What does research say about AI and productivity at work?

A year ago, claims about AI and productivity were mostly guesses. Now there is a small but solid body of research in which real workers did real tasks, some with AI and some without, and researchers compared the results. Three studies stand out, and together they give a more careful picture than either the enthusiasts or the skeptics usually offer.

All three use some form of randomized comparison: participants are assigned by chance to work with or without the AI tool, so differences in results can be credited to the tool rather than to who chose to use it. That design is what makes them worth reading closely.

Three studies at a glance

What each study measured

Writing tasks. Shakked Noy and Whitney Zhang asked marketers, grant writers, data analysts, managers and others to do two 20-to-30-minute writing tasks. Half were shown ChatGPT before the second task. Besides speed and quality, they tracked how people spent their time. With ChatGPT, drafting time fell by more than half and editing time more than doubled. Many participants pasted in the first output with little change.

Customer support. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a company that rolled out a generative AI chat assistant to its agents in stages, which let them compare agents before and after access. Beyond the 14% gain in issues resolved per hour, customers responded more positively, fewer asked to speak to a manager, and more agents stayed in the job.

Consulting. The study led by Fabrizio Dell'Acqua gave consultants GPT-4, some with a short training on prompting. Inside what the authors call the "frontier" of AI ability, such as generating and developing ideas for a new shoe, AI users did more, faster and better. Then came one task built to sit outside that frontier: a recommendation that depended on spotting subtle clues in interview notes that contradicted a spreadsheet. Here the AI users did worse: their rate of correct answers was 19 percentage points below that of consultants working alone.

The limits to keep in mind

Each study looks at specific tasks over hours or months, not whole careers or whole companies. The writing tasks were short and self-contained. The support study covers one firm and one tool. The consulting tasks were realistic but designed for an experiment. None of them shows what happens to wages, hiring or job design over years, and none can tell you whether your own work falls inside or outside the frontier. That is something you have to test.

Why do some workers gain more than others?

The most consistent finding is that AI narrows the gap between less and more experienced workers. In the support study, the biggest gains went to newer and lower-skilled agents, while the most experienced saw little change. The authors suggest the assistant passes on the know-how that top agents have built up, so a new agent gets some of their judgment on day one. Noy and Zhang saw something similar: those who scored lowest on the first task improved most. In the consulting study, people below the average baseline improved by 43%, against 17% for those above it.

The second reason is about judgment, not skill level. The AI's abilities are uneven, which the consulting authors call a "jagged" frontier: two tasks that look equally hard to a person can be easy for the AI or beyond it. Workers who know where that edge lies gain the most. The researchers observed two successful styles. Some consultants, nicknamed "Centaurs," split the work, handing certain parts to the AI and keeping others. Others, called "Cyborgs," worked with the AI continuously, checking and steering as they went.

What the studies warn against is the third pattern: accepting output without checking it on a task the AI can't really do. The tool's confidence doesn't change with its accuracy, so the gains go to people who keep asking whether this particular task is one it handles well.