71%. That's the median productivity gain Stanford researchers found in companies running agentic AI - where software owns a task completely, from start to finish, without stopping to ask a human for approval at each step. Companies using AI in a support role, where humans stay in the loop at every decision, averaged 40% gains.
That's a 31-point gap from the same underlying technology, across 51 real deployments - not pilots, not survey responses, not vendor case studies. Actual production systems at actual companies.
What "Agentic" Actually Means
Most people interact with AI as a smart assistant: you write a prompt, it produces something, you review it, maybe ask for changes, and the cycle repeats. The human is the decision-maker; the AI is the helper.
Agentic AI flips this. You define the goal - "process these invoices" or "research 200 leads and flag the qualified ones" - and the AI handles the entire workflow without requesting approval at each step. It reads files, makes decisions, takes actions, and delivers a finished result.
The obvious downside: when something goes wrong, it has gone further wrong than you'd like before anyone notices. But the Stanford data suggests the efficiency trade-off is real.
Where the Gap Comes From
The 31-point difference has structural explanations.
Human approval loops add time. Every checkpoint where a person must review and greenlight the next step is a pause in the workflow. In agentic setups, those pauses disappear. The AI runs the whole task at machine speed.
Context-switching costs compound. When AI acts as an assistant, humans still have to hold the mental model of the project, digest the output, and decide what happens next. With agentic workflows, humans shift from operators to supervisors - checking outcomes rather than managing each step.
There's also a selection effect worth naming: companies that have successfully deployed fully autonomous AI workflows have almost certainly done harder integration work upfront. The productivity gain partly reflects organizational discipline, not just software capability.
From Assistant to Autonomous
The 40% figure isn't bad. For most people using ChatGPT, Claude, or similar tools today, that's roughly the range of real gain - faster drafts, faster research, faster first passes. Meaningful.
But if you're running a business and thinking about where AI creates competitive separation, the study suggests the answer is: give AI more autonomy, not more prompts. The companies seeing 71% gains aren't prompting harder - they've built systems where AI runs processes rather than assists with them.
That's harder to set up. It requires clear success criteria, good error handling, and tolerance for occasional autonomous mistakes. But 51 real deployments of data suggest it's where the outsized gains actually live.