OpenAI launched Presence on July 22, 2026 — a platform for deploying AI voice and chat agents inside enterprise organisations. The announcement generated the expected volume of coverage. What most of that coverage missed is a structural design detail that tells you more about the maturity of enterprise AI agents than anything in the press release.
What Actually Happened
According to OpenAI's announcement and reporting by VentureBeat, Presence is a managed platform that lets large organisations run AI agents across customer support, outbound sales, and internal operations channels — both voice and chat. The system runs on GPT-5.6, the model family OpenAI released earlier this month, and combines model reasoning with company-defined policies, guardrails, and escalation rules. Before any agent goes live, teams can simulate interactions and test against edge cases using built-in graders — a mechanism for catching failure modes before they reach customers.
The early customer list includes SoftBank in Japan, BBVA Mexico, and IAG. OpenAI reported that its own phone support channel now handles 75% of inquiries autonomously using Presence.
There is no self-service tier. Deployments are led by OpenAI's Forward Deployed Engineers and select global systems integrators. Pricing has not been disclosed — VentureBeat reported asking twice without receiving an answer. The Register characterised the model as "consulting rates for boots-on-ground deployment," which is an accurate description of what limited general availability with individually scoped engagements actually means.
The Governance Problem OpenAI Just Admitted
The structural choice that defines Presence — no self-serve, no self-deployment, OpenAI staff in the room — is not a business model preference. It is an engineering admission. Getting an AI agent into production in an enterprise environment involves a class of problems that a well-designed API alone cannot solve: what does the agent do when a customer says something the training data never anticipated? What happens when company policy changes and agent behaviour needs to update across every live instance? Who is accountable when the agent takes an action that causes a downstream problem?
Presence's architecture addresses these questions directly. Policies and guardrails are defined by the enterprise and can intervene mid-interaction. A simulation layer tests agents against edge cases before deployment. There is a human checkpoint between every policy change and production. These are not features bolted onto a model — they are the infrastructure of operational trust, and they require significant implementation work.
The consulting model exists because that implementation work is not currently automatable. The enterprises that have already discovered this — by attempting to build production AI agents without adequate governance infrastructure — will recognise the problem immediately. The enterprises that have not yet discovered it may underestimate what Presence is actually selling.
What Presence is selling is not the model. GPT-5.6 is available via API without a consulting contract. What Presence is selling is the guarantee that the agent will behave predictably in production, at enterprise scale, across real customer interactions with real consequences. That guarantee requires human expertise to establish and maintain. OpenAI has concluded, based on running agents inside its own operations, that this expertise is worth packaging and charging for separately.
The Enterprise Lens
If you are evaluating whether to deploy an AI agent to handle customer interactions — phone calls, chat support, or internal helpdesk requests — the launch of Presence is a useful calibration point. Not because you need Presence specifically, but because OpenAI's design choices define what an honest scope of an agent deployment actually looks like.
The business case for customer-facing AI agents is straightforward: if your support team spends the majority of its time on predictable, repeatable queries — order status, policy questions, account changes, delivery updates — there is real cost reduction available. The question is not whether to pursue it. The question is what it takes to do it reliably. Presence answers that question with staffed implementation, a governance layer, and a simulation environment before go-live.
If your current plan to deploy a customer-facing agent does not include equivalent provisions — a defined process for what happens when the agent gets it wrong, a mechanism for updating agent behaviour as your policies change, and a test environment that reflects real customer conversations — your current plan is underspecified. The practical question to put to your technology partner is not "can we build this" but "what is our governance model for agent behaviour, and what happens on day sixty when a policy changes and we need to update the agent."
What to Watch
- Whether Microsoft Copilot Work and Google's enterprise agent offerings respond with comparable governance infrastructure, or continue emphasising self-service deployment — the gap between the two approaches will become visible as production failure rates surface in the next six to twelve months
- How OpenAI's consulting model scales: the Forward Deployed Engineer model is expensive and hard to staff at volume, and OpenAI will need to automate the implementation layer or certify partner networks before Presence can reach mid-market enterprise customers
- Whether Presence's simulation and grading capabilities are made available as standalone developer tools — if they are, that would significantly change the accessibility of production-grade agent governance for teams that cannot afford the full consulting engagement
Sources
- Introducing OpenAI Presence — OpenAI, July 22, 2026
- OpenAI unveils Presence — VentureBeat, July 2026
- OpenAI tries the consulting path with 'Presence' — The Register, July 22, 2026
- OpenAI Launches Presence, an Enterprise AI Agent Platform — MLQ News, July 2026