OpenAI introduces Frontier to manage AI agents

OpenAI has introduced Frontier, a platform that lets companies create, deploy and manage artificial intelligence agents capable of carrying out real tasks within an organization.

The proposal aims to address a problem that emerges when a company moves from testing isolated agents to using them regularly: data is spread across many applications, permissions are difficult to control and each agent often operates without knowing the context of the rest of the business.

From experiments to working agents

According to OpenAI, 75% of employees at companies say AI has helped them perform tasks they could not do before. The company also says it has seen these uses at more than one million businesses over the past few years.

OpenAI cites several business examples, although it does not identify all the companies involved:

• At one manufacturer, agents reduced production optimization work from six weeks to one day.
• At a global investment company, agents automated the entire sales process and freed up more than 90% of salespeople's time to speak with customers.
• At an energy company, they helped increase production by up to 5%, which OpenAI says represents more than $1 billion in additional revenue.

The company argues that the main obstacle is no longer just model capability, but the difficulty of turning models into reliable, controllable systems within an organization.

A shared context for the entire company

Frontier connects sources such as data warehouses, customer management systems, support tools and internal applications. The idea is to create a shared information layer so agents understand how the business works, where its data is and which decisions matter.

That can prevent a very specific problem: an agent responsible for responding to a customer might have access to the CRM, but not to the internal documents or incident history it needs to provide a correct answer.

The platform also lets agents work with files, run code and use software tools. They can operate in local environments, on a company's infrastructure or on systems hosted by OpenAI, without requiring all data to be moved to a single location.

OpenAI says Frontier uses open standards and can integrate with existing applications and agents. Users could interact with them through ChatGPT, via workflows with Atlas or within the business applications they already use.

Permissions, memory and supervised learning

Each agent has its own identity, defined permissions and operating limits. This is intended to let it work in regulated sectors or with sensitive information without the company losing control over what it does.

Frontier also includes memory, evaluations and tools for measuring performance. In practice, those responsible for the agents can review what works, identify errors and adjust the agent's behavior over time.

OpenAI gives the example of an engineering case involving millions of failed hardware tests. An agent brought together simulation logs, documents, internal processes and code to investigate the causes. The company says the time needed to identify the source of each failure fell from around four hours to a few minutes.

The first companies and availability

HP, Intuit, Oracle, State Farm, Thermo Fisher and Uber are among the first companies to adopt Frontier. BBVA, Cisco and T-Mobile have also tested the platform's approach in AI projects.

OpenAI will provide support through its deployment engineers, who will work with each company's teams and pass their findings on to the company's research teams. It has also created a partner program with companies including Abridge, Clay, Ambience, Decagon, Harvey and Sierra.

Frontier is available as of today to a limited group of customers. General availability will arrive over the coming months.

For your day-to-day life, this does not mean that every company will suddenly have autonomous artificial employees tomorrow. It means OpenAI is trying to turn agents from one-off demonstrations into tools connected to an organization's real systems, with permissions, oversight and measurable goals. The important question will be whether companies can deploy them safely and whether the reported results hold up beyond the cases selected by the company.