INTERNATIONAL RESEARCH / PEER-REVIEWED ARTICLE

Does AI really make customer-service agents more productive?

Data from 5,172 agents and three million chats showed a 15% average rise in issues resolved per hour with AI assistance. We examine why less-experienced workers gained most.

ACADEMIC WORK REVIEWED

Generative AI at Work

ResearchersErik Brynjolfsson, Danielle Li and Lindsey Raymond.

The Quarterly Journal of Economics · 4 February 2025 · 140(2):889–942 · Peer-reviewed article · Source language: English

Open the original publication Find on Google Scholar

ENGLISH REVIEW AND COMMENTARY: DIJITALPI

We explain the paper in plain English, discuss what it may mean for organisations and identify our own examples separately.

Read the explanation Go to 3 business examples

English review: 10 September 2026 · The source date appears in the citation above.

Published review

Source and editorial checks are complete. DijitalPi did not conduct a customer experiment for this paper.

LET'S READ THE RESEARCH TOGETHER

We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.

01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?

How do AI suggestions change customer-service speed and resolution quality?

Brynjolfsson, Li and Raymond studied the effect of an AI chat assistant on productivity, quality and learning in real customer service.

Explaining the context · DijitalPi commentary

Closing more chats is insufficient if problems remain unresolved or customers return.

The system learned patterns from successful conversations and provided suggestions in real time, making the study partly about how tacit knowledge spreads.

02 / HOW WAS THE RESEARCH CONDUCTED?

How did they test the question?

The researchers observed a staggered rollout at a Fortune 500 software company. The data covered about three million chats from 5,172 agents and compared access timing with worker outcomes.

Understanding the method · DijitalPi commentary

This was an econometric analysis of a staggered rollout, not a randomized trial. It provides comparison within teams but remains one company.

The primary productivity measure was successfully resolved chats per hour, combining duration, simultaneous chats and resolution.

03 / RESEARCH FINDINGS

Resolved work per hour rose 15% on average, with unequal gains across workers.

AI access increased resolved chats per hour by 15% on average. Less-experienced and lower-skilled workers gained more in speed and quality, while the most experienced group had small speed gains and slight declines on some quality measures.

Agents: 5,172Customer-service workers observed at one company.
Average productivity: +15%Average change in successfully resolved chats per hour.

Source: QJE full article and published results ↗

04 / DIJITALPI'S EXPLANATION

How should we interpret these findings?

Value may come from transferring patterns used by experienced workers to newer staff.

An average effect is not an equal effect. Training, targets and review should vary with worker experience.

If top workers’ conversations train the system, the ownership and freshness of that knowledge also need management.

05 / CONCLUSION AND OPEN QUESTIONS

What did we learn, and what do we still not know?

AI assistance improved productivity and learning in this customer-service setting. One company, an older model and a staggered rollout mean that 15% is not a ready-made promise for every support team.

Which workers speed up and which errors grow in a Turkish support workflow? Test with your own documents, repeat-contact records and quality measures.

Return to the researchers' original publication

QJE full article and published results NBER record and version history

DIJITALPI'S APPLICATION COMMENTARY

What could this look like in your organisation?

We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.

EXAMPLE 01

The average becomes everyone’s target

The company adds the 15% result to every employee’s performance target.

Should every experience group show the same gain?

Open the recommendation for example 01

Break the results down by experience.

Report speed and quality separately for new, mid-level and experienced workers. Do not turn an average effect into an individual promise.

EXAMPLE 02

The chat is shorter, but the problem returns

Conversations end faster, yet some customers return with the same issue.

Is a shorter chat productive by itself?

Open the recommendation for example 02

Measure successful resolution end to end.

Track first-contact resolution, seven-day repeat contact and escalation alongside time. Do not move the problem into the next conversation.

EXAMPLE 03

Expert knowledge trains the system

The AI learns patterns from successful agents’ conversations, but their contribution is invisible.

Is this only a software investment?

Open the recommendation for example 03

Manage the human contribution to knowledge.

Define quality checks for training examples, worker contribution and update ownership. Do not freeze yesterday’s best practice into a permanent rule.

TRY IT WITH YOUR TEAM

Measure one support queue by experience.

In a four-week pilot, separate workers by experience. Put resolution quality and repeat contact next to speed to see who the system helps and how much.

Explore our services

Source and ownership

Erik Brynjolfsson, Danielle Li and Lindsey Raymond.
Stanford University; Massachusetts Institute of Technology; National Bureau of Economic Research.

The Quarterly Journal of Economics · 4 February 2025 · 140(2):889–942 · Peer-reviewed article. Open the original publication · Find on Google Scholar

The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.

← Return to all research articles