OpenAI highlights three patterns for turning AI agents into company workflows

OpenAI’s September 1, 2026 report, “How AI-native companies turn workflows into operating capability”, argues that the most advanced enterprise AI users are moving beyond chat-based assistance. The top 10% of companies by AI usage now produce 8.3× more output tokens per active user than typical companies, up from 2.6× in January.

The report points to Basis, Clay, and Exa Labs as examples of that shift. Their implementations differ, but each connects an agent to company context, repeatable processes, and the tools needed to complete bounded work.

Three operating patterns

Basis uses Codex to guide new employees through company-specific onboarding and background integration setup. The process takes about 30 minutes, compared with two hours previously. HR can update the reusable onboarding skill when recurring questions or exceptions emerge.

Clay gives every account a persistent workspace and subagent. These agents review CRM, email, Slack, call, and presentation data overnight, while a coordinating agent produces daily sales priorities with supporting evidence. Clay says the workflow saves roughly an hour of nightly inbox triage.

Exa Labs uses Codex to monitor for promising developer integrations, gather context, create pull requests, run tests, and prepare weekly updates. People still decide which opportunities to pursue, and human review remains part of the release process.

Why it matters for builders

The common lesson is to start with one consequential, repeatable workflow and define its owner, baseline, KPI, permissions, evidence requirements, and review points. Agents become more useful when they have persistent context and clear triggers—but autonomy should expand only as testing and oversight show the process is reliable.

For teams building AI tools, the opportunity is not simply generating more output. It is creating workflows that preserve evidence, expose exceptions, and improve with each run.

Source: OpenAI News