OpenAI Says AI-Native Companies Are Turning Workflows Into Operating Capability
OpenAI highlights how Basis, Clay, and Exa Labs use AI agents to automate repeatable work while preserving context, testing, and human oversight.
OpenAI is pointing to a new phase of enterprise AI adoption: companies are moving beyond assistants that answer questions and toward AI agents that execute repeatable workflows. In a new OpenAI News report, the company profiles Basis, Clay, and Exa Labs as examples of organizations embedding agents into onboarding, account management, and developer integrations.
The report argues that the important shift is not simply higher AI usage. It is the conversion of successful experiments into repeatable operating processes with defined triggers, access to company context, connected tools, measurable outcomes, and explicit human review. That distinction matters to enterprise buyers deciding whether AI should remain an individual productivity tool or become part of how work is run.
From usage to operating capability
OpenAI’s accompanying Enterprise Signals analysis says companies in the top 10% of AI usage now generate 8.3 times as many output tokens per active user as typical firms, compared with 2.6 times in January. OpenAI presents the widening gap as evidence that leading organizations are connecting agents to internal context and tools, delegating more substantial work, and making effective workflows easier to repeat.
Those figures are vendor-reported and do not, on their own, show that higher token output produces better business results. OpenAI’s own guidance acknowledges that organizations need to measure completed tasks, quality, cost, cycle time, risk, exceptions, and review effort rather than treating activity volume as a proxy for value.
The practical model is closer to a job description for software. A team defines what starts the work, what information and permissions the agent needs, what “done” means, what evidence it must produce, and where a person must take over.
Three workflows, three levels of delegation
At Basis, which builds AI agents for accounting firms, OpenAI says employee onboarding has been reduced from two hours to 30 minutes. New employees receive access to Codex and a company-specific onboarding skill, described as a reusable set of instructions and resources for a particular workflow. The agent introduces company concepts and completes integration setup, while HR updates the skill when new exceptions appear.
The example shows how a one-time demonstration can become a reusable process. The workflow has a trigger, known steps, tool access, and a completion condition. Human staff remain available for unusual or sensitive cases, but onboarding no longer depends entirely on one person’s availability.
Clay applies a different pattern to sales. According to OpenAI, the company uses a persistent workspace and a dedicated subagent for each account. These subagents review primary sources and update deal folders overnight. A coordinating agent then produces daily priorities, such as answering a customer question or identifying a missing member of the buying committee.
Clay says the process saves one go-to-market engineer roughly an hour of nightly inbox triage. That is a company claim, not an independently verified productivity study. Its operational significance lies in the supporting evidence: sellers can inspect the source material behind a recommendation before acting, while account context remains available to other authorized teams.
Exa Labs uses agents to pursue its “Exa everywhere” objective for its search API. OpenAI says Codex monitors for potential integrations, gathers context from systems including Slack and Notion, creates pull requests, runs tests, and prepares weekly updates. It can also draft announcements, but people decide which opportunities matter and what the company should commit to externally.
This workflow goes further toward execution than the Basis example, but it remains bounded by tests and review. The agent can carry an opportunity from discovery to a tested artifact without making the final business or relationship decision.
What the evidence does—and does not—show
The strongest evidence in the report is descriptive: OpenAI documents how three startups have structured particular workflows. The performance figures come from the companies themselves or from OpenAI’s analysis, and the source does not provide independent validation, comparable baselines across businesses, or evidence that the same results will transfer to larger enterprises.
Still, the cases identify recurring design choices. Basis formalizes a stable process as a reusable skill. Clay gives an agent persistent context and a refresh cadence for work that changes over time. Exa connects signals to tools, tests, and review before anything reaches production or an external audience.
OpenAI also cites research showing that, six months after adoption, early-career employees sent 13 more messages per week than executives. The company uses that finding to argue that employees closest to daily work should have room to test new applications. That conclusion is plausible, but experimentation will only create durable value when teams capture the underlying process, evidence, controls, and ownership.
What this means for builders and enterprises
For AI builders, the report reinforces that the difficult product problem is not only model capability. It is workflow design. Agents need reliable access to the right sources, permission boundaries, evaluation methods, exception handling, and a clear handoff to people. Without those elements, automation can increase activity while making accountability harder to locate.
For enterprises, the examples suggest starting with one consequential workflow rather than attempting a broad automation program. The candidate should recur often enough to generate feedback and matter enough to justify redesign. Leaders should name an accountable owner, establish a baseline, and track both business outcomes and the operational burden of reviewing agent work.
The cases also show why deployment should expand gradually. An agent that summarizes account activity may later recommend actions, prepare changes, or open a pull request. Each step increases the importance of permissions, evidence, testing, and decision rights. The fact that an agent can perform an action does not establish that it should perform it without approval.
What to watch next
The next signals will be whether OpenAI or the featured companies publish independent outcome data beyond token usage and time-saved claims. Enterprise buyers should also watch for clearer evaluation methods covering error rates, exception frequency, review load, and the cost of maintaining agent context.
Product teams should examine whether Codex and similar tools gain stronger controls for permissions, audit trails, reusable skills, and cross-system testing. More examples of agents moving from recommendations to bounded execution would indicate that the pattern is becoming operational rather than promotional.
Creati.ai perspective
OpenAI’s report is most useful when read as a workflow-design case study, not as proof that AI agents are delivering uniform enterprise gains. Basis, Clay, and Exa Labs are showing different versions of the same operating principle: automation becomes more valuable when context, tools, evidence, and human judgment are designed together.
The competitive question for AI-native companies will be how quickly they can turn successful agent experiments into maintainable systems without losing control. The teams that document ownership, permissions, evaluation, and review points will be better positioned to scale execution than those that measure adoption only by how much employees ask models to do.
