OpenAI Enterprise AI Agents Come with Dedicated Engineering Support

The latest approach to acquiring enterprise AI agents from OpenAI represents a significant shift from traditional online purchasing models. OpenAI Presence, officially announced on July 22, operates as a managed solution delivered through a limited general availability program. Unlike conventional self-service products, this offering requires direct engagement with OpenAI's Forward Deployed Engineers and carefully selected global systems integrators.

This marks a strategic departure for a company traditionally built on API keys and seat licensing models. Presence functions as a project-based implementation rather than an off-the-shelf product. Each deployment begins with a specific business objective, such as:

• Resolving billing disputes
• Processing insurance claims
• Managing employee IT service requests

The AI agent receives precisely scoped access—only the knowledge and system permissions necessary for its designated task. Customers maintain full control by defining operational rules, approval workflows, and human escalation triggers. Following deployment, Codex analyzes production sessions and escalation patterns, proposing refinements that customer teams review and approve before implementation.

📋 Transparent Implementation Requirements

OpenAI's documentation demonstrates unusual candor regarding the implementation effort involved. The help center outlines a comprehensive six-stage deployment process:

1. Business outcome scoping
2. Security, privacy, and legal review
3. Simulation and acceptance testing
4. Staged rollout procedures
5. Post-launch iteration cycles

The documentation explicitly states that document ingestion alone does not create production-ready agents—a critical acknowledgment of real-world deployment complexity.

🎯 Addressing Real Market Challenges

The managed delivery model addresses documented industry failures. Gartner has projected that over 40% of agentic AI projects will be cancelled by end of 2027, attributing failures to:

Governance gaps, undefined business value, and weak operational discipline—rather than technological limitations

Presence directly targets these pain points through:

• Pre-deployment simulation and grading to verify correct outcomes, policy compliance, and appropriate escalation behavior
• Real-time guardrails that intervene when interactions exceed defined parameters
• Comprehensive session records and action histories for audit purposes
• Structured escalation handoffs providing context beyond raw transcripts
• Controlled version rollout with rollback capabilities

Enterprises have discovered that production agent challenges center on integration, permissions, and change management—not core AI capabilities. By deploying engineers to handle this work directly, OpenAI addresses actual implementation failures rather than shifting responsibility to customers.

⚖️ Delivery Capacity Constraints

Access eligibility depends on three factors: workflow compatibility, implementation readiness, and available delivery capacity.

Delivery capacity represents a consulting bottleneck. While software scales infinitely, engineers with clearance to access banking core systems do not. The Forward Deployed Engineer role, borrowed from Palantir's playbook, involves staff embedded in customer operations for extended periods. This model operates under fundamentally different economics than metered API inference.

By positioning proprietary FDEs and designated partners at every deployment's forefront, OpenAI enters territory traditionally occupied by systems integrators—creating a workable arrangement at current volumes but presenting scalability questions as demand grows.

This structure also raises accountability considerations: when the model vendor simultaneously serves as implementation partner, contractual responsibility for production policy failures requires explicit documentation rather than assumptions.

📊 Early-Stage Customer Deployments

OpenAI characterizes Presence as "battle-tested," basing this claim on years of enterprise agent deployments predating the formal product packaging. This assertion references accumulated operational experience rather than extended market availability—a reasonable but important distinction.

The primary proof point is OpenAI's own English-language support line (1-888-GPT-0090). According to company metrics, the agent:

• Met or exceeded internal human support benchmarks within weeks
• Currently resolves 75% of inbound issues without human intervention
• Reduced human handoffs by 15 percentage points in 10 days through Codex optimization

Note: These figures represent OpenAI's internal measurements against proprietary grading criteria and lack independent verification.

The three publicly named customers occupy earlier deployment stages than launch messaging might suggest:

• BBVA – Exploring voice support for everyday banking services in Mexico
• SoftBank – Testing Japanese-language conversational capabilities
• IAG – Investigating support during high-demand scenarios like severe weather events

Daniel Ordaz, head of AI transformation at BBVA Mexico, describes the bank as a design partner helping shape voice experiences for financial customer service. Design partnerships are standard during limited GA phases. However, none of these organizations are presented as operating Presence at production scale—a significant caveat when evaluating "proven" claims.

🔒 Undisclosed Critical Details

Pricing remains unpublished. Implementation scope and costs are determined per customer and deployment—standard practice for enterprise services, yet leaving buyers without public benchmarks for cost-per-resolved-contact comparisons against incumbent contact center vendors.

Model specifications are not disclosed. Documentation states that Presence utilizes OpenAI models with configurations selected for specific workflows and subject to evolution as those workflows develop. While this flexibility represents sound engineering practice—avoiding the degradation associated with frozen model versions—teams that have invested in version-specific evaluation suites will require contractual clarity regarding performance standards when configurations change.

Channel support during limited GA encompasses voice or chat, with contact center integration, routing, authentication, and handoff design confirmed on a deployment-by-deployment basis. Data handling follows identical patterns, with signed architecture documents and contracts serving as governing records rather than published policies.

Presence operates separately from ChatGPT Workspace Agents, which remain the self-service path for teams building within ChatGPT and Slack environments. Voice customers retain API access to OpenAI's frontier models. The company now offers substantially similar capabilities through three distinct pathways, differentiated primarily by who performs the implementation work rather than technological differences.

This structure positions buyers to evaluate based on delivery capacity as much as model capability—and by OpenAI's own acknowledgment, delivery capacity represents the rationed resource.