gains, who carries extra monitoring or correction work, and whether performance remains strong after the initial period. Worker voice is relevant because people closest to the workflow can identify when assistance is helpful, distracting or unsafe. Evaluation should also distinguish changes caused by the tool from changes in training, staffing or management.

The goal should be durable human capability

A workplace system can raise immediate output while weakening learning if workers become dependent on suggestions they do not understand. It can also support learning when it explains reasoning, exposes useful examples and gives people room to revise recommendations. The appropriate design depends on the task and on who remains responsible for the outcome. High-impact or unusual cases require stronger review than routine support. The public-interest objective is not maximum automation but better work: competent people, reliable services, fair opportunities to develop and clear accountability when the system is wrong.

Alternative views & uncertainty

What this evidence does not settle

AI assistance may lower barriers to competent performance, but it could also narrow discretion or standardise work around patterns that do not serve every customer or culture.
Requiring extensive measurement and consultation can slow adoption, yet deploying without them may produce hidden quality failures, unequal burdens or workforce resistance.

Public-interest implications

What this means for different stakeholders

public

Customers should receive reliable service and appropriate disclosure or human escalation where AI assistance materially shapes an important interaction.

organisations / industry

Employers should pilot by task, compare quality and learning as well as speed, involve workers and preserve accountable human handling for exceptions.

policy

Workplace governance should consider monitoring, fairness, skill development, worker participation and responsibility rather than treating productivity gains as the only public outcome.

research

Further studies should test different occupations, organisations and time horizons, including effects on expertise, service quality, wellbeing and inequality.

What to watch next

Whether measured gains persist after novelty, workflow adaptation and changes in the underlying AI system.
Whether less experienced workers build independent expertise or remain dependent on automated guidance.
How benefits, monitoring burdens, job redesign and decision authority are distributed across workers and managers.

Conclusion

The study gives credible reasons to take human–AI collaboration seriously while resisting exaggerated claims. In the observed customer-support setting, assistance improved average throughput and particularly helped workers with less experience. That finding points toward AI as a mechanism for making useful knowledge more available. It does not settle how other jobs should be redesigned. Organisations should treat adoption as a learning programme with evidence, worker participation and safeguards, aiming to extend human capability rather than merely accelerate activity.

Independence and relevance disclosure

This is FUURAA’s independent editorial analysis of the cited study published in The Quarterly Journal of Economics. It does not imply approval, participation or endorsement by the authors, journal, publisher or participating company.

Forward view

How organisations can learn responsibly

Target the task

Identify where guidance and knowledge retrieval help, rather than forcing AI into every activity.

Measure distribution

Track who benefits, who is burdened and whether gains persist after the initial deployment.

Protect learning

Design AI assistance to strengthen human capability instead of creating silent dependency.