B2B Signals: How Frontier Enterprises Are Scaling AI Adoption in 2026

OpenAI's new enterprise telemetry report contains a number that should reset every CEO's AI scorecard. Frontier firms now consume 3.5x more intelligence per worker than typical firms, and the gap is accelerating. But the real story is not the headline ratio — it is what the 36/64 split between volume and depth reveals about why most enterprises are asking the wrong question.

The Number That Should Reset Your AI Strategy

On May 6, 2026, OpenAI published the first edition of B2B Signals, a recurring research initiative built on privacy-preserving, aggregated telemetry from enterprise OpenAI usage. The headline finding landed in engineering managers' Slack channels within hours: frontier firms — organizations at the 95th percentile of AI usage — now consume 3.5 times as much intelligence per worker as typical firms.

A year ago, that ratio was 2x. The gap has nearly doubled in twelve months.

The natural reaction is to treat this as a procurement gap. Frontier firms bought more AI. They deployed more seats. They have bigger budgets. If we just increase our AI spend, we can close the gap.

That reaction is exactly wrong. And OpenAI's own data proves it.

Message volume explains only 36% of the 3.5x frontier advantage. The remaining 64% comes from something else entirely: richer prompts, more complex tasks, longer reasoning chains, delegated agentic workflows, and more substantive outputs. In other words, frontier firms are not just using AI more often. They are using it at a fundamentally different depth.

This distinction changes everything. If the gap were primarily about volume, the solution would be simple: buy more licenses, deploy more seats, encourage more usage. But since 64% of the gap is about depth, the solution is organizational, not financial. It is about how work is designed, how teams are structured, how governance is built, and how AI is integrated into core processes rather than bolted onto them as an afterthought.

The question for enterprise leaders is no longer "how much AI are we using?" It is "how deeply is AI embedded into how we actually work?"

What B2B Signals Actually Measures

Before interpreting the data, it is worth understanding what B2B Signals is and what it is not.

B2B Signals is the enterprise extension of OpenAI's consumer-side Signals research program, which began publishing in late 2025. The methodology uses aggregated, anonymized telemetry across ChatGPT Enterprise, ChatGPT Team, the API, and Codex to derive percentile-ranked behavioral metrics. No OpenAI employee reviews individual enterprise data. Message content is classified using automated systems. The analysis is purely statistical.

The "intelligence per worker" metric is a composite combining message volume, prompt complexity, output length, and feature adoption. It explicitly weights depth heavily. A worker who sends one long, context-rich prompt that generates a substantive multi-step output scores higher than a worker who sends ten short chat messages. This weighting is deliberate: OpenAI designed the metric to capture not just activity but the complexity of the work being delegated to AI.

The frontier cohort is defined as the top 5% of enterprises by this intelligence-per-worker metric. The typical firm is at the 50th percentile. The first report covers usage from late 2025 through Q1 2026.

This methodology has important limitations. B2B Signals only measures OpenAI product usage. It cannot capture AI adoption through competing platforms like Anthropic's Claude for Work, Google's Gemini Enterprise, or open-source models deployed on-premises. It also cannot measure the business value generated by AI usage — only the intensity and depth of that usage. An organization could score high on intelligence per worker while generating minimal business impact, or vice versa.

These caveats matter, but they do not invalidate the core finding. Even if the absolute numbers shift when non-OpenAI usage is included, the directional signal is clear: the organizations using AI most deeply are pulling away from the median, and the gap is compounding.

The 36/64 Split: Why Volume Is the Wrong Metric

The most important number in the B2B Signals report is not 3.5x. It is 36%.

That is the share of the frontier advantage attributable to message volume alone. If a typical firm somehow matched the frontier firm's message-sending rate, it would close only 36% of the 3.5x gap. The other 64% would remain.

This decomposition is the operational hinge of the entire report. It tells us that frontier firms are winning not because their employees log into ChatGPT more often. They are winning because each interaction is doing more of the actual work.

What does "deeper usage" look like in practice? OpenAI's data points to several concrete patterns:

Longer, more contextual prompts. Frontier firm workers provide richer context in their prompts — background documents, specific constraints, prior decisions, success criteria. They are not asking AI to "write an email." They are asking AI to "draft a follow-up to the Q2 pricing discussion with Acme Corp, referencing the attached contract terms, maintaining the tone of my previous messages, and flagging any terms that deviate from our standard agreement."

