The AI gap is now 3.5x: what B2B Signals reveals about frontier firms!

Hi everyone, it's me, Shii-chan! Today I've got a slightly eye-opening story about how the gap between companies is widening based on how deeply they use AI.

OpenAI News openai.com

What was announced?

From OpenAI's News, there's a new research report called B2B Signals. It's a privacy-preserving, aggregated look at how AI is spreading inside companies and how deeply it's being used, refreshed on a recurring basis. It's a business extension of OpenAI's Signals.

The short version: the thing that separates leaders from everyone else has shifted from "who has access" to "how deeply you use it."

Why it matters

Not long ago, enterprise AI was mostly about access — how many seats were deployed and whether people were trying it. That still counts, but access itself is no longer the differentiator.

What creates the gap now is depth. The report calls the top 5% (95th percentile) of firms frontier firms, and they now use 3.5x as much intelligence per worker as typical firms. A year ago (April 2025) it was 2x, so the gap is widening and starting to compound — that's the core message here.

What changes

Here's the interesting part: the gap isn't just about sending more messages. Message volume explains only 36% of it, and most of the rest comes from how much real work each interaction does.

In other words, typical firms are using AI to answer questions, while frontier firms are using it to help execute complex work. The question for leaders is moving from "how many people have access" to "where is AI deepening our workflows."

Dive Deep

The report uses tokens generated as a proxy for how much intelligence is being demanded. Tokens aren't business value directly, but they help measure how much work people are handing to AI.

The next sign of maturity is agentic workflows. Codex shows the biggest gap — frontier firms send 16x as many Codex messages per worker as typical firms. ChatGPT Agent, Apps in ChatGPT, Deep Research, and GPTs show similar directional patterns, pointing to more delegated use: coding, multi-step tasks, and context-aware research.

There's a concrete example too. Cisco uses Codex for complex software work across a large engineering org, and reported reducing build times by about 20%, saving 1,500+ engineering hours per month, and increasing defect-resolution throughput by 10-15x. The biggest gains came when they treated Codex a certain way.

the biggest gains came when they treated Codex as "part of the team."

Usage is broad but increasingly specialized. It's widest in writing and communication, but IT and Security lean into how-to and procedural guidance, Software Development and Data Science lean into coding, and Finance leans into analysis and calculation — AI is moving closer to each function's core work.

In insurance, Travelers built an AI Claim Assistant with OpenAI, and expects it to handle roughly 100,000 first notice of loss calls in its first year.

And one of the clearest markers of leaders was education and learning — they use AI not only to get work done, but to help employees build the skills and habits to use AI well.

Wrap-up

• Frontier firms (top 5%) use 3.5x as much intelligence per worker, and the gap is starting to compound
• Most of the gap is about depth — message volume explains only 36%
• The next maturity sign is agentic use; Codex shows a 16x gap, and Cisco reported cutting build times ~20%
• To move toward the frontier: measure depth, build governance for production, invest in enablement, scale the teams that work, and move from chat to delegated work with agents

This is less a new feature and more a data-driven map of where enterprise AI is heading. If you're an engineering leader trying to deepen AI adoption, or figuring out where an agent like Codex fits inside your org, this should be a helpful compass.