Enterprise AI Agents, Engineers Included

OpenAI is taking a hands-on approach to deploy its new enterprise AI agents, a significant shift from its previous reliance on API keys and seat licenses. The newly announced “OpenAI Presence” is a managed service, currently in a limited general availability program, and it’s not yet a self-serve product. Deployments are spearheaded by OpenAI’s own Forward Deployed Engineers (FDEs) and a curated group of global systems integrators.

This new model positions Presence as a project rather than a mere product. Each engagement is tailored to a specific business challenge, such as resolving a billing dispute, processing an insurance claim, or streamlining employee IT service requests. The AI agent is granted only the necessary knowledge and system access pertinent to its assigned task. Crucially, customers define the operational parameters, including when an agent requires human oversight or when a human intervenes in the process. Post-deployment, the system leverages production session data and escalations to identify areas for improvement, with proposed changes then vetted and approved by the customer’s team before implementation.

OpenAI’s documentation for Presence is notably transparent about the comprehensive process involved. The company outlines a six-stage journey, beginning with the definition of business outcomes, followed by rigorous security, privacy, and legal reviews, simulation and acceptance testing, a phased rollout, and continuous post-launch iteration. This indicates that achieving production readiness for a Presence agent involves far more than simply ingesting vast amounts of data.

**Addressing the Critical Bottlenecks in Enterprise AI Adoption**

The necessity of this managed model becomes clearer when examining the challenges plaguing enterprise AI adoption. Gartner has projected that over 40% of agentic AI projects could face cancellation by the end of 2027. The primary culprits cited are not limitations in AI model capabilities, but rather issues with governance, poorly defined business value, and a lack of robust operational discipline.

OpenAI Presence appears to be directly engineered to counter these very issues. The inclusion of simulation and grading mechanisms ensures that an agent achieves the desired outcomes, adheres to policy, utilizes its tools effectively, and escalates appropriately before interacting with external parties. “Guardrails” are implemented to prevent interactions from deviating from predefined boundaries. Comprehensive session records and action histories provide a clear audit trail, while structured escalation paths offer human agents contextual information rather than simple transcriptions. Furthermore, new versions are deployed through controlled rollouts with the safety net of rollback capabilities.

Enterprises have spent the last two years grappling with the complexities of integrating AI agents into production environments, managing permissions, and implementing effective change management strategies. By deploying its own engineers and integrating with select partners to handle these intricate operational aspects, OpenAI is directly addressing buyer pain points that have historically led to project failures, rather than simply offering another dashboard and deeming the remaining challenges as customer responsibility.

**The Strategic Constraint: Delivery Capacity**

The constraints on Presence are evident in its eligibility criteria, which OpenAI states are based on workflow fit, implementation readiness, and available delivery capacity. Delivery capacity, in this context, represents a significant consulting bottleneck. Unlike software that can scale exponentially, engineers with the necessary security clearances to access core banking systems, for instance, are inherently limited.

The role of Forward Deployed Engineer (FDE) itself borrows from the Palantir model, signifying personnel deeply embedded within customer operations for extended periods. The economics of this approach are vastly different from the predictable costs associated with metered AI inference. By positioning its FDEs and strategic partners at the forefront of every deployment, OpenAI is stepping into a market segment traditionally dominated by systems integrators. While this arrangement can function effectively at smaller volumes, it presents a more intricate scaling challenge as demand grows.

This model also raises important questions regarding contractual clarity. When the AI model vendor also acts as the implementation partner, the lines of accountability for any policy misapplications in a production environment must be explicitly defined, rather than assumed.

**Early Stages of Enterprise AI Agent Deployment**

OpenAI describes Presence as “battle-tested,” asserting that the product has been refined through years of deploying AI agents with enterprise clients before being formally packaged and branded. This claim hinges on accumulated practical experience rather than the product’s recent market entry, and it’s a reasonable assertion to make.

A compelling, albeit self-reported, proof point is OpenAI’s own English-language phone support line. The company states that the agent met or exceeded its internal benchmarks for frontline human support within weeks. It now resolves 75% of inbound inquiries autonomously, and through the Codex improvement loop, has reduced human handoffs by 15 percentage points in just ten days. It is important to note that these figures are provided by OpenAI and measured against its own internal criteria and grading systems, and have not been independently verified.

The three named customers—BBVA, SoftBank, and IAG—represent earlier stages of adoption than the launch messaging might suggest. BBVA is exploring voice support for routine banking services in Mexico, SoftBank is piloting Japanese-language conversational AI, and IAG is investigating support during high-demand periods, such as severe weather events. Daniel Ordaz, Head of AI Transformation at BBVA Mexico, describes the bank as a design partner actively shaping and refining voice experiences for financial customer service. While design partnerships are common and valuable during limited general availability, none of these customers are currently operating Presence at a large scale, a point worth considering alongside the claim of proven deployment.

**Undisclosed Key Details**

Pricing for OpenAI Presence is not publicly disclosed. Implementation scope and associated costs are determined on a per-customer, per-deployment basis. While this is standard for enterprise services, it leaves potential buyers without a public benchmark for cost-per-resolved contact compared to incumbent contact center vendors.

The specific underlying AI model utilized by Presence is also not explicitly named. OpenAI’s documentation indicates that Presence employs OpenAI models, with configurations optimized for the specific workflow and subject to evolution as those workflows change. This flexibility is a prudent engineering approach, as locking a production agent onto a static model version can lead to obsolescence. However, teams investing in building evaluation suites against particular model versions will likely require contractual assurances regarding what they are committing to as configurations evolve.

Channel support during the limited general availability phase currently encompasses voice and chat, with contact center integration, routing, authentication, and handoff design being confirmed on a deployment-by-deployment basis. Data handling follows a similar pattern, with signed architectures and contracts serving as the governing documents rather than general published policies.

OpenAI Presence is positioned distinctly from ChatGPT Workspace Agents, which continue to serve as the self-serve pathway for teams building within ChatGPT and Slack. For voice-based customer interactions, clients retain API access to OpenAI’s frontier models. This tiered offering suggests OpenAI is now providing similar core capabilities through three distinct channels, differentiated more by the delivery model and engagement approach than by fundamental technological limitations. Ultimately, prospective buyers are being compelled to evaluate potential vendors based on their delivery capacity as much as their model capabilities, and by OpenAI’s own account, it is this delivery capacity that is currently the most constrained resource.