More complex, multi-step tasks. Frontier firms delegate work that spans multiple files, tools, and reasoning steps. A typical firm might use Codex to generate a function. A frontier firm delegates an entire feature: "Implement user authentication for the admin dashboard, including JWT token handling, role-based access control, and integration with our existing OAuth provider. Write tests for all edge cases and update the API documentation."

Agentic workflows over chat assistance. The single largest gap in the entire dataset is Codex usage: frontier firms send 16 times as many Codex messages per worker as typical firms. Codex is an agentic coding tool, meaning engineers describe a task and the AI completes it autonomously over multiple steps. This is structurally different from chat-based assistance, where the human drives each interaction. The 16x gap suggests frontier engineering organizations have moved past assisted coding into agent-delegated coding — a different productivity regime entirely.

Substantive outputs, not answers. Typical firms use AI to answer questions. Frontier firms use AI to produce artifacts: completed code reviews, drafted research reports, validated data pipelines, generated test suites. The output is not information — it is work product that moves a project forward.

The 36/64 split reframes the entire enterprise AI conversation. For the past two years, most organizations have measured AI adoption in terms of access: percentage of employees with AI licenses, number of active users, monthly message volume. These are Driver-stage metrics. They tell you whether people have AI tools. They tell you nothing about whether AI is changing how work gets done.

Frontier firms have moved beyond access metrics to depth metrics. They measure prompt complexity, delegation ratio, agentic workflow adoption, and output substantiveness. These are Architect-stage metrics. They reflect not just whether AI is present in the organization but whether the organization has been restructured around AI's capabilities.

The Agentic Gap: 16x Codex and What It Means

No single data point in the B2B Signals report is more striking than the Codex gap. Frontier firms send 16 times as many Codex messages per worker as typical firms. This is the largest separation in any workflow category OpenAI measured.

Why is the Codex gap so extreme? And what does it reveal about the broader enterprise AI landscape?

Codex is OpenAI's agentic coding product. Unlike chat-based coding assistance, where a developer asks a question and receives an answer, Codex accepts a task description and works autonomously across files, codebases, and tools to complete that task. The developer reviews the result and provides feedback, but the execution is delegated.

This delegation model is the key. Chat-based AI assistance augments the developer's existing workflow: they still write most of the code, but they get help with specific questions. Agentic AI delegation redesigns the workflow: the developer defines the task and evaluates the result, while the agent handles the execution.

The productivity delta between these two models is substantial. Cisco's engineering organization, cited in the B2B Signals report, uses Codex in production workflows and reports 20% faster build times, 1,500+ engineering hours saved per month, and 10-15x improvement in defect-resolution throughput. As Cisco's team described it, the biggest gains came when they treated Codex as "part of the team" rather than a tool.

The 16x gap suggests that frontier engineering organizations have crossed a threshold that typical organizations have not: they have built the infrastructure to delegate meaningful work to AI agents. This infrastructure includes:

• Clear task specifications with verifiable success criteria
• Codebase context that agents can access and understand
• CI/CD integration that validates agent outputs automatically
• Review workflows that catch errors without creating bottlenecks
• Trust mechanisms that allow incremental delegation from low-risk to high-risk tasks

Typical firms lack this infrastructure. They may have purchased Codex licenses, but their developers still use it as an enhanced autocomplete — asking for function implementations, not feature completions. The agentic capability is present but underutilized because the organizational harness around it is missing.

This pattern extends beyond coding. ChatGPT Agent, Apps in ChatGPT, Deep Research, and Custom GPTs all show similar directional gaps, though smaller in magnitude than Codex. Frontier firms are better at adopting any tool that requires delegation of multi-step tasks, application of company context, or complex research workflows. The Codex gap is simply the most extreme expression of a broader pattern: frontier firms have learned to delegate, while typical firms have not.

The implication is stark. If the 16x Codex gap reflects a genuine productivity regime difference — and Cisco's data suggests it does — then frontier engineering organizations are not just 16x more efficient at coding tasks. They are operating on a different production possibility frontier entirely. They can attempt projects that would be economically infeasible for typical organizations, iterate faster, and accumulate technical debt at a slower rate because agentic refactoring is built into their workflow.

The Task-Level Frontier: Education, Coding, and Functional Specialization

B2B Signals breaks down the frontier advantage by task type, revealing a nuanced picture of where AI adoption is deepest and where it remains shallow.

