OpenAI's B2B Signals reframes AI maturity around depth of use, not seat count

What B2B Signals actually measures

OpenAI has launched B2B Signals, a recurring report built from de-identified, aggregated enterprise usage of its products. The stated goal is to track how AI is diffusing inside businesses: how deeply it is used, which tools correlate with leading adoption, and where use cases are broadening across industries and functions.

The headline metric is unusual. Rather than counting seats or message frequency, the report uses tokens generated as a proxy for 'intelligence demanded.' OpenAI is explicit that tokens are not a direct measure of business value, but treats them as a signal of how much work employees are asking the model to do per interaction.

That choice matters because it reframes the adoption question. The first phase of enterprise AI was about access — who had tools deployed. OpenAI's argument is that access is no longer the differentiator, and its new metric is designed to capture what it thinks now is.

The depth claim and the 36% figure

OpenAI reports that firms at the 95th percentile of usage — its definition of 'frontier' — now demand 3.5x as much intelligence per worker as typical firms, up from 2x in April 2025. The company presents this widening as evidence that the frontier advantage is 'beginning to compound.'

The most load-bearing statistic is smaller and more specific: message volume explains only 36% of that gap. If accurate, the majority of the difference between leading and typical firms comes not from sending more messages but from each message doing more — richer context in, more substantive output back.

This is the distinction OpenAI draws between using AI to answer questions and using it to execute complex work. It is worth noting the whole framing rests on tokens as a stand-in for complexity, a proxy the report itself flags as imperfect.

Codex and the shift toward delegation

The sharpest gap appears in agentic and advanced tools. Frontier firms send 16x as many Codex messages per worker as typical firms — the largest single disparity in the report. ChatGPT Agent, Apps in ChatGPT, Deep Research, and GPTs show similar directional patterns.

OpenAI cites Cisco as its Codex production example: roughly 20% faster build times, 1,500-plus engineering hours saved per month, and a 10-15x increase in defect-resolution throughput. The report attributes the biggest gains to treating Codex as, in Cisco's words, 'part of the team.'

The firms moving first are building the operating muscle to use AI not just as a faster interface, but as a way to redesign work from the ground up.Montana Labs

On the production side, Travelers Insurance's AI Claim Assistant guides customers through first notice of loss and creates claims inside Travelers' systems; OpenAI says it expects the assistant to handle roughly 100,000 such calls in its first year. These are the two named proof points for the delegation thesis.

Reading a vendor's own usage data with care

B2B Signals is a measurement product built by the vendor whose consumption it measures, and its central unit — tokens — is also OpenAI's billing unit. The narrative that depth beats access, and that Codex adoption marks maturity, aligns neatly with buying more of OpenAI's advanced tools.

That does not make the findings wrong, but it sets the terms of engagement for teams reading them. The useful, testable claims here are narrow and specific: that per-interaction depth, not message count, drives most of the observed gap, and that agentic tooling shows the widest spread between leading and typical firms.

The concrete implication for engineering leaders is a measurement one. If depth is the differentiator OpenAI describes, then tracking seats and message counts will systematically understate — or overstate — where a team actually stands. The harder work is instrumenting which workflows have moved from chat assistance to delegated, tool-using execution, and validating that token-heavy use maps to real business outcomes rather than just larger bills.