OpenAI’s September 1 enterprise note reads less like a product drop and more like three desk photos pinned to a wall. Basis cut accounting-firm onboarding from about two hours to thirty minutes; Clay gives every account an overnight-updating subagent; Exa folds “spot an integration → gather context → open a PR → run tests” into a Codex workflow. The hard number: firms in the top 10% of AI usage now generate 8.3× as many output tokens per active user as typical firms, up from 2.6× in January. The gap is not who chats more — it is who wired agents into company context and tools.
[1]Per the official write-up, the three paths map to three difficulties. Basis demonstrates onboarding once, then packages a skill with trigger, steps, tool access, and a definition of done; new hires get Codex plus a company skill on day one while integrations finish in the background and exceptions feed back into the skill. Clay faces deal context scattered across CRM, email, Slack, and calls, so each account gets a persistent workspace and subagent that refreshes the deal folder overnight; a coordinating agent turns that into a short, inspectable priority list — about an hour of inbox triage saved each night, Clay says. Exa’s “Exa everywhere” goal becomes a defined Codex path: monitor high-priority integrations, gather context, open PRs, run tests, draft weekly updates; humans still decide which relationships to commit to.
[1]None of this is a model IQ contest. It is a contest over whether work is teachable, testable, and handoff-ready. Basis standardizes a one-shot process; Clay keeps evolving context fresh overnight; Exa cuts research–engineering–comms handoffs. OpenAI’s six-step checklist — pick a consequential end-to-end workflow, name KPIs and guardrails, specify triggers/context/stop points, put frontline people in the design loop, package wins as skills/Plugins, then reuse the boundary — is the real product. Author’s take: enterprises do not need another chat box; they need a way to turn “it worked once” into something that still runs tomorrow.
[1]For buyers, the fights are practical. Token-volume gaps can become vanity KPIs unless paired with cycle time, quality, revenue, or risk — OpenAI says as much. Clay-style account subagents immediately raise permission and customer-data questions. Exa-style PR-and-test agents only work if review standards live inside the workflow, or “visible tests” become rubber stamps. Whether regulated giants can copy overnight account agents remains open; the case studies stop at startups with compressed ownership.
[1]Treat the piece as a scorecard. If your agent demo still ends at “draft an email” while peers refresh deal folders and open tested PRs, the 8.3× gap is an operating fact, not marketing. Official claims cover three startup cases plus Enterprise Signals stats; larger regulated rollouts are not proven here. Bottom line: agent-era competitiveness starts looking like whether a workflow can improve itself overnight.
[1]