Independent questions. Shared human futures.Powered by The Human Question ↗
Work & everyday lifeContext, evidence, perspective
Where AI Fits.

Understand the change. Choose the part you play.

Work

What a fifteen percent gain can tell a team

A workplace study offers a useful result. Applying it elsewhere begins with understanding what was measured.

Imagine a support team looking at a queue of unanswered questions. One person wants quicker replies. Another worries that a confident but incorrect answer will create tomorrow's complaint. Both concerns are reasonable. The question is not simply whether an assistant can produce more words. It is whether the whole service can improve while people can still understand and check what happened.

Brynjolfsson, Li and Raymond studied 5,172 customer-support agents at one business-software firm. AI assistance increased average issues resolved per hour by 15%, with effects varying by experience and skill. This is a workplace finding, not a universal promise.

The measure was resolved issues, not generated words. These are the published figures; earlier working-paper versions differ.

Suppose your team is considering an assistant for a routine inquiry. Start by describing an acceptable response before trying the tool. What source should support it? What would count as a serious error? When must the conversation move to a person? These are proposed planning questions. The study does not supply your answers, and this imagined inquiry is not an observation from the research.

Next, consider a bounded comparison with your usual process. Keep track of the entire attempt: finding information, drafting, checking, correcting and following up. An attractive first response can still require substantial repair. A small trial may reveal a practical problem, but it cannot establish an organization-wide effect or substitute for a careful evaluation.

Ask the people doing the work to help define that evaluation. They may notice that an exception is common, that a customer needs explanation rather than speed, or that an instruction is easy to misread. Their observations can change the plan before a small problem becomes a routine part of the workflow.

The published result gives the discussion a concrete reference point. It does not relieve anyone of choosing a sensible task, a fair comparison and an appropriate quality standard. The most useful question to carry into the next meeting is specific: what would a better resolved issue look like here, and how would we know we had achieved it?

Sources & further reading

Original editorial analysis informed by the sources below. A source’s date, method and limits matter.

  1. Generative AI at Work Scope & limits · S09