Agents Move From Pilots

OpenAI is no longer selling companies a chatbot demo and calling it transformation. On April 8, Denise Dresser, OpenAI’s chief revenue officer, said customers are moving from isolated tests to company-wide deployments built around ChatGPT Enterprise, Codex, and a new agent platform called Frontier. (openai.com) Frontier is OpenAI’s attempt to become the control room for workplace artificial intelligence agents. OpenAI says the product is designed to build, deploy, and manage agents that share company context, follow permissions, and work inside existing business systems. (openai.com) That changes what “using artificial intelligence at work” means. Instead of asking one model one question, a finance team can run agents that pull data from systems of record, produce forecasts, and hand results to a human manager inside the same workflow. (openai.com) OpenAI is pitching the same setup to support and operations teams. Its Frontier materials describe end-to-end agents for customer support, revenue operations, procurement, and software engineering, which is much closer to hiring digital staff than buying a search box. (openai.com) The company is also building a services layer around that software push. In March, OpenAI introduced Frontier Alliance partners including McKinsey, Boston Consulting Group, Accenture, and Capgemini to help large companies redesign workflows and move agents from pilot projects into production. (openai.com) That timing matters because the hard part has shifted. OpenAI’s February launch post said the main bottleneck is no longer model intelligence but how agents are onboarded, given context, assigned permissions, and improved with feedback once they start doing real work. (openai.com) At the same moment the industry is pushing harder on adoption, it is getting more cautious about release speed. On April 7, Anthropic said it would give its new Mythos model only to a small set of partner organizations because of concerns about the model’s ability to find and exploit software flaws. (axios.com) Anthropic’s first users for Mythos include companies like Microsoft, Amazon, Apple, CrowdStrike, and Palo Alto Networks under a program called Project Glasswing for defensive cybersecurity work. The point of the limited rollout is to study a powerful model in a narrow setting before broader access. (cnbc.com) Put those two moves together and the pattern is clear: vendors want agents embedded deeper into payroll, legal review, customer service, and code, but they do not want every new model released to everyone on day one. One side of the business is accelerating deployment while the other side is adding gates, partners, and staged access. (openai.com) (axios.com) That leaves companies with a new job they did not have during the chatbot phase. If an agent can read internal data, call tools, and complete multi-step work across systems, a company now needs approval rules, monitoring, audit trails, and human override points built into the workflow itself. (openai.com 1) (openai.com 2) The race is no longer just about who has the smartest model. It is about who owns the layer that decides which agent can do what, in which system, with whose permission, and under what supervision. (openai.com) (futurumgroup.com)