Education and learning shows the largest frontier advantage: frontier firms send 7 times as many education-related messages per worker as typical firms. This is a revealing finding. It suggests that leading firms use AI not only to complete work but to help employees build the skills, habits, and confidence needed to use AI well. They are investing in AI literacy as core infrastructure, not as a one-time training push.

This aligns with what we know about technology adoption more broadly. The electric motor took 60 years to transform factory design not because the technology was inadequate but because organizations needed time to develop the complementary skills, processes, and organizational structures. AI is following a similar trajectory. Frontier firms are accelerating through this transition by using AI itself as the training mechanism — employees learn to use AI by using AI, in a continuous feedback loop.

Coding shows the second-largest frontier advantage at 4x messages per worker, consistent with the broader pattern of advanced tool adoption. Software development and data science teams concentrate their AI usage heavily in coding tasks, reflecting the maturity of AI coding tools and the measurable productivity gains they deliver.

How-to guidance and writing and communication show the smallest frontier gaps. These are the most accessible and familiar uses of AI — asking for help with a procedure, drafting an email, summarizing a document. Most firms have already adopted these use cases, so there is less differentiation between frontier and typical organizations.

The functional specialization pattern is equally informative. IT and Security teams concentrate their queries heavily in how-to and procedural guidance. Software Development and Data Science teams show high coding usage. Finance teams use AI for analysis and calculation. The pattern suggests AI is moving beyond general productivity and into work more closely tied to each function's core responsibilities.

There is no single AI adoption leaderboard across industries. Professional, Scientific, and Technical Services ranks first in both Codex adoption and API intensity, indicating relatively advanced use in developer and product-integrated workflows. Finance and Insurance leads in broad ChatGPT adoption due to large-scale deployments. Educational Services has the highest message intensity, suggesting deeper per-person usage. Retail Trade and Health Care rank highly in API intensity despite lower rankings on other measures.

This heterogeneity matters. It means organizations have multiple entry points to frontier adoption: scale access, deepen usage, adopt agentic tools, or build AI directly into products and systems. The right path depends on the organization's industry, function, and existing technical infrastructure.

From Experimentation to Production: The API Signal

B2B Signals tracks not just ChatGPT and Codex usage but also API deployments — where enterprises integrate OpenAI models directly into their products, services, and internal systems. This is the clearest signal of AI moving from experimentation to production.

Common API use cases include in-app assistants, coding and developer tools, customer support, research workflows, and workflow automation. These are not pilots or proofs-of-concept. They are repeatable workflows with measurable operational impact.

Travelers Insurance provides a concrete example. Its AI Claim Assistant, built with OpenAI, guides customers through first notice of loss, answers policy questions, gathers the information needed to start a claim, and creates claims directly inside Travelers' systems. The assistant is expected to handle approximately 100,000 first notice of loss calls in its first year. This is not a chatbot experiment. It is a production system processing real customer transactions at scale.

The API signal is important because it reveals a different dimension of frontier adoption. Some organizations may have high ChatGPT usage but low API adoption, indicating strong individual productivity gains without systemic integration. Others may have lower overall message volume but high API intensity, indicating deep embedding of AI into core business processes. Both paths can lead to frontier status, but they require different organizational investments.

The Five Practices That Move Organizations Toward the Frontier

OpenAI's report distills five practices that appear to help firms build momentum over time. These are not theoretical recommendations. They are observed behaviors of the frontier cohort, derived from usage patterns and enterprise interviews.

1. Measure Depth of Use, Not Just Access

The relevant signal is not how many employees have AI accounts but whether teams are using AI more substantively over time. Organizations should track whether AI use is becoming more frequent, more complex, and more closely tied to valuable workflows.

This is a harder metric to collect than seat count, but it is the only metric that correlates with frontier status. Leading firms track prompt complexity, output length, delegation ratio, and the share of work products that pass through AI-assisted workflows. They treat these as leading indicators of organizational AI maturity.

2. Build Governance That Enables Production Use

Leading firms are not avoiding governance. They are using it to make agentic AI more deployable. This means clear rules for where agents can operate, what information they can use, when they should advise rather than act, and how humans review higher-risk decisions.

The key insight is that governance is not a constraint on AI adoption but an enabler of it. Without governance, organizations cannot safely delegate meaningful work to agents. They remain stuck in chat-assisted mode because the risk of autonomous action is unbounded. Frontier firms define governance standards as part of the deployment process, so governance becomes a way to expand adoption safely rather than slow it down.

