How AI-native companies turn workflows into operating capability

This is OpenAI's own write-up of its Enterprise Signals data, so the framing serves the vendor. The number underneath it is still striking: the top ten percent of firms by AI usage generate 8.3 times the output tokens per active user of typical firms, up from 2.6 times in January. Whatever else that measures, it is a widening gap over eight months.

The three examples are meant to explain the gap operationally. At Basis, which builds agents for accounting firms, first-day onboarding takes 30 minutes instead of two hours: new hires get Codex and a company-specific onboarding skill that welcomes them, explains key concepts, and completes integration setup on their machine. At Clay, a go-to-market engineer gave every account a persistent workspace and a dedicated subagent that reviews primary sources overnight; a coordinating agent turns those updates into a morning list of priority moves, saving roughly an hour of nightly inbox triage. At Exa Labs, Codex monitors for integration opportunities, gathers context, opens pull requests, runs tests, and prepares weekly updates from sources like Slack and Notion, with humans deciding what matters and reviewing before anything ships.

Token volume is a strange proxy for capability, and it is worth naming what it actually measures: how much work is being handed to models, not how much value comes back. OpenAI has an obvious interest in a metric where more consumption equals more maturity. Read the 8.3x as evidence of delegation depth, not of results.

That said, the three cases share a structure worth stealing, and it is more specific than most enterprise AI advice. Basis demonstrated a process once and turned it into a reusable skill with a clear trigger, known steps, tool access, and a definition of done. Clay's insight is persistence — context that refreshes on a cadence and keeps the supporting evidence next to each recommendation so a seller can check primary sources before acting. Exa's is bounded execution — tests and review points that make the agent's work visible before it ships.

Signal Desk's read: the common ingredient is not the model, it is a process someone could already write down. Basis had a repeatable onboarding. Exa had a consistent sequence from discovery to implementation. Clay had a known set of accounts and sources. Agents converted stable processes into repeatable ones; they did not invent the process. The organizations getting nothing from the same tools are usually the ones with no describable workflow to hand over, and no amount of token volume fixes that.

The honest caveat is scale. Basis, Clay, and Exa are AI-native startups where the person designing the workflow is often the person doing the work. The line about permissions, evidence, and decision rights becoming a larger part of workflow design as work gets consequential is the whole enterprise problem stated in one sentence and then left there.

Enterprise Signals is OpenAI's own usage telemetry, so it describes behavior on OpenAI products rather than the market. The interesting movement is the direction of travel across eight months, from a 2.6x gap to 8.3x, which suggests the leaders are compounding rather than the laggards catching up.