OpenAI's new 'Presence' resolves 75% of support calls without a human

OpenAI is officially productizing the hard-won lesson that shipping a reliable AI agent is a systems problem, not just a model problem. The company just announced Presence, a battle-tested product designed to help enterprises deploy AI agents that don’t just chat—they actually get stuff done. The headline number is a strong one: Presence already powers OpenAI’s own English-language phone support at 1-888-GPT-0090, where it resolves 75% of inbound issues without any human assistance. Within weeks of launch, that phone line met or exceeded the company’s benchmarks for grading frontline human support quality.

The secret isn’t a smarter model in a vacuum. It’s a tight integration of model reasoning with hard-coded policies, guardrails, and escalation rules that keep the agent on a short leash. Each deployment is scoped to a specific job, like resolving billing issues or handling IT service requests, and the agent only gets access to the knowledge and systems required for that narrow task. The company sets the boundaries: what the agent can do, when it needs a human’s approval, and when a person should just take over entirely. This is a direct response to the enterprise fear of giving an LLM unchecked access to critical systems.

What’s maybe more interesting is the post-launch improvement loop. Production sessions inevitably reveal gaps as customer behavior changes. Presence uses a system called Codex to investigate those signals and propose specific updates—a new policy tweak, a revised escalation path. Teams can then test those proposed changes against the live version and approve a controlled rollout. OpenAI claims this loop reduced human handoffs for their own support line by 15 percentage points in just 10 days. Early enterprise customers are already kicking the tires, with BBVA exploring it for voice banking in Mexico, SoftBank testing natural Japanese-language conversations, and IAG looking at support during high-stress events like severe weather. It’s a clear signal that OpenAI sees the messy, unglamorous work of deployment and evaluation as a moat, not an afterthought.