OpenAI presented three cases of companies introducing artificial intelligence agents into everyday activities. Basis, Clay, and Exa Labs use them to support employee onboarding, organize sales accounts, and prepare software integrations.
What tasks can agents handle?
These systems do more than answer questions. Each agent receives a defined assignment, consults authorized information, and completes steps within a process. Its scope depends on the instructions, tools, and permissions defined by each organization.
The goal is to reduce repetitive work without giving the system decisions that still require human judgment.
Employee onboarding at Basis
Basis turned part of its new employee welcome process into a reusable skill for Codex. The agent introduces internal concepts and helps configure tools needed on the first day.
According to the case published by OpenAI, Basis reduced one onboarding activity from two hours to thirty minutes. The figure comes from the company and describes a particular workflow, not an automatic result for every team.
Sales priorities at Clay
Clay uses agents to gather data spread across CRM records, email, Slack, calls, and other sources. With that context kept current, the system prepares a list of priority actions for each account.
Clay estimates that this workflow avoids approximately one hour of manual inbox review each day. Staff can inspect the information supporting every recommendation before deciding whether to act.
Software integrations at Exa Labs
Exa Labs uses Codex to identify integration opportunities, gather documentation, prepare code changes, and run tests. The agent can carry an opportunity from initial research to an artifact ready for review.
The team still decides the integrations to pursue and reviews changes before publication. Codex prepares part of the work, but it does not replace technical approval or the company’s external decisions.
Human oversight remains
All three cases retain checkpoints for people. Basis allows staff to address exceptions, Clay keeps sources alongside recommendations, and Exa requires testing and review before modifications are incorporated.
The usefulness of these agents does not depend only on the AI model. It also requires a specific assignment, limited access to necessary context, and clear rules defining when the system must stop and hand the result to a person.
Results are not universal
OpenAI presents these examples as business experiences, not as the launch of an automatic feature for all users. Each organization designed its own workflow and defined the corresponding permissions.
The time savings were reported by Basis and Clay and were not accompanied by independent evaluations. They should be treated as results from specific implementations, without assuming that every company will achieve the same productivity gains.
