OpenAI Presence: Enterprise AI Agent Platform Explained

OpenAI is no longer content to sell enterprises a model and wish them luck. On Tuesday, July 22, 2026, the company introduced OpenAI Presence, a managed platform for building, deploying, governing, and continuously improving production-grade AI agents that handle real customer conversations over voice and chat. It is the clearest signal yet that OpenAI’s enterprise strategy is shifting from raw API access toward owning the full stack that sits between a language model and a live support line.

The pitch is aimed squarely at the gap that has stalled most corporate agent rollouts: a capable model is easy to demo and hard to trust in production. Presence bundles the guardrails, permissions, escalation logic, and evaluation tooling that companies would otherwise have to assemble themselves, and wraps them in a deployment process OpenAI runs alongside the customer. The launch also lands one day after a widely reported breach involving an AI agent at Hugging Face, sharpening the timing of a product whose central promise is control.

What Presence actually is

Presence is a deployment platform, not a new model. OpenAI describes it as the components teams need to run agents in production, brought together in one place: company policies and standard operating procedures, guardrails, approved actions, pre-deployment simulations, evaluation tools, and a Codex-powered improvement process.

The design leans hard on constraint. According to OpenAI, Presence agents can access only the data and systems a specific workflow requires — billing records for a billing agent, claims systems for an insurance agent, IT tickets for a service-desk agent — rather than a broad grant of enterprise access. On top of those permissions sit escalation rules that hand a conversation to a human when the agent hits a scenario it should not handle alone, and simulations that test an agent against common requests, edge cases, and high-risk situations before it ever touches a customer.

The pieces map onto the broader industry conversation about how to make AI agents safe enough to run unattended. Presence is essentially OpenAI’s opinionated answer to the same question that AI guardrails frameworks and agent standards have been circling for a year: how do you let a probabilistic system act on a company’s behalf without letting it act badly?

The Codex improvement loop

The most distinctive component is the feedback mechanism. OpenAI’s Codex agent reviews past interactions, identifies weaknesses in how the deployed agent behaved, and proposes behavioral improvements. Crucially, those changes do not ship automatically — staff members must test and approve each proposed adjustment before it goes live.

OpenAI put numbers behind the loop using its own operations. The company has been running Presence internally to power its English-language phone support line, and says the system met or exceeded the quality benchmarks used to grade human support staff within a few weeks of launch. Presence now resolves roughly 75% of inbound issues without human assistance, and OpenAI says the Codex-powered improvement loop reduced human handoffs by 15 percentage points in just 10 days.

That framing — an agent that measurably gets better week over week, under human supervision — is the part OpenAI most wants enterprises to notice. It reframes an agent deployment from a one-time integration into a managed service with a visible learning curve, and it positions Codex as the engine that keeps the curve moving.

Where it runs first

Presence targets high-volume, well-defined workflows: customer support, outbound sales development, procurement, IT service requests, and HR functions. These are the domains where the cost of human staffing is high, the interactions are repetitive enough to model, and the escalation paths are clear.

OpenAI named several early customers spanning finance, telecom, and insurance:

• BBVA Mexico is using Presence for faster, more personalized customer interactions.
• SoftBank Corp. has deployed Japanese-language agents that the company says deliver natural and accurate conversations — a notable data point for anyone worried that voice agents degrade outside English.
• Retail Insurance Australia, part of the IAG group, is using the platform for customer-facing support.

The voice capability builds on OpenAI’s push into real-time, full-duplex voice, the technology that lets an agent listen and speak simultaneously rather than trading rigid turns. Combining that with policy enforcement and escalation is what turns a voice demo into something a bank will put on its main support number.

Not self-service — and that matters

Presence is launching through a limited general availability program, and it is deliberately not a self-service product. OpenAI’s Forward Deployed Engineers and a set of select global systems integrators lead the deployments, working directly with each customer to identify high-value workflows, connect internal systems, set permissions and policies, test the agent, and support it after launch.

OpenAI has not published standard pricing for Presence. The high-touch, engineer-led model implies enterprise contracts negotiated per deployment rather than a public per-seat or per-minute rate card.

That go-to-market choice is a strategy statement in itself. It tells you OpenAI believes the hard part of enterprise agents is not the model but the last mile — the integration, the policy design, the testing, the ongoing tuning. By staffing that last mile with its own engineers, OpenAI captures more of the value and more of the control, at the cost of the scalability that self-service usually brings.

The competitive backdrop

Presence enters a crowded field. Enterprise agent platforms have become one of the most contested categories in AI, with incumbents bundling agent tooling into existing CRM and support suites while a wave of startups sells standalone orchestration. The broader ecosystem has also coalesced around interoperability efforts like the Model Context Protocol and emerging agent standards for enterprise deployments, which aim to let agents connect to tools and data without bespoke integration for each one.

OpenAI’s bet is differentiation through operational rigor rather than raw capability. Anyone can wire a model to a phone line; the harder sell is a system that a compliance officer, a security team, and a head of support will all sign off on. Presence is built to be that system — and to make OpenAI the vendor of record for the entire agent, not just the model underneath it.

What it means

Presence is best read as OpenAI moving up the value chain. Selling tokens is a commodity business with thinning margins as rivals match capability and undercut on price; selling a governed, continuously improving agent that a Fortune 500 support organization depends on is a far stickier, higher-margin relationship. The company is trading some of the scalability of a pure API for the durability of a managed platform, and it is using its own support operation as the reference implementation.

Who benefits: large enterprises that have wanted agents but balked at the integration and governance burden now have a turnkey path, complete with OpenAI engineers on site. The named launch customers — a Mexican bank, a Japanese telecom, an Australian insurer — signal that OpenAI is prioritizing regulated, multilingual, high-volume operations, exactly the buyers with budget and risk tolerance to move first.

Who feels the pressure: independent agent-orchestration startups and the guardrails vendors whose entire product Presence now absorbs as a feature. When the model provider ships the governance layer for free with the model, the standalone version of that layer gets harder to sell. Contact-center and CRM incumbents also now face a first-party OpenAI offering aimed at the same workflows they monetize.

What to watch next. Three things. First, pricing and availability — whether Presence stays engineer-led and bespoke or eventually opens to self-service will determine how much of the market it can actually reach. Second, the safety record — a platform that markets itself on control will be judged harshly on its first public failure, and the Hugging Face incident a day earlier is a reminder that agent security is now a headline risk. Third, the Codex loop’s durability — the 15-points-in-10-days figure is a compelling internal result, but enterprises will want to see the improvement curve hold up across messier, less cooperative domains than OpenAI’s own support queue. If it does, the managed-agent model becomes the template every rival has to answer. If you want to understand the machinery underneath, our guide to building your own AI agent walks through the same components Presence is now selling as a service.