OpenAI: Leading Companies Use AI 3.5 Times More Deeply and Adopt Codex More Actively

OpenAI has launched B2B Signals, a new research format on how businesses actually adopt AI inside their organizations. The first issue, published on May 6, 2026, shows that the gap between typical and advanced organizations is growing fast, and the main factor is not access to tools but how deeply they are used.

The gap is widening

In the report, OpenAI compares "frontier" companies, organizations at the 95th percentile of AI use, with typical companies at the median. The difference is already striking: leaders use 3.5 times more "intelligence" per employee than the average business. A year earlier the gap was half that, around 2x.

The report's logic is simple: access to ChatGPT or enterprise licenses alone no longer creates an advantage. What matters more is how deeply AI is embedded in teams' real work.

OpenAI stresses that message volume is not the main indicator of maturity. It explains only 36% of the gap between leaders and the "middle of the market." The rest comes from more substantive use: employees give models more context, set harder tasks and get more useful results.

In other words, the typical company still often uses AI as a smart search engine or a helper for quick answers, while leading companies already hand it chunks of real, complex work.

The agentic stage

OpenAI saw the largest gap in agentic and advanced tools. According to the report, frontier companies send 16 times more messages per employee to Codex than typical ones. Similar patterns appear with ChatGPT Agent, Apps in ChatGPT, Deep Research and GPTs.

This is an important shift: the market is moving away from the "ask, get an answer" model toward one where AI helps write code, carry out multistep tasks, use a company's internal context and bring work to a more finished result.

For business, this means not just buying new licenses but restructuring processes. The better AI gets at working with files, codebases and long chains of actions, the more important governance, access rules and the habit of delegating a full task, not a snippet of text, to the machine become.

OpenAI states plainly that leaders are building an operational muscle: they use AI not as a faster interface but as a way to reorganize how teams work.

What companies do in practice

The report shows that AI adoption is becoming both mass-market and specialized. The broadest use cases so far involve writing and communication, but differences by function emerge beyond that: IT and security teams more often use models for procedural instructions, developers and data science for code, and finance for analysis and calculations.

There is no single leaderboard: one company may be stronger in ChatGPT, another in the API, a third in Codex or internal AI services.

• Cisco reports roughly a 20% reduction in build time when using Codex in production scenarios.
• The company also estimates savings of more than 1,500 engineering hours per month.
• Cisco's defect-fixing throughput rose 10–15 times.
• Travelers expects its AI Claim Assistant to process about 100,000 first-notice insurance claims in its first year.

OpenAI notes one more sign of maturity: leaders use AI not only to perform tasks but also to train employees. In education and learning scenarios in particular, the gap between advanced and typical companies was among the most pronounced.

This is a good signal for large organizations: the payoff comes not when a tool is simply handed to everyone, but when people have learned to phrase complex requests, trust agents with part of the process and regularly return to these scenarios in their work.

What this means

The key conclusion from B2B Signals is that the next round of competition in enterprise AI will be over the quality of adoption, not the number of users. If OpenAI's data is accurate, the advantage goes to companies that move AI from assistant mode to executor mode sooner than others, especially in development, analytics, support and other functions where results can be measured in hours, speed and throughput.