Netomi’s lessons for scaling agentic systems into the enterprise

Netomi builds AI agents that handle customer service for large companies, and OpenAI just published a case study on how the company keeps those agents from falling apart under real-world load. The headline detail is that Netomi runs two different OpenAI models in tandem: GPT-4.1 for faster, high-concurrency tasks, and GPT-5.2 for the harder multi-step reasoning that customer issues often require.

That split matters more than it sounds. Enterprise support isn't one job, it's thousands of small jobs happening at once, some trivial, some requiring the agent to chain together several decisions before it gives an answer. Netomi's approach treats concurrency and reasoning depth as separate problems, routing simpler high-volume requests to the leaner model while saving heavier reasoning for cases that actually need it. It's a pragmatic answer to a cost and latency problem that a lot of companies deploying agents at scale are quietly wrestling with.

Governance is the other piece OpenAI highlights. Netomi didn't just wire up a model and let it loose on customer chats. The company built structure around how agents make decisions, presumably guardrails, escalation paths, and checks that keep the system from going off script when it hits an edge case. That's the unglamorous part of agentic AI that rarely makes headlines but determines whether a deployment survives contact with actual customers.

What's notable here isn't a new model or a flashy demo. It's OpenAI using an enterprise partner to show, in plain terms, that scaling agents isn't about picking the smartest model and shipping it. It's about architecture: knowing which tasks need raw reasoning power, which need speed, and building enough oversight that the whole thing doesn't quietly drift into bad answers nobody notices until a customer complains.