À suivreGuide suivant
How Deep Research Agents Work
Applications
GUIDE DE LA SOCIÉTÉ
What Research Says About AI and Workplace Productivity
Controlled field experiments show generative AI can make workers substantially faster and better at specific tasks such as customer support, writing and some coding, with the largest gains usually going to less experienced workers.
Sur cette page4 minutes de lecture
Aperçu
The gains are uneven, though. AI can make people worse on tasks outside its abilities, and broad studies of whole workforces and economies so far find much smaller effects than lab and single-task studies do.
Plongée profonde
The best-known evidence comes from field and controlled experiments on well-defined tasks. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied more than 5,000 customer support agents who were given an AI assistant. On average they resolved about 14 percent more issues per hour. Novice and lower-skilled agents gained around 34 percent, and the most experienced agents gained little. The researchers argue the AI spread the know-how of top performers to everyone else. Shakked Noy and Whitney Zhang, publishing in Science in 2023, found that ChatGPT cut the time college-educated professionals spent on writing tasks by about 40 percent and raised rated quality by about 18 percent. A GitHub Copilot experiment found developers finished a set programming task about 55 percent faster. The 2023 Boston Consulting Group study with Harvard and other researchers added an important caveat, which it called the 'jagged technological frontier'. On tasks inside the AI's abilities, consultants with GPT-4 finished more tasks, faster and at higher quality. On a task designed to sit outside those abilities, they were less likely to get the right answer than consultants without AI, because they trusted plausible but wrong output. More recent findings are more modest. In a 2025 randomized trial by METR, experienced developers working in their own large codebases were about 19 percent slower with AI tools while believing they had been faster. Studies of whole labor markets, such as research on Danish workers, have found small effects on earnings and hours so far. Why the gap? Experiments isolate tasks that suit AI. Real jobs mix many tasks, time saved is not always reused productively, and organizations need to redesign workflows before gains show up. Economists saw a similar lag in measured productivity after electricity and computers arrived. The common misconception is that one headline percentage applies to every job.
Impact stratégique
Risques et sécurité
Les dommages catastrophiques et quotidiens causés par l’IA dépendent tous deux de la personne qui comprend les risques et qui peut agir.
Décisions plus claires
Les connaissances du public et des professionnels déterminent si une politique de sécurité forte est politiquement possible.
Passer à travers le battage médiatique
Des explications claires réduisent la capture par le battage médiatique, les relations publiques en laboratoire et le théâtre d'éthique vague.
The Future of What Research Says About AI and Workplace Productivity
The evidence base is growing quickly, and the models studied in 2023 are already outdated, so results may shift as tools, training and workflows improve. The key open questions are whether task-level gains add up to firm-level and national productivity growth, whether benefits keep concentrating among less experienced workers, and how deskilling or over-reliance affects people in the long run. Careful researchers expect a lag like the one seen with earlier general-purpose technologies. Treat both very large and near-zero headline numbers with caution until more long-term, economy-wide data exists.
Mise en œuvre dans le monde réel
In a large customer support study, agents given an AI assistant resolved more issues per hour on average, and the newest, least experienced agents improved the most.
In an experiment with professional writing tasks, participants using ChatGPT finished faster and produced work that graders rated higher in quality.
Consultants at Boston Consulting Group did better with GPT-4 on tasks inside the AI's abilities, but were more likely to reach wrong answers on a task deliberately designed to fall outside them.
Experienced open-source developers in a 2025 randomized trial by METR took longer to finish tasks with AI tools, even though they believed the tools had made them faster.
Risques et garde-fous
Traiter le risque existentiel comme de la science-fiction alors que les capacités s’accroissent.
Confondre sécurité des produits de surface et alignement sous haute autonomie.
Laisser le public non anglophone et non expert avec uniquement des sources de mauvaise qualité.
Feuille de route de mise en œuvre
Séparez les dommages causés aux produits, leur mauvaise utilisation et les risques de perte de contrôle/désalignement.
Demandez quelles preuves pourraient changer votre point de vue sur les délais et la gravité.
Préférez les sources primaires et les évaluations concrètes aux allégations marketing.
Identifiez une voie d’action : carrière, politique, financement ou compétences – et pas seulement la sensibilisation.
Continuez à explorer
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the What Research Says About AI and Workplace Productivity quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Questions fréquemment posées
What is What Research Says About AI and Workplace Productivity?
Controlled field experiments show generative AI can make workers substantially faster and better at specific tasks such as customer support, writing and some coding, with the largest gains usually going to less experienced workers. The gains are uneven, though. AI can make people worse on tasks outside its abilities, and broad studies of whole workforces and economies so far find much smaller effects than lab and single-task studies do.
In the Brynjolfsson, Li and Raymond customer support study, which agents gained the most from AI?
Novice and lower-skilled agents improved by around 34 percent, while the most experienced agents gained little. The average gain was about 14 percent.
What did Noy and Zhang's 2023 Science study find about ChatGPT on writing tasks?
Professionals finished writing tasks much faster, and graders rated the quality higher.
What does the 'jagged technological frontier' describe?
AI's abilities are uneven. Inside the frontier it helps, and outside it people who trust it can do worse.
In the BCG study, what happened on the task designed to fall outside AI's abilities?
Consultants using GPT-4 on that task were less likely to be correct, because they trusted plausible but wrong output.
What did METR's 2025 randomized trial with experienced developers find?
Measured completion times were slower with AI, yet the developers believed it had sped them up. This shows how unreliable self-reports can be.
Continuez à apprendre
Guides associés
Plus de guides sélectionnés pour ce sujet