The Vertical Integration of Agent Infrastructure
The launch of OpenAI Presence on July 22, 2026, marks a definitive shift in the trajectory of the agent economy. By moving beyond the role of a pure model provider to become an infrastructure-layer orchestrator, OpenAI is signaling that the most critical battleground for enterprise AI is not the model itself, but the deployment layer. Presence functions as a governance-focused control plane, designed to bridge the gap between large language models and the rigid, often fragmented, reality of enterprise legacy systems.
At its core, Presence is an enterprise AI agent deployment platform tailored for voice and chat workflows. However, its primary value proposition is not the underlying intelligence, but the governance layer. This architecture provides policy enforcement, guardrails, identity verification, and granular permission controls. These features ensure that agents operate strictly within an approved set of actions, effectively creating a safety perimeter that is essential for enterprise adoption. The platform also includes evaluation and simulation testing tools, allowing organizations to stress-test agents against edge cases and high-risk scenarios before they are exposed to production environments.
The operational model behind Presence mirrors the Palantir playbook: high-touch, human-led engineering. Rather than offering a standardized, self-service protocol, OpenAI relies on Forward Deployed Engineers (FDEs) to embed directly with clients. This approach was bolstered by the acquisition of Tomoro, a London-based applied AI consulting firm, which brought approximately 150 FDEs into the fold. These engineers are responsible for the complex task of integrating agents with legacy infrastructure, a process that remains manual and bespoke.
A key component of this infrastructure is the Codex-powered improvement loop. This system reviews production sessions and escalations to propose behavioral fixes, which are then tested and approved by staff before deployment. OpenAI reports a 75% resolution rate on its own English-language phone support using this system, though it is important to note that this figure is self-reported and lacks a standardized definition of what constitutes a ‘resolution.’ Within 10 days of implementing this feedback loop, the company observed a 15 percentage point reduction in human handoffs.
This deployment strategy creates structural tension with incumbent enterprise software providers. By building its own governance and deployment layer, OpenAI is effectively bypassing the traditional application-layer integration points typically managed by platforms like Salesforce Agentforce, ServiceNow, Zendesk, Genesys, NICE, and Amazon Connect. The governance layer acts as the connective tissue between the model and the enterprise, and by controlling this layer, OpenAI is positioning itself to dictate the terms of agent interaction.
The agent infrastructure landscape is rapidly consolidating. The MCP protocol finalized its stateless specification in late July, while Snowflake shipped a centralized MCP gateway at Black Hat USA. The governance layer that Presence is building sits on top of these protocol and gateway layers — it’s the control plane that decides what agents can actually do once they’re connected.
Despite the ambition, the platform currently faces significant infrastructure gaps compared to established incumbents. Presence lacks native workforce management, interaction routing engines, quality management, and omnichannel reporting. These are the foundational elements of modern enterprise contact centers, and their absence suggests that Presence is currently optimized for specific, high-value workflows rather than broad-spectrum enterprise replacement.
The current deployment landscape also highlights the experimental nature of this infrastructure. While early customers include BBVA Mexico for banking voice support, SoftBank for Japanese-language conversations, and IAG for Australian insurance, there is no external US enterprise customer identified to date. OpenAI’s own phone support remains the primary documented deployment, suggesting that the platform is still in a phase of internal validation and high-touch refinement.
The formation of the OpenAI Deployment Company in May 2026, with a $14 billion valuation and $4 billion in initial investment, underscores the scale of this infrastructure play. With 19 investors and TPG-anchored backing, the entity is clearly designed to support the capital-intensive nature of embedding engineers within global enterprises. This is not merely a software play; it is a service-heavy infrastructure strategy.
The lack of protocol interoperability and standardization creates a risk of new silos, even as enterprises attempt to automate workflows. If the governance layer is the primary point of control, then the model provider has successfully captured the most valuable real estate in the enterprise stack. But without a common language for agent governance, the fragmentation problem may simply move up the stack.
By embedding governance directly into the deployment platform, OpenAI is forcing a confrontation with the established software giants. Whether this high-touch, FDE-led model can scale to meet the demands of the broader enterprise market, or if it will remain a boutique solution for the most complex integrations, remains the central question.