AI in Business: 8.3× More Outputs for the Top 10%

A recent assessment from OpenAI describes a shift from assistance to execution, with a widening gap between pioneers and followers. Three concrete cases—Basis, Clay, and Exa Labs—demonstrate how agents are becoming operational capabilities, with human checkpoints and precise metrics. Steps for experimentation and model transfer are proposed to industrialize what works.

Measuring, Governing, and Replicating AI-Enhanced Workflows

Scaling involves visible and measurable trials, followed by converting the most promising ones into reusable practices. Recommendations include selecting a strategic and frequent workflow, defining an outcome, a responsible party, and indicators, and drafting a precise job description for the agent, including triggers, evidence to produce, and checkpoints for human review. The construction of the human system around the agent is detailed: responsibilities for business outcomes, business logic, access and controls, adoption, and daily use, with formalized decision-making rights as the scope expands. Experimentation must be captured and packaged into skills, plugins, or shared spaces, relying on complementary modes (Chat for quick exchanges, Work for multi-step tasks, and Codex for technical execution). Six months after adoption, entry-level employees sent 13 more messages per week than executives, according to OpenAI, reflecting a dynamic of ownership. Testing, review points, and management of permissions, evidence, and decision rights make the agents' work visible and auditable, facilitating adjustments before each new execution. Finally, transferring the operational model requires preserving context, controls, metrics, and proven roles to apply them to a new value surface.

Accelerated Onboarding at Basis, Without Sacrificing Consistency

At Basis, the onboarding of new hires has been reduced to 30 minutes instead of two hours. Human resources can reallocate time to culture and support, while employees gain access to Codex and a dedicated onboarding skill from day one. This skill consolidates the necessary instructions and resources; Codex welcomes the employee, presents key concepts, and finalizes background configuration. HR updates the skill based on recurring questions or exceptions, making the process less dependent on individual availability. Formalized with triggers, steps, tools, and a definition of completion, the journey becomes repeatable and improvable, while providing a concrete model of collaboration with AI from the outset.

At Clay, a Sales Context Maintained by Sub-Agents

Clay faces a fragmented sales context across CRM, messaging, and multiple exchanges. A go-to-market engineer has established a persistent workspace and a sub-agent for each account, which aggregates key sources every night and updates a business file. In the morning, a coordinating agent converts these updates into a short list of priorities, such as responding to a customer question or following up with a prospect. According to the company, this setup saves about an hour of nightly email sorting and facilitates tracking micro-actions over long sales cycles. Recommendations come with consultable evidence, and the approach could be extended to other sales roles, subject to account permissions. Clay summarizes the needs of scalable work: stable structure, regular refresh, shared evidence, and human judgment at the moment of action.

At Exa Labs, Opportunities Converted into Tested Artifacts

Exa Labs, aiming for "Exa everywhere" with its search API, has formalized a path previously managed by DevRel and account teams. The transition from discovery to implementation has been structured within a Codex workflow, integrating explicit priorities, access to appropriate sources, and human validation before each submission. Codex now tracks integrations deemed priority, collects context, creates pull requests, conducts tests, and generates weekly reports from Slack and Notion; if necessary, it anticipates drafting the next step, up to the first announcement, for review. This setup reduces transfers between functions and transforms a signal into a tested artifact. The choices of opportunities, commitments, and external relationships remain human, while testing and reviews make the agent's work visible and adjustable. As the stakes grow, permissions, evidence, and decision rights weigh even more heavily in the design of the workflow.

A Widening Gap Between Pioneers and Followers

Organizations in the top 10% generate 8.3 times more output tokens per active user than the average, compared to 2.6 times in January, according to OpenAI. This dynamic reflects a shift from assistance to execution, where leaders connect agents, context, and tools, delegate more broadly, and replicate winning paths. The challenge lies in making this work reliable, measurable, and improvable, while maintaining a testing space, even for cases where value is not immediately apparent. The implementations at Basis, Clay, and Exa illustrate a common trajectory: teaching a stable process to the agent, maintaining a persistent context, and then converting signals into proven actions. They show that it is possible to integrate agents at the core of existing activities and improve workflows over iterations.