On May 6, OpenAI published "How Frontier Enterprises Are Building an AI Advantage" — its first systematic B2B Signals report measuring AI adoption depth across enterprise customers. The headline finding: frontier firms are not just using more AI than typical firms. They are using qualitatively different AI, in qualitatively different ways, and the gap is measurable and growing.

OpenAI has 1 million enterprise customers. Until yesterday, it had never published a systematic comparison of how the top-performing organizations use its tools versus the average. The B2B Signals report is that comparison — and its most important finding is not a single number but a pattern. Frontier firms lead not just in overall usage volume but specifically in the tools that require the most organizational discipline to deploy: Codex, ChatGPT Agent, Deep Research, and GPTs.

The 16x Codex gap is the most striking single statistic. Frontier firms send sixteen times as many Codex messages per worker as typical firms. ChatGPT Agent, Apps in ChatGPT, and Deep Research show similar directional patterns — all substantially higher at frontier firms. OpenAI's analysis of what separates these organizations is precise: frontier firms are better at adopting tools that help workers code, delegate multi-step tasks, apply company context, and conduct complex research. They have moved from "using AI for help on tasks" to "managing teams of agents to do tasks for them."

The report includes specific production deployments. Travelers Insurance built an AI Claim Assistant that guides customers through first notice of loss, answers policy questions, gathers claim information, and creates claims directly inside Travelers' systems. Travelers expects it to handle approximately 100,000 first notice of loss calls in its first year. Cisco uses Codex to speed up complex software work across a large enterprise engineering organization.

The report's most counterintuitive finding is about industry leadership. There is no single AI adoption leaderboard. Some industries lead in broad ChatGPT adoption, others in Codex use, others in API intensity. The relevant question is not "how much AI are we using" but "which AI tools require the most organizational discipline to deploy, and how do we compare to frontier firms on those specific tools?"

OpenAI explicitly frames the adoption gap as closable: "The gap between frontier firms and typical firms should not be read as a fixed divide." The organizations that use the next eighteen months to close the discipline gap will establish advantages that compound. The ones that don't will find the gap harder to close as frontier firms' institutional knowledge deepens.

Sources: OpenAI B2B Signals Report, May 6, 2026 / OpenAI Enterprise Phase report

reduction in hallucinated claims on high-stakes prompts in medicine, law, and finance — GPT-5.5 Instant vs. GPT-5.3 Instant.

This is the number that matters for enterprise procurement. A 52.5% reduction in hallucinated claims in the specific domains where AI errors generate the most liability is a concrete operational improvement with direct implications for how organizations govern AI-assisted work. Combined with a 37.3% reduction in inaccurate claims on challenging conversations users flagged for factual errors, GPT-5.5 Instant represents the first ChatGPT default model that enterprise risk management teams can point to specific, quantified reliability gains in regulated domains. For organizations that have been waiting for AI reliability to reach a threshold sufficient for high-stakes deployment, this week's release is a meaningful signal.

GPT-5.5 Instant replaced GPT-5.3 Instant as the default ChatGPT model for all users on May 5, 2026. Beyond the hallucination reduction: the personalization layer now uses context from past conversations, uploaded files, and connected Gmail. The new "memory sources" feature shows users which specific past chats or saved memories shaped a response. Output style is more compressed: 30.2% fewer words and 29.2% fewer lines on the same prompt. For organizations that have built workflows around GPT-5.3 Instant's output format — particularly prompt templates and downstream parsing logic — the format change deserves a verification pass before the three-month retirement window closes. API access is via "chat-latest"; GPT-5.3 remains available to paid users for three months through model configuration settings.

Sources: OpenAI official release / TechCrunch, May 5 / Axios, May 5

OpenAI's B2B Signals report identifies a qualitative shift in how the most advanced enterprise AI users operate: "The people who are furthest ahead have gone from using AI for help on tasks, to managing teams of agents to do tasks for them." Frontier firms deploy Codex to write, review, and deploy code autonomously across large engineering organizations. They use ChatGPT Agent to delegate multi-step research and analysis. They apply company context — internal documents, institutional knowledge, customer data — to AI workflows in ways that make output genuinely differentiated from generic model output. The operational implication: the AI "assistance" frame is how typical firms think about the technology. The AI "delegation" frame is how frontier firms operate it.

