OpenAI’s contact center once had the same problem every enterprise support line has: too many routine calls, not enough context sharing between systems, and agents who need months of training before they can resolve anything on their own. In July 2026, OpenAI turned its answer to that problem into a product. Presence is a managed platform for deploying voice and chat agents that connect to a company’s own systems, follow defined policies, and improve after launch instead of staying frozen at whatever quality they shipped with. It is already running OpenAI’s own English-language phone support line, where it resolves 75% of inbound calls without a human. For engineering teams evaluating whether to build agent integrations in-house or adopt a managed platform, the interesting part of Presence is the architecture behind that number.
What Presence actually packages
Presence is not a single model call wrapped in a chat widget. It bundles the pieces a team would otherwise assemble separately: policies and standard operating procedures, guardrails, approved actions, simulations, evaluation tooling, and a Codex-powered improvement process. Each deployment is scoped to one job, such as billing questions, insurance claims, or IT service requests, and the agent gets only the data and system access that job requires, according to VentureBeat’s coverage of the launch. That scoping decision matters more than it sounds. The hard engineering work isn’t the voice model itself; it’s defining the boundaries around what the agent can query, execute, or must route to a person.
Before any deployment goes live, it runs through simulations and graders that test whether the agent reaches correct outcomes, follows policy, calls tools correctly, and escalates when it should. The same VentureBeat report describes this as testing against “common scenarios, edge cases, and high-risk situations.” After launch, the loop doesn’t stop: Codex reviews production sessions and escalations, proposes behavioral changes, and OpenAI staff test and approve them before they ship. Help Net Security reports that this loop cut human handoffs by 15 percentage points in just 10 days on OpenAI’s own line.
The voice model underneath: Realtime-2
Presence sits on top of OpenAI’s Realtime API, which was split into three specialized models rather than one general-purpose voice model: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. GPT-Realtime-2 is the one built for agent conversations. It brings GPT-5-class reasoning into live audio for the first time, scoring 15.2% higher than GPT-Realtime-1.5 on the Big Bench Audio benchmark, per Enterprise DNA. It can call several tools in a single turn and narrate what it’s doing mid-call (“checking your account now”) so the caller isn’t left with dead air during a lookup.
This separation is useful context for teams building their own agent stack instead of waiting on Presence’s white-glove rollout. The reasoning, translation, and transcription layers are available independently through the Realtime API, so the same architectural pattern Presence uses can be built directly by any team with the engineering discipline for it, on infrastructure they control.
Why it isn’t self-serve — and what that signals
Presence launched through a limited general-availability program, not as a self-service product. Deployments are led by OpenAI’s Forward Deployed Engineers or select systems integrators, and there is no published pricing. CIO reports early adopters include BBVA testing customer support in Mexico, SoftBank running Japanese-language voice agents, and Australian insurer IAG using it to manage claims surges during severe weather. That deployment model, hands-on integration work rather than an API key and a dashboard, is itself the signal. Connecting an agent to real account systems, claims databases, or ticketing platforms safely is still bespoke integration work, and OpenAI is choosing to do that work directly rather than ship a toolkit and hope teams get the guardrails right on their own.
The gap between OpenAI’s number and everyone else’s
The 75% resolution figure is self-reported and comes from OpenAI’s own support line, which is a best-case environment: clean internal systems, a team that built the model, and months of tuning. CIO’s reporting is direct about the caveat: most organizations should expect a lower starting automation rate that improves over time as the agent gets refined, and legacy system fragmentation plus compliance requirements will keep typical enterprise results below OpenAI’s internal benchmark.
That gap tracks with a broader pattern in enterprise AI adoption. Databricks’ research found that companies with formal AI governance programs pushed 12 times more agent projects into production than those without one, and teams using structured evaluation tooling moved nearly 6 times more systems to production. Yet 40% of organizations surveyed said their own governance, meaning data definitions, guardrails, and accountability, is still insufficient. That points to a constraint on shipping enterprise voice agents that has little to do with a model’s conversational quality: it comes down to whether a team has the policy definitions, evaluation harness, and audit trail to trust the agent with real account actions.
Conclusion
Presence is a preview of where enterprise voice agents are heading: less about which model sounds most natural, more about who has the governance, evaluation, and integration discipline to deploy one safely at scale. Whether a company waits for OpenAI’s Forward Deployed Engineers or builds on Realtime-2 with its own team, the underlying checklist stays the same: scoped permissions, pre-launch simulation, and a reviewed improvement loop. If your engineering team is scoping a voice or chat agent rollout and needs nearshore capacity to build that governance layer properly, Luby’s engineering team can help design and ship it.