The AI productivity gap is widening, not closing — and it is compounding fast

One year ago, OpenAI's enterprise data showed frontier firms operating at roughly twice the intelligence-per-worker of typical companies. The gap has since widened dramatically. A new report published Tuesday finds that leading enterprises are now running at 3.5 times the intelligence capacity per worker compared to the median firm, a 75 percent jump in twelve months.

The finding comes from OpenAI's first B2B Signals report, drawn from anonymized usage data across its enterprise customer base. It covers the period from April 2025 to April 2026. The headline number is striking, but its implications are more complicated than a simple leaderboard.

Message volume explains only about 36 percent of the gap between frontier and typical firms, according to the report. That means the difference is not simply that leading companies are using AI more, or that their workers are more active on AI platforms. The data suggests they are using it differently, and more effectively. Frontier firms send sixteen times as many Codex messages per worker as the typical enterprise, a measure that points to sustained, programmatic engagement rather than occasional prompts.

The case studies in the report are where the abstract number becomes concrete. Cisco reduced engineering build time by 20 percent and is now saving more than 1,500 engineering hours per month across its teams. Defect resolution throughput improved by ten to fifteen times. Travelers Insurance deployed an AI Claim Assistant expected to handle roughly 100,000 first notice of loss calls per year, a volume that would represent a substantial share of its total claims intake.

These are real deployments producing measurable outcomes. McKinsey's long-range estimate puts the total AI productivity opportunity at $4.4 trillion across the global economy. If the Cisco and Travelers numbers are representative of how frontier firms are capturing that opportunity, the scale is not surprising. What is surprising is how concentrated the gains appear to be.

There is a counterweight worth taking seriously. A separate survey from Writer.com found that 75 percent of executives describe their own AI strategy as performative, and 79 percent report meaningful adoption challenges. Only 29 percent say they are seeing significant return on investment from generative AI, despite reporting five-times productivity gains at the individual worker level. The gap between what individual employees can do with AI tools and what organizations can capture institutionally is wide, and it is not closing on its own.

The compounding framing matters here. If AI adoption followed the pattern of electricity or the internet, we would expect adoption to spread relatively evenly across firms, with early advantages eroding as the technology became ubiquitous. The OpenAI data points in the opposite direction. The intelligence gap is not stable, it is widening, and it is widening fast. A firm that was twice as productive as the median a year ago is now three and a half times as productive. That is not diffusion. That is compounding.

The obvious caveat is that OpenAI is reporting on its own enterprise customers. The 3.5x figure reflects the usage patterns of companies paying for premium AI access, and those companies are not a random sample of the economy. They skew larger, more technically sophisticated, and more likely to have already invested heavily in the infrastructure to absorb AI effectively. The widening gap could partly reflect a self-selection effect: the firms most capable of using AI are also the ones most likely to buy it from OpenAI. That is a real limitation, and any story built around this data should say so plainly.

But even accounting for that selection bias, the trajectory is significant. The report's methodology tracks the same cohort of firms over time, so the 75 percent widening of the gap is not a composition effect. Within this specific group of enterprise customers, the leading firms pulled further ahead over twelve months.

What does this mean for the typical enterprise? If the compounding dynamic is real and durable, it implies a structural repricing of AI strategy for the roughly three-quarters of organizations that are not frontier adopters. AI was supposed to democratize capability. What the data suggests instead is that it may be stratifying it. CTOs and HR leaders who treated AI as a seat-based cost to be managed rather than a compounding advantage to be built are facing a narrowing window to catch up.

The $4.4 trillion opportunity is real. The question the OpenAI data raises is whether most of it will be captured by a shrinking set of firms that figured out how to use it first.