OpenAI outlines agent workflows at Basis, Clay and Exa Labs
OpenAI maps a path from AI assistance to repeatable execution
OpenAI has set out how AI-native companies are embedding agents in operational workflows, citing Basis, Clay and Exa Labs as examples. Its latest Enterprise Signals report finds that frontier firms—the top 10% by AI usage—generate 8.3 times as many output tokens per active user as typical firms, compared with 2.6 times in January.
The company argues that the widening gap reflects more than greater use of AI. Leading organisations connect agents to company context and tools, delegate substantive work and make successful processes repeatable. The practical challenge for enterprise leaders is to create work that can be trusted, measured and improved.
Across the three examples, OpenAI identifies a progression: teach an agent a stable process, maintain persistent context as work changes, and enable it to take an opportunity into tested action. Human involvement remains central where exceptions, commitments and consequential decisions arise.
Basis turns onboarding into a reusable skill
Basis, which builds AI agents for accounting firms, uses Codex and a company-specific onboarding skill to support new employees on their first day. The company says first-day onboarding now takes 30 minutes rather than two hours.
The skill welcomes employees, introduces company concepts and completes integration setup on their computer in the background. Basis demonstrated the process once and converted it into a reusable workflow with a trigger, defined steps, appropriate tool access and a clear completion condition.
Recurring questions and exceptions can be used by HR to update the skill for the next intake. The process no longer depends on a single person being available, while staff can still intervene for complex issues.
Clay and Exa apply context, tools and review
Clay uses a persistent workspace and a dedicated subagent for each account to address deal information scattered across CRM records, email, Slack, calls and other conversations. The subagent refreshes each deal folder overnight from primary sources, while a coordinating agent produces a morning list of priority actions.
Clay says this approach saves one GTM engineer roughly an hour of inbox triage every evening. Recommendations retain the supporting evidence so sellers can inspect the underlying sources before acting, with account access remaining subject to existing permissions.
Exa Labs, a web-search infrastructure provider for AI agents, formalised its developer integration process for Codex. Codex monitors high-priority opportunities, gathers context, creates pull requests, runs tests and prepares weekly updates using sources including Slack and Notion. It can also draft an initial announcement for team review when appropriate.
Exa retains human decisions over opportunity selection, commitments and external relationship management. Tests and review points expose the agent's work before anything is shipped and provide evidence for refining the workflow.
Start with a measurable workflow
OpenAI recommends selecting a consequential end-to-end workflow with clear systems, handoffs, controls and measurable stakes. Teams should define an accountable owner, KPI, baseline and guardrails, then assess outcomes such as cycle time, quality, cost, revenue or risk rather than output volume alone.
A practical implementation should specify the agent's trigger, outcome, context, tools, permissions, evidence requirements and points for human review. Businesses can then preserve the context, evaluations, decision rights and measures that work, using each proven workflow as a foundation for the next one.