AI-Assisted Customer Service: What the Numbers Really Show

No other area of work has been studied as closely as customer service. Tickets can be counted, handling times are measured anyway, and the tasks repeat. That makes it the best place to see what AI does at work when you measure it instead of guessing.

What the research shows

The most important study so far comes from Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in 2025 in the Quarterly Journal of Economics. They analysed an AI assistant used by 5,179 support agentsS. The assistant suggested replies; people checked and sent them. On average, teams resolved 14 % more issues per hourS. New agents gained 34 %S, very experienced agents barely changed. Customer sentiment improved, escalations fell and so did staff turnover.

The lesson: the assistant spreads the know-how of the best people to everyone else. If you want new hires productive faster, this is the strongest lever.

The Klarna counterexample

Klarna took the other route. In February 2024 the company said its AI assistant had handled 2.3 million conversationsH in its first month, the equivalent of about 700 full-time agents. From May 2025, Klarna started hiring people for customer service again. CEO Sebastian Siemiatkowski conceded that the strong focus on cost had led to lower quality. Klarna also says AI now does the work of more than 800 rolesH.

Both can be true. Full automation cuts costs. Without quality measurement, you notice the price late.

A shop-floor example

ClickUp shows how concrete the time savings in ticket preparation can be. According to a Zapier case study, research time per ticket fell from 15 to 4 minutesH, adding up to more than 917 hours a month. That is a company claim without independent review. The direction matches the research, though: gains appear where people used to search and collect information by hand.

The workflow

This workflow combines classification, knowledge search and a draft reply. People stay in charge of sending and of every sensitive case.

The last step matters most. Measure handling time, repeat-contact rate and customer satisfaction before and after the rollout. Without a baseline there is no reliable comparison.

Limits and risks

• Hallucinations: Generate replies only from approved help-centre articles and show the source in the draft.
• Transparency: Art. 50 of the AI Act applies since 2 August 2026. If customers talk to a chatbot directly, you must tell them.
• Data protection: Tickets contain personal data. No ticket goes into an AI tool without a data processing agreement and a business licence.
• Quality: If you only measure handling time, you optimise for speed. Repeat contacts and satisfaction belong on every dashboard.

How to start

1. Pick one ticket category with high volume and clear answers, such as delivery status or password resets.
2. Measure handling time and repeat-contact rate for two weeks without AI.
3. Introduce draft replies with sources. The team still sends.
4. Compare after four weeks and decide per category whether to expand.

Sources

1. SBrynjolfsson, Li, Raymond: Generative AI at Work. Quarterly Journal of Economics 140(2), 2025
2. MForbes: Klarna Reverses On AI, Says Customers Like Talking To People, 18 May 2025
3. MFinTech Weekly: Klarna Reverses Course on AI Customer Support, 12 May 2025
4. HZapier Customer Stories (ClickUp)
5. RRegulation (EU) 2024/1689 (AI Act), Art. 50