Source: OpenAI B2B Signals Report, May 6, 2026

Travelers Insurance's AI Claim Assistant handles first notice of loss calls end-to-end — guiding customers through intake, answering policy questions, gathering claim information, and creating claims directly inside Travelers' systems — with human oversight at exception points rather than at every step. Travelers expects 100,000 first notice of loss calls handled in its first year. The deployment template: a high-volume, structured, repetitive workflow with clear intake-to-output flow, in a regulated industry with explicit audit requirements. This is the architecture that makes enterprise AI deployment viable in regulated contexts — not full automation, but full workflow coverage with human escalation at exception points.

Source: OpenAI B2B Signals Report, May 6, 2026

OpenAI released a dedicated version of the ChatGPT iOS app for Microsoft Intune on May 4, resolving a persistent friction point for enterprise IT teams that manage iPhones and iPads through Intune. The Intune-specific version supports iOS 17 and above and allows organizations to deploy ChatGPT under the same MDM policies governing every other managed application. For organizations in financial services, healthcare, and government that operate under strict MDM requirements, the Intune integration removes the last common objection to enterprise-wide ChatGPT deployment on managed devices.

Source: VoIP Review, May 6, 2026

The OpenAI B2B Signals report documents a genuine competitive advantage at frontier firms. It also documents something that deserves equal scrutiny: frontier firms have built workflows, operating muscle, and institutional habits around OpenAI's rapidly changing model stack. That depth of integration is the source of their advantage. It is also a source of risk.

GPT-5.3 Instant, the default model until this week, is now on a three-month deprecation clock. GPT-4o was deprecated in February 2026 despite genuine user attachment. The observable cadence is roughly one major model rotation every two months, with each rotation changing response format, personality, verbosity, and reliability characteristics.

Organizations that have built parsing logic, prompt templates, or downstream workflows around a specific model's output format face a recurring disruption cycle that is difficult to plan around. The frontier firms generating 16x Codex usage have enough engineering capacity to absorb these transitions. The organizations in the catch-up cohort — mid-market enterprises building their first serious AI workflows — may find that the model they built around no longer exists by the time they reach production scale.

The enterprise risk of deep AI platform dependency is not adequately captured in the B2B Signals report's framing of the adoption gap as simply a question of organizational discipline to close. It is also a question of whether the speed of model iteration at AI labs is compatible with the governance and change management requirements of regulated enterprise deployment.

Sources: OpenAI B2B Signals Report / TechCrunch deprecation timeline analysis

The OpenAI B2B Signals report's most important insight is buried in its language rather than its numbers. The organizations generating the most value from AI have made a specific organizational shift: they have moved from "using AI for help on tasks" to "managing teams of agents to do tasks for them." That is not a technology change. It is a management philosophy change.

The Travelers Insurance case makes the distinction concrete. The AI Claim Assistant does not assist human agents with first notice of loss calls. It handles first notice of loss calls — all the way through claim creation inside Travelers' systems — with human oversight at exception points. The design decision that made this possible is the same decision most organizations have not yet made: defining, precisely, which parts of the workflow AI owns end-to-end and which parts require human judgment.

This week's governance story connects to this in a specific way. The ServiceNow Knowledge 2026 announcements from yesterday — AI Control Tower, Action Fabric, Autonomous Workforce — are the infrastructure layer that makes the "managing teams of agents" operating model viable at enterprise scale. You cannot delegate to teams of agents without a control plane that can see what the agents are doing, enforce permission boundaries, and produce audit trails. The organizations that combine the Travelers workflow design philosophy with the ServiceNow governance infrastructure are building the complete architecture for compounding AI advantage.

The GPT-5.5 Instant default rollout is the mechanical expression of the same underlying trend. The model running under every ChatGPT user became substantially more reliable this week in the specific domains where enterprises have been most cautious. The question is whether the organizations around them are building the structures — workflow design, governance infrastructure, trust frameworks — that allow that improved reliability to produce compounding business value. Or whether the default upgrade is happening beneath a set of organizational practices that are still treating AI as an experiment rather than a reliable operating component.

The default is now AI. The question that determines whether your organization leads or follows is not whether you're using AI. It is whether you're designing for it.

AK / Spearhead / Building AI systems, not tools