What the numbers say about agents at work

Someone looking for proof that agents raise output will find two kinds of evidence. They do not measure the same thing.

An experiment gives the tool to some people and not to others, then counts what those people finish. A survey asks companies whether cost or revenue moved. An experiment can show a change in one job. A survey can show whether companies believe they are seeing one. Neither figure is a forecast for the next company that switches a tool on.

The numbers in this piece are public. Each one points at the paper or the survey it comes from. The links are at the end.

The clearest production figure comes from customer support. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond followed the staggered rollout of a generative assistant across 5,179 agents at a Fortune 500 software company. The paper is NBER Working Paper 31161. The assistant suggested replies. The agent stayed responsible for the conversation and could ignore the suggestion.

Issues resolved per hour rose by about 14 percent. For newer and lower-skilled agents the rise was about 34 percent. An agent with two months on the job, using the tool, resolved issues at the rate of an untreated agent with more than six months of tenure. The paper also reports better customer sentiment and higher employee retention. For the most experienced agents the productivity change was small, and in places the quality of their conversations slipped.

That is a production result. More finished work per hour, concentrated where people had the most to learn, with a person still on the conversation.

Fabrizio Dell’Acqua and colleagues ran a preregistered experiment with 758 consultants at Boston Consulting Group. On 18 realistic tasks inside what the model could do, people using GPT-4 completed 12.2 percent more tasks and finished them 25.1 percent faster. BCG’s account of the same experiment reports quality about 40 percent higher on a creative product task.

On a managerial task chosen to sit outside that reach, people using the tool were 19 percent less likely to produce a correct solution. The tool stays fluent when it is wrong. For administration, analysis, and process work, the dividing line is not whether the task matters. It is whether a wrong answer can be recognized without doing the job again.

McKinsey’s global survey was in the field from June 25 to July 29, 2025. It heard from 1,993 people in 105 countries. Eighty-eight percent said their organization uses AI in at least one function. Sixty-two percent said they are at least experimenting with agents: 23 percent scaling an agent somewhere in the enterprise, and 39 percent still in experiments. In any single function, no more than 10 percent said they are scaling agents. Seven percent said AI is fully scaled across the organization.

Thirty-nine percent attributed any effect on EBIT to AI, and most of those put it under 5 percent of EBIT. Sixty-four percent said AI is enabling innovation. Reported cost decreases showed up most often in software engineering, manufacturing, and IT. These figures are what respondents said. They are not what an experiment counted. They are still the widest public picture of where companies stood in 2025.

McKinsey calls a high performer a respondent who attributes more than 5 percent of EBIT to AI and also reports significant value. That group is about 6 percent of the survey. Fifty-five percent of them say they have fundamentally redesigned workflows, against 20 percent of everyone else. Sixty-five percent have a defined process for when a person must check a model’s output, against 23 percent.

The pattern matches the experiments. The tool raises output on work that has a check. It does not raise the company’s result until the workflow is rebuilt around that check, with a person who can reject the output.

Scraplay does not transfer these percentages onto a client. A 34 percent gain belonged to newer support agents in one company, with one assistant, in the years that paper covers. The next company’s number will be different.

What we take from the evidence is the shape of the work. We look for a station you already know how to accept: a support reply, a process step, a report, a handoff of data. An agent can take that station. A person keeps the rejection. We measure time, completions, and errors, the same classes of measure these studies used. After that, the figure that matters is the one from your own workflow.

If the task has no check, the consulting experiment is the warning. That work stays with a person.