3. Treat Enablement as Core Infrastructure

As AI capabilities improve, both workers and organizations need systems that help them keep pace. Frontier firms do not treat enablement as a one-time training push. They build continuous learning into deployment through role-specific training, use-case workshops, hackathons, internal champion networks, dedicated experimentation time, and shared repositories of workflows, best practices, and skills.

The 7x education gap suggests this investment pays off. Firms that use AI to help employees learn AI create a compounding effect: better AI users generate better outputs, which improves the training data and context for subsequent AI use, which makes the AI more effective, which makes users more confident, which increases adoption depth.

4. Identify Frontier Teams and Scale Their Impact

In many organizations, the most advanced usage is concentrated in a small number of teams. Those teams can reveal which workflows, habits, and operating models are working. Leaders should identify these teams, understand and scale the conditions behind their success, and help them share insights and examples of deeper AI use with the rest of the firm.

This is a bottom-up complement to the top-down governance and enablement investments. Frontier teams often emerge organically when a particular group discovers a high-value use case and iterates on it. The organizational challenge is recognizing these pockets of excellence, understanding what makes them work, and replicating those conditions elsewhere without crushing the organic experimentation that produced them.

5. Move Beyond Chat to Delegating Work

Enterprise AI is shifting from chat assistants to work that can be delegated to agents. Software engineering illustrates this trend with Codex, but delegated work is spreading across functions. The pattern is consistent: define the task, provide the context, let the agent work across files and tools, review the result, and refine the workflow with feedback.

This shift from chat to delegation is the single most important behavioral change separating frontier firms from typical firms. It is also the hardest because it requires not just new tools but new workflows, new trust mechanisms, new evaluation criteria, and new management practices. An organization cannot delegate work to agents until it knows how to evaluate agent outputs, how to catch errors before they propagate, and how to maintain human accountability for AI-driven decisions.

The Compounding Gap: Why Waiting Is Not Neutral

The 12-month delta from 2x to 3.5x is the most underrated number in the B2B Signals report. At that rate of divergence, the gap between frontier and typical firms doubles roughly every 18 months.

This is not a temporary disequilibrium that will self-correct as AI tools become more accessible. It is a compounding advantage driven by structural factors that typical firms cannot easily replicate:

Data flywheels. Each agent interaction produces structured logs that feed back into the model and workflow, improving accuracy over time. Frontier firms have been accumulating these flywheels for 12-18 months longer than typical firms. The data advantage compounds.

Institutional knowledge in agent workflows. Frontier firms have encoded their domain expertise, coding standards, review criteria, and business logic into agent-accessible formats. This is not just documentation — it is executable context that agents can use to produce higher-quality outputs. Building this context takes time and cannot be purchased from a vendor.

Team skill at working with agents. The developers, analysts, and operators at frontier firms have developed intuitions about what agents do well, what they struggle with, how to specify tasks clearly, and how to review agent outputs efficiently. This tacit knowledge is hard to transfer and takes months of hands-on practice to develop.

Infrastructure maturity. Frontier firms have built the CI/CD integrations, governance frameworks, evaluation pipelines, and monitoring systems that make agentic deployment safe and scalable. Typical firms are still designing these systems. The infrastructure gap is not just about engineering effort — it is about organizational learning that happens through iteration and failure.

The implication is that waiting is not a neutral choice. If the gap went from 2x to 3.5x in 12 months, another 12 months of inaction could push it to 5x or 6x. The organizations most likely to close the gap are those that start now, not those that wait for the technology to mature further. By the time the technology is "mature enough" for conservative adopters, frontier firms will have accumulated another year of compounding advantage.

What Existing Coverage Misses

The B2B Signals report has received substantial coverage since its release. Most analyses have focused on the headline 3.5x ratio, the 16x Codex gap, and the five practices for moving toward the frontier. These are important findings, but several deeper implications have been underexplored.

The organizational design dimension. Most coverage treats the frontier gap as a technology adoption problem: buy better tools, train more employees, deploy more agents. But the 64% depth share suggests the gap is primarily an organizational design problem. Frontier firms have restructured workflows, redefined roles, rebuilt evaluation criteria, and redesigned trust mechanisms around AI capabilities. This is not something that can be solved by a larger AI budget. It requires the same kind of organizational transformation that accompanied the shift from craft production to factory manufacturing — a transition that took decades, not quarters.

