Enterprise leaders have spent the last year oscillating between the promise of autonomous AI agents and the reality of fragile prototypes that collapse under the weight of real-world edge cases. The gap between a successful API demo and a production-ready system that can actually handle a customer's angry phone call remains the primary hurdle for the Fortune 500. This week, the conversation shifted from how to build these agents to how to actually deploy them without breaking the business.

The High-Touch Architecture of OpenAI Presence

OpenAI Presence arrives not as a self-service tool, but as a managed deployment platform designed to bridge the gap between raw model capability and enterprise reliability. The platform allows qualified corporate clients to deploy AI agents across customer-facing interfaces and internal business workflows, governed by strict company-defined policies, permissions, and evaluation standards. These agents are designed to answer queries and execute authorized actions within company systems, with a built-in fail-safe that routes requests to human agents the moment a task exceeds the agent's defined scope.

Unlike the typical OpenAI product rollout, Presence is not available via a sign-up page. It is currently distributed through a limited General Availability program where the deployment process is entirely driven by OpenAI's Forward Deployed Engineers (FDEs) and a select group of global system integrators. This high-touch model ensures that the integration is not merely a technical connection but a strategic alignment of the AI's behavior with the company's operational goals. While the technical capabilities are clear, OpenAI has kept the specifics of pricing, geographic availability, and the exact costs associated with the FDE engineering hours under wraps.

The Codex Loop and the End of Static Bots

The fundamental difference between Presence and a standard chatbot lies in its continuous improvement loop. Most enterprise bots are static; they are trained, deployed, and then slowly drift into obsolescence as user behavior changes. Presence treats every operational session, every escalation to a human, and every quality signal as a data point for optimization.

This is where Codex enters the ecosystem. By utilizing a Presence plugin, Codex analyzes these quality signals to identify exactly where the agent is failing or where the logic is lagging. Instead of a developer manually guessing why a customer was frustrated, Codex proposes specific, data-driven updates to the agent's configuration. The operations team then tests these proposed changes against the version currently in production, ensuring that a fix for one problem does not create three new ones. Once validated, the update is rolled out in a controlled manner. This transforms the AI agent from a fixed piece of software into a living system that evolves based on real-time interaction data.

OpenAI has already stress-tested this loop on its own infrastructure. By applying Presence to its English-language phone support channel at 1-888-GPT-0090, the company achieved a 75% resolution rate for inbound issues without any human intervention. The system handled open-ended requests, verified callers, and used account context to perform authorized tasks. Most tellingly, the Codex-driven improvement loop reduced the rate of agent transfers by 15 percentage points in just 10 days. While these figures come from internal reporting and lack independent verification, they provide a blueprint for what is possible when engineering resources are embedded directly into the deployment process.

This model is already attracting global interest from sectors where the cost of failure is high. Mexico's BBVA is exploring the platform for voice-based banking support, while Japan's SoftBank is evaluating it for complex Japanese-language customer dialogues. In Australia, the insurer IAG is looking at Presence to manage the massive spikes in inquiry volume that typically follow natural disasters. For these companies, the draw is not the LLM itself, but the FDE-led deployment model that mitigates the risk of a public-facing AI failure.

Currently, the service fully supports real-time voice and chat experiences. While OpenAI has signaled an ambition to expand into broader channels including email, that functionality remains a future roadmap item rather than a day-one feature. For any organization considering adoption, the decision rests on three non-negotiable constraints: the absence of self-service, the requirement for FDE-led deployment, and the currently unconfirmed status of compatibility with external models.

The shift toward Forward Deployed Engineering suggests that the era of the plug-and-play enterprise agent is over, replaced by a model of high-density engineering partnerships.