OpenAI Presence: Inside Its New Enterprise AI Agent Platform

On July 22, 2026, OpenAI introduced Presence, a platform built for a narrower and more specific job than "access to a language model." Presence lets enterprises build, deploy, govern, and continuously improve production voice agents and chatbots, wrapping model reasoning in a company's own policies, guardrails, testing and evaluation systems, and rules for when a human needs to step in. It is a shift in what OpenAI is actually selling: not raw model access billed by the token, but a managed system that a bank, an airline, or a telecom can point at its phone line and trust to handle real customers.

OpenAI

Google Cloud

What Presence actually does

Presence is not a new model. It is an orchestration and governance layer that sits on top of OpenAI's models and is meant to close the gap between a model that can hold a conversation and a system a company is willing to put in front of paying customers. According to OpenAI's announcement, that layer includes pre-deployment simulation and testing against edge cases, permission controls that define what an agent is and is not allowed to do, and escalation rules that route a conversation to a human when it goes outside the agent's competence. The pitch is that governance, not raw conversational ability, is what has been missing from most enterprise voice and chat deployments.

75%Auto-resolved on OpenAI's own support lineWithout human help

15 ptsHandoff reductionVia the Codex improvement loop, in 10 days

July 22, 2026Launch dateLimited general availability

3+Named early adoptersBBVA, SoftBank, IAG

OpenAI's own support line as the proof point

Rather than lead with a customer case study, OpenAI pointed to its own operations. The company says Presence already powers its English-language phone support line, reachable at 1-888-GPT-0090, and that it resolves about 75% of inbound issues without a human ever getting involved. That figure alone is a claim worth treating with the appropriate caution, it is self-reported and has not been independently audited, but it is a specific, falsifiable number rather than a vague marketing claim, and OpenAI is willing to run its own support operation on the product it is selling.

The more interesting number is the second one. OpenAI describes a continuous improvement loop in which its Codex coding agent reviews production sessions and escalations, identifies where the support agent is failing, and proposes behavioral or policy changes, which human staff then test and approve before they go live. OpenAI says that loop cut human handoffs by 15 percentage points within 10 days. If accurate, that is a genuinely fast iteration cycle for a customer-facing system, the kind of week-over-week tightening that would take a traditional support organization a quarter or more of workflow redesign to achieve manually.

At BBVA, we are working closely with OpenAI to explore how trusted customer agents can help shape the future of financial services.
Daniel Ordaz, Head of AI Transformation, BBVA Mexico

Early adopters, and what they are actually testing

OpenAI named a small set of enterprise partners at launch rather than a broad customer list, consistent with a product still in limited general availability. BBVA is exploring voice-based banking support in Mexico. SoftBank is testing natural, accurate Japanese-language conversations for its own customer-facing workflows, an application area where language quality has historically been a harder problem than in English. IAG, the Australian insurer, is looking at using Presence to handle customer support during high-demand periods such as severe weather events and natural disasters, when call volume spikes sharply and consistent, policy-compliant responses matter most.

Rows of cubicles with headset-wearing agents at a traditional human-staffed call center — A traditional, human-staffed call center floor in Sao Paulo, Brazil, 2012. Not affiliated with OpenAI, BBVA, SoftBank, or IAG; shown to illustrate the kind of high-volume, human-staffed support operation that platforms like Presence are aimed at partially automating. Photo by Carlos Ebert via Wikimedia Commons, CC BY 2.0.

None of these three deployments is described as fully live and unsupervised at scale. They read as pilots inside real enterprises rather than finished case studies, which is consistent with OpenAI's own framing of the launch as "limited general availability" for eligible customers, with deployments led by OpenAI's Forward Deployed Engineers and select systems integrators rather than sold as a self-serve product. That is a services-heavy motion, closer to how a company like Palantir sells software with embedded delivery teams than how OpenAI has historically sold API access.

The gap Presence is explicitly built to close

The timing lines up with a specific, widely cited piece of research. MIT's Project NANDA published a report in mid-2025 titled "The GenAI Divide: State of AI in Business 2025," based on 52 executive interviews, a survey of 153 leaders, and analysis of 300 public AI deployments. Its headline finding was that 95% of generative AI pilots at the enterprises studied delivered no measurable profit-and-loss impact, while only about 5% extracted significant, integrated value. The report's diagnosis was not that the underlying models were too weak, but that most pilots failed on integration: brittle workflows, no persistent memory of context, and tools that did not fit how the organization actually worked day to day.

Presence is a direct answer to that diagnosis. Rather than handing a company an API and letting it build its own governance, testing, and escalation logic from scratch, OpenAI is selling the surrounding infrastructure as a packaged product, with its own operations as the reference deployment. Whether that closes the pilot-to-production gap at scale, or simply moves the same integration risk from the customer's engineering team to OpenAI's Forward Deployed Engineers, is not yet answerable from a two-day-old launch. It is, however, a coherent response to a specific, sourced failure mode rather than a vague promise of "enterprise readiness."

