BIGGER PICTURE Learning & work
Helping a new support agent find the expert’s next move
When a customer is stuck, a new support worker may spend precious minutes searching for the right explanation. In one company, an AI assistant brought suggested responses and relevant documentation into the live conversation.

- Workers studied
- 5,172 customer-support agents
- Average result
- 15% more issues resolved per hour
- Largest gains
- Less-experienced and lower-skilled workers
How the idea works
Read the support context
The assistant uses the current customer conversation.
Offer a useful next response
Suggested replies connect to technical documentation.
Let the worker decide
Agents accept, adapt or ignore suggestions; outcomes are measured.
Knowledge at the moment of need
The assistant watched the support exchange and offered a possible reply together with relevant technical material. Customers did not interact directly with it. The worker could use, change or ignore each suggestion. In this design, the AI acted as a bridge between the company’s accumulated knowledge and the person answering a customer right now.
Sources: [1]
The rollout created a comparison
Researchers analysed a staggered introduction across 5,172 agents. Access was associated with 15% more resolved issues per hour on average, with the biggest improvements among newer and lower-skilled workers. This main analysis used the rollout’s timing rather than a large randomized assignment, an important distinction when interpreting the result.
Sources: [1]
A better conversation, not only a shorter one
The peer-reviewed study also examined what happened inside conversations: customers were more positive, and requests to escalate to a manager fell. The pattern suggests that assistance can spread useful communication practices, not merely accelerate typing. For a novice facing a frustrated customer, finding an understandable explanation can matter as much as finding the correct technical fact.
Sources: [2]
Experienced workers still need room to disagree
Benefits were uneven, and the most skilled workers showed small quality declines. That argues against treating every suggestion as best practice. The practical lesson is to support judgement rather than enforce an AI script, and to measure outcomes separately for beginners and experts. One company’s support operation does not establish the same effect in every workplace.
Sources: [2]
The original sources
- Brynjolfsson, Li and RaymondGenerative AI at Work: full author manuscript
- Brynjolfsson, Li and Raymond / arXivGenerative AI at Work: publication record and abstract