OpenAI’s own phone support line now resolves 75 percent of English-language calls without a single human stepping in. That number comes from the same system OpenAI just opened up to the rest of the world. On July 22, 2026, the company introduced the OpenAI Presence AI agent platform, a deployment product built specifically to run AI agents inside real business workflows, not just demo them.

Presence is not a new model. It is a foundation for company context, policies, permissions, guardrails, and evaluations so an agent behaves consistently whether a customer reaches it by voice, chat, or another channel. OpenAI is positioning it for customer support, sales development, procurement, IT, and HR, the kind of high-stakes internal work that most companies have been afraid to hand to AI until now. For a site built around tracking how agentic AI actually gets adopted, this launch is worth a close look, because it answers a question a lot of readers have been asking: what does a production-grade AI agent deployment actually require beyond a good model?

What the Enterprise AI Agent Platform Actually Does

Presence bundles several pieces that companies previously had to stitch together on their own. It ships with a shared layer of company policies and standard operating procedures, a guardrail system that limits what an agent can say or do, a defined set of approved actions, simulation tools for testing, and evaluation graders that check outcomes before anything reaches a real customer.

Before a deployment goes live, teams can run it against common requests, tricky edge cases, and higher-risk scenarios. The graders check whether the agent reached the correct outcome, followed policy, used its tools correctly, and escalated to a human when it should have. That escalation piece matters. Presence agents only get access to the systems and data needed for a specific workflow, such as billing records, insurance claims, or IT tickets, rather than a standing connection to everything a company owns.

After launch, the system keeps improving itself. Codex, OpenAI’s coding agent, reviews production sessions and escalations, then proposes changes to the agent’s behavior. Human staff still test and approve those changes before they go live, which keeps a person in the loop even as the agent gets better on its own.

Early Results From AI Customer Support Agents in Production

Three named companies are already running Presence, and their use cases show how differently the same platform can be applied. BBVA Mexico is using it for faster, more personalized customer interactions at one of Latin America’s largest banks. SoftBank Corp has deployed Japanese-language voice agents that the company says hold natural, accurate conversations, a meaningful bar to clear in a language where tone and formality carry real weight. Retail Insurance Australia, part of the IAG group, is running Presence for customer-facing insurance support.

None of this is available off the shelf yet. OpenAI is rolling Presence out through a limited general availability program, and deployments are led by OpenAI’s own forward deployed engineers alongside select systems integrators rather than through self-service signup. That is a notable choice. It mirrors a pattern this site has covered before with Microsoft’s Frontier Company push into forward deployed AI engineering and TCS converting thousands of staff into forward deployed AI engineers: the vendors best positioned to sell AI agents right now are the ones willing to sit inside a client’s systems and do the integration work themselves, not just hand over an API key.

As VentureBeat reported on the launch, the 75 percent resolution rate on OpenAI’s own support line, plus a 15 percentage point drop in handoffs to humans within ten days, is the kind of number every customer support leader wants to see before committing budget. Whether that holds up across industries with more regulatory complexity than tech support remains the open question.

What Businesses Evaluating AI Agent Guardrails Should Know

If you are a business leader looking at Presence or a similar enterprise AI agent platform, the details that matter most are not the flashy parts. They are the boring ones: what data can the agent touch, how is that scoped per workflow, who approves behavior changes, and what happens when the agent is unsure.

Presence’s answer, at least on paper, is restrictive by design. Agents get workflow-specific access rather than broad permissions, every behavior change runs through simulation and human approval, and the system is built to escalate rather than guess. That approach directly addresses a governance gap this site flagged just two days ago: a 2026 OutSystems survey found 96 percent of enterprises already running AI agents in production, but only 12 percent say they can actually govern them. Presence is essentially OpenAI’s answer to that gap, built into the product rather than bolted on afterward. See our full breakdown of how leading AI agent governance platforms compare for more on how rivals are approaching the same problem.

For any business comparing platforms, the practical questions are the same regardless of vendor: does the platform separate policy from model, can you test before you ship, and does a human stay in the approval loop for changes. Those three questions apply whether you are looking at Presence, Google’s Gemini Enterprise Agent Platform, or Anthropic’s Claude Cowork, which we compared in our piece on AI agent work platforms and the ChatGPT Work versus Cowork race.

The Bigger Shift Behind Enterprise AI Agent Platforms

Presence is a signal of where the AI agent market is heading in the second half of 2026. The competitive fight is no longer just about which model scores highest on a benchmark. It is about who can wrap that model in enough policy, testing, and access control that a risk-averse enterprise will actually deploy it in front of real customers.

That shift favors companies willing to do unglamorous integration work over companies that only ship a chat interface and hope for the best. It also raises a fair concern: as more of these platforms position themselves as the trusted layer between agents and company systems, the vendor that wins the deployment contract may end up owning a lot of a company’s operational context. That is a trade-off worth weighing carefully, not a footnote.

Conclusion

OpenAI Presence packages policies, guardrails, simulation, and continuous improvement into one enterprise AI agent platform, and its early results at BBVA, SoftBank, and IAG suggest the approach works well enough to trust with real customers. The rollout also confirms that deployment expertise, not just model quality, is becoming the deciding factor in enterprise AI agent adoption. And the emphasis on guardrails and scoped access shows vendors are finally responding to the governance gap that has followed agentic AI all year.

Want to keep up with every major AI agent launch, framework, and governance shift as it happens? Explore more tools, articles, and deployment guides at BigAIAgent.tech.

If a platform like Presence can already resolve three out of four support calls without a human, what part of your own business would you trust it with first?