OpenAI: frontier firms now generate 8.3x more AI output
OpenAI's Enterprise Signals series reports that enterprise AI use is splitting into two speeds. Frontier firms, defined as the top 10% of AI usage, now generate 8.3 times as many output tokens per active user as typical firms, up from 2.6 times in January. OpenAI attributes the widening gap not to more usage alone but to a different way of working: leading firms connect agents to company context and tools, delegate more substantive work to them, and turn workflows that succeed once into processes that repeat.
The piece illustrates that shift with three startups. At Basis, which builds AI agents for accounting firms, first-day employee onboarding now takes 30 minutes instead of two hours. New hires get immediate access to Codex and a company-specific onboarding skill: a reusable set of instructions and resources for that workflow. Codex introduces key company concepts and handles integration setup in the background, while HR updates the skill whenever recurring questions or exceptions come up, rather than repeating the walkthrough by hand each time.
Clay, which builds a self-learning revenue engine for go-to-market teams, tackled a different problem: deal context scattered across CRM records, email, Slack, calls, presentations, and conversations with internal teams and customer champions. One of Clay's GTM engineers set up a persistent workspace with a dedicated subagent for every account; each subagent reviews primary sources and updates its account's deal folder overnight. A coordinating agent then turns those updates into a short list of priority actions each morning, such as answering a lingering customer question or filling a gap in the buying committee. According to Clay, the workflow saves that engineer roughly an hour of inbox triage every night, and the supporting evidence stays attached to each recommendation so sellers can check the primary sources before acting. OpenAI suggests enterprises could extend this kind of shared context to account executives, business development reps, solutions engineers, and sales leaders, within existing account permissions.
Exa Labs, which builds web search infrastructure for AI agents, wanted its search API integrated wherever developers might use it, a goal the team calls "Exa everywhere." It turned the previously manual process, developer relations staff monitoring repositories and the wider ecosystem for promising integrations, into a defined Codex workflow. Codex now watches for high-priority integration opportunities, gathers context, opens pull requests, runs tests, and prepares weekly updates drawing on sources like Slack and Notion; it can also draft materials such as an initial announcement for the team to review. People still decide which opportunities matter and which commitments Exa makes, and every change goes through tests and human review before it ships.
OpenAI draws a general playbook from the three cases: pick one consequential, frequently repeated workflow; define the outcome, the accountable owner, and the metrics to track (cycle time, quality, cost, revenue, or risk, not just output volume); write out the agent's job description, including required context, tools, permissions, and where it must stop for human review; build the human system of ownership and decision rights around the agent; and capture what works as reusable skills, Plugins, or shared workspaces, distributed across OpenAI's Chat, Work, and Codex products for different kinds of tasks. The piece also cites separate OpenAI research finding that six months after adoption, early-career employees sent 13 more weekly messages to AI tools than executives did, framed as evidence that giving employees room to experiment early pays off.
Key facts
- Frontier firms, the top 10% by AI usage, now generate 8.3x as many output tokens per active user as typical firms, up from 2.6x in January, per OpenAI's Enterprise Signals.
- At Basis, an AI-agent startup for accounting firms, Codex-driven onboarding cut first-day setup from two hours to 30 minutes using a reusable, HR-editable onboarding skill.
- Clay gives every sales account a persistent subagent that reviews CRM, email, Slack, and call data overnight; the company says this saves one GTM engineer roughly an hour of inbox triage each night.
- Exa Labs runs its 'Exa everywhere' integration push through Codex, which monitors for opportunities, drafts pull requests, runs tests, and prepares weekly updates, with people still deciding what ships.
- Six months after AI adoption, early-career employees sent 13 more messages per week to AI tools than executives did, according to OpenAI research cited in the piece.
Why it matters
The headline number, an 8.3x output gap that has more than tripled from 2.6x since January, is OpenAI's evidence that enterprise AI adoption is bifurcating rather than converging. The firms pulling ahead are not simply using AI more; they are changing how work gets structured: agents get persistent access to company context and tools, get handed more substantive tasks, and successful one-off workflows get turned into repeatable processes. That framing sets up the three case studies as a how-to for closing the gap rather than just a scoreboard.
Who it affects
The direct subjects are three OpenAI-ecosystem startups and their own customers: Basis builds agents for accounting firms and used one on its own HR onboarding; Clay builds a go-to-market platform and applied the pattern to enterprise sales reps managing multiple accounts; Exa Labs builds search infrastructure for AI agents and applied it to its own developer relations pipeline. The intended audience is broader: enterprise leaders deciding whether and how to hand agents ownership of a workflow rather than isolated tasks.
How to use it
OpenAI lays out a five-step approach drawn from the cases: choose one consequential, frequently repeated workflow rather than spreading effort thin; name an accountable owner and define success with concrete metrics (cycle time, quality, cost, revenue, or risk, not raw output volume); write the agent's job description, specifying triggers, required context, tools, permissions, and the points where it must stop for human review; build the surrounding human system of ownership and decision rights, since startups can compress these roles into a few people but larger enterprises need them made explicit; and capture what works as a reusable skill, Plugin, or shared workspace, using OpenAI's Chat for quick questions, Work for multi-step deliverables, and Codex for technical execution.
How solid is it
This is an OpenAI-published case study built around three of its own ecosystem partners, all Codex users, so it should be read as a success narrative rather than independent research. The 8.3x and 2.6x figures come from OpenAI's own Enterprise Signals data with no disclosed sample size, firm count, or methodology. The company-level numbers are self-reported: Basis's onboarding-time drop and Clay's roughly-an-hour nightly saving are both attributed to the companies themselves, with Clay's figure explicitly qualified 'according to Clay.' No dollar figures, revenue, or customer counts are given for any of the three startups, and neither the current measurement date nor the year of the cited January baseline is stated.
Risks and caveats
Every case in the piece keeps a human in the loop by design, not as an afterthought: HR handles onboarding exceptions at Basis, sellers check primary sources before acting on Clay's recommendations, and Exa Labs requires human review before any Codex-generated pull request or announcement ships. That framing cuts against reading the piece as evidence of full automation. The three companies are also OpenAI customers being showcased by OpenAI, which is a promotional incentive worth keeping in mind when weighing the reported time savings and the size of the frontier-versus-typical-firm gap.
“Exa everywhere”
— Exa Labs, describing its goal of getting its search API integrated wherever developers might use it