A four-way race over different things

Presence did not launch into an empty market. Four major enterprise agent platforms shipped within roughly a month of each other in 2026: Presence on July 22, Google's Gemini Enterprise Agent Platform (a rebuild of Vertex AI and Agentspace) at Google Cloud Next on April 22, Meta's Business Agent Platform opened to all businesses on WhatsApp, Instagram, and Messenger on July 1, and the NVIDIA and ServiceNow Project Arc, announced in May, which runs long-lived desktop agents secured by Nvidia's runtime inside ServiceNow's existing AI Control Tower.

Four platforms, four different control points

All four launched within roughly a month of each other in 2026. None compete head to head on model quality alone, each is betting on a different lever to make its platform sticky.

OpenAI Presence

July 22, 2026

Managed services

A fully managed voice/chat agent system sold with governance, evals, and human escalation, deployed by OpenAI Forward Deployed Engineers rather than self-serve.

Google Gemini Enterprise Agent Platform

April 22, 2026

Governance

Rebuilt from Vertex AI and Agentspace into one platform, pitched on enterprise-grade oversight and control across many agents at once.

Meta Business Agent Platform

June 3, 2026

Distribution

Free to start, opened to all businesses on WhatsApp, Instagram, and Messenger on July 1, 2026, monetizing the reach of Meta's existing messaging apps.

NVIDIA / ServiceNow Project Arc

May 2026

Containment

Long-running desktop agents secured by NVIDIA's runtime and governed inside ServiceNow's existing AI Control Tower, extending a platform enterprises already license.

Distribution, governance, containment, and managed services are four adjacent bets, not one market with a single winner. Launch dates and framing per VentureBeat, TechTimes, and Being Shivam, July 2026.

What is notable is how little these four platforms actually compete on the same axis. Google is selling governance across many agents at once. Meta is selling distribution, agents that reach customers where the customers already are, monetized through the reach of its messaging apps rather than a per-seat fee. Nvidia and ServiceNow are selling containment, an agent platform that extends a governance relationship enterprises already have rather than asking them to adopt a new one. OpenAI is selling a managed service with people attached. Four different bets on what will actually make an enterprise agent platform sticky, running in parallel rather than converging on one obvious winner.

The buyer's tension

There is a real tension underneath all four of these launches that none of the vendors highlight in their own materials. A fully managed, single-vendor agent platform is genuinely convenient, it minimizes the integration work that MIT's research suggests is where most pilots actually fail. But it also concentrates a company's customer-facing infrastructure, its governance rules, its escalation logic, its evaluation data, inside one lab's roadmap and pricing. That is a meaningfully different commitment than picking a model for a single task.

The alternative some enterprises are exploring is keeping the orchestration, governance, and evaluation layer model-agnostic, so a workflow can be pointed at whichever model or provider fits it best today without re-architecting the surrounding system if that changes tomorrow. That is the same principle Metir AI applies at the workspace level for teams choosing between leading AI models day to day: the governance and orchestration layer is worth owning independently of any single model provider's roadmap, even as platforms like Presence make the case that a fully managed, single-vendor stack is the fastest way to close the pilot-to-production gap MIT's research documented.

The honest read

Presence is a real, specific product with a real, specific reference deployment, OpenAI's own support line, rather than a slide deck of hypothetical use cases. The 75% auto-resolve figure and the 15-point handoff reduction are self-reported and unaudited, and the named enterprise partners are still described as exploring or testing rather than running Presence at full production scale. What is clear is the strategic direction: OpenAI, like Google, Meta, and Nvidia's ServiceNow partnership, has concluded that the money in enterprise AI is no longer only in the model, it is in the governance, evaluation, and escalation layer wrapped around the model. Whether any one of these four platforms becomes the default the way Presence's Forward Deployed Engineer motion suggests OpenAI hopes it will, or whether enterprises keep that layer deliberately provider-agnostic instead, is the open question the next few quarters of actual deployments, not launch announcements, will answer.

Sources:

Image credits

Header image: Sam Altman, CEO of OpenAI, speaking on a panel, via Wikimedia Commons, photo by Village Global, licensed under CC BY 2.0. In-body photograph: a traditional, human-staffed call center floor in Sao Paulo, Brazil, photographed in December 2012 by Carlos Ebert via Wikimedia Commons, licensed under CC BY 2.0. This photograph is not affiliated with OpenAI, BBVA, SoftBank, or IAG and does not depict Presence in operation; it illustrates the kind of human-staffed support environment the platform is aimed at partially automating.