The middle management compression signal. B2B Signals does not explicitly measure this, but the task-level data hints at it. The smallest frontier gaps are in how-to guidance and writing/communication — tasks traditionally handled by middle managers who route information, coordinate timing, and reduce friction between teams. If AI is increasingly handling these coordination tasks, the economic rationale for large middle-management layers erodes. The frontier firms may be ahead not just in AI adoption but in organizational flattening.

The verification infrastructure bottleneck. The 16x Codex gap is only possible because frontier firms have built verification infrastructure that makes agentic coding safe at scale. Without automated testing, code review workflows, and CI/CD validation, delegating coding tasks to agents would generate more errors than it prevents. Most coverage celebrates the productivity gains without acknowledging the infrastructure prerequisite. This creates a dangerous expectation gap: typical firms may attempt to replicate frontier Codex usage without building the verification harness first, leading to quality failures that set back adoption.

The talent asymmetry. OpenAI's data shows that companies with over 500 engineers adopt agentic workflows at three times the rate of companies with fewer than 50 engineers. This suggests a talent and infrastructure asymmetry that may permanently divide the enterprise landscape. Small and mid-size enterprises may lack the engineering bandwidth to build and maintain agentic systems, creating a two-tier market where large enterprises compound their advantage while smaller competitors fall further behind.

The measurement problem. B2B Signals measures AI usage intensity, not business value. An organization could score high on intelligence per worker while generating minimal ROI. Conversely, an organization with lower usage intensity might be extracting more value per token because it has focused on high-impact use cases. The frontier metric is a necessary but not sufficient condition for AI success. Coverage that treats frontier status as synonymous with competitive advantage oversimplifies a more nuanced reality.

The Real Question for Enterprise Leaders

B2B Signals provides a clear answer to one question and raises a harder one.

The answered question: Is the AI adoption gap between leading enterprises and everyone else real? Yes. It is measurable, it is widening, and it is driven primarily by depth of integration rather than volume of usage.

The harder question: What does it take to close the gap?

The honest answer is that no one fully knows yet. Frontier firms have reached their status through a combination of early investment, organizational experimentation, technical infrastructure, and cultural adaptation. Some of these factors are replicable. Others may be path-dependent — the result of specific decisions made at specific moments that cannot be recreated.

What we do know is that the gap is not primarily about technology procurement. It is about organizational transformation. The enterprises that will close the gap are those that treat AI not as a productivity tool to be deployed but as a new operating model to be built. They will invest in governance before scaling, enablement before delegation, and measurement before expansion. They will accept that the transition from Driver to Architect takes time, and that attempting to skip stages produces the failed-pilot pattern that Gartner forecasts will cancel 40% of agentic AI projects by 2027.

The B2B Signals data confirms what we have seen across hundreds of enterprise deployments: the firms pulling ahead are not doing anything exotic. They made a decision to move from AI tools to AI operations, and they started doing it 12 to 18 months ago.

The question for every enterprise leader is which side of the 3.5x gap you want to be on in another 12 months.

References

• OpenAI. "How frontier firms are pulling ahead." B2B Signals, May 6, 2026. https://openai.com/index/introducing-b2b-signals/
• OpenAI. "B2B Signals Dashboard." https://openai.com/signals/b2b/
• Enterprise DNA. "OpenAI Data: Frontier Firms Now Use 3.5x More AI Per Worker." May 8, 2026. https://enterprisedna.co/resources/news/openai-b2b-signals-frontier-firms-enterprise-2026/
• IvriTech. "OpenAI B2B Signals: Frontier Firms Use 3.5x More AI." May 12, 2026. https://ivristech.com/openai-b2b-signals-frontier-firms-3-5x/
• Vectrel. "Depth Beats Volume: What OpenAI's New B2B Signals Report Reveals." May 12, 2026. https://www.vectrel.ai/blog/openai-b2b-signals-depth-volume-frontier-firms
• Beri, Rajesh. "Frontier Firms Use 3.5x More AI: Score Your Gap." THE D[AI]LY BRIEF, May 9, 2026. https://www.beri.net/article/2026-05-09-openai-b2b-signals-frontier-firms-3-5x-ai-gap-readiness-assessment
• AI Herald. "OpenAI B2B Signals: How Enterprises Build AI Advantage With Codex." May 6, 2026. https://artificialintelligenceherald.com/news/openai-b2b-signals-frontier-enterprises-ai-advantage-2026
• Gartner. "Lack of AI-Ready Data Puts AI Projects at Risk." February 2025.
• Writer. "2026 Enterprise AI Adoption Survey." April 2026.