The latest way to buy an enterprise AI agent from OpenAI doesn’t require you to buy anything online. Announced on July 22nd, OpenAI Presence is a managed product offered through a limited general availability program, and the company clarifies that it is not yet available as a self-service product. The deployment will be led by OpenAI’s forward deployment engineers and a select set of global system integrators.
This is a major shift for a business that has operated primarily on API keys and seat licenses. Presence is sold as a project, not a product. Each engagement begins with a single job, such as resolving a billing dispute, processing an insurance claim, or processing an employee’s IT service request.
Agents are given only the knowledge and access to systems they need to do their jobs, and customers create the rules that govern what agents can do, when sign-off is required, and when someone takes over. After launch, Codex reads production sessions and escalations and suggests changes for the customer’s team to test and approve before rollout.
OpenAI’s documentation is unusually frank about the effort involved. The company’s Help Center sets out a six-step process, from scoping business outcomes to security, privacy, and legal reviews, simulation and acceptance testing, phased deployment, and post-release iterations. The Presence agent states that just ingesting documents is not production-ready.
The problem this was created to solve is real
Management models are easy to interpret cynically, but difficult to reject on the basis of evidence. Gartner warns that more than 40% of agent AI projects will be canceled by the end of 2027, with failures due to poor governance, undefined business value, and operational discipline rather than model capabilities.
Almost all Presence bundles are aimed directly at that diagnosis. Simulations and scorers test whether agents reached the correct outcome, followed policy, used tools correctly, and escalated when needed before anyone outside the company speaks to them. Guardrails intervene when interactions cross defined boundaries. Session records and behavioral history provide reviewers with audit targets. Escalation paths pass structured context to users rather than cold transcripts, and new versions are published through controlled rollouts with rollbacks.
The company spent two years discovering that the hard part about operating agents is integration, permissions, and change management. Vendors that send engineers to do that work are responding to what the buyer is actually failing, rather than shipping another dashboard and calling the gap a customer problem.
Where constraints exist
This trade-off manifests itself in eligibility criteria. According to OpenAI, access is determined by workflow suitability, implementation readiness, and available delivery capacity.
Delivery ability is a consulting constraint. Software scales. This is not the case for engineers who are qualified to work on a bank’s core systems. Forward Deployed Engineer is a title borrowed from Palantir to describe a staff member who is integrated into a customer’s operations for months at a time. The economics associated with it bear no resemblance to the economics of pay-as-you-go reasoning. By putting its own FDE and designated partners at the forefront of all deployments, OpenAI has stepped into a segment of the market dominated by integrators who also rely on it to scale up. This is a viable arrangement while volumes are small and becomes more complex later on.
It also raises questions for those considering the scope of the contract. If the model vendor is also an implementation partner, the scope of liability for policies incorrectly applied in production should be written down, rather than assumed.
Enterprise AI agents on display are still in their infancy
OpenAI describes Presence as battle-tested, and the case for the language was assembled from years of experience deploying agents to enterprise customers before the product was packaged and given a name. This claim is not about when the product came to market, but about accumulated practice, which is a reasonable one.
The strongest single point of evidence is OpenAI’s proprietary English phone support line, 1-888-GPT-0090. The company says its agents met or exceeded internal benchmarks for front-line human support within weeks, now resolving 75% of inbound issues without human assistance and reducing human handoffs by 15 percentage points in 10 days through the Codex improvement loop. These are OpenAI numbers measured against OpenAI’s own scoring criteria in OpenAI’s own channels. While transparency is welcome, the numbers have not been independently verified.
The three named customers are earlier in the cycle than the launch framework indicates. BBVA is considering voice support for day-to-day banking operations in Mexico. Softbank is testing Japanese conversation. IAG is looking at support during high demand events such as severe weather. Daniel Ordaz, head of AI transformation at BBVA Mexico, describes the bank as a design partner that helps shape and refine voice experiences for financial customer services. Design partners are normal and useful in limited GA. However, none of the three have been shown to be running a large-scale presence, which is worth holding alongside the word proven.
What has not been disclosed
Price not disclosed. Deployment scope and costs are set per customer and per deployment. While this is normal for enterprise services, buyers are still left without a public reference point for cost per contact resolved against their existing contact center vendors.
The model has no name. According to the documentation, Presence uses an OpenAI model, and its configuration is chosen for your workflow and may change as your workflow evolves. This flexibility is defensible engineering, as locking production agents into frozen model versions would be horribly stale. The team has spent the past year building an evaluation suite for a specific version, but they still want the contract to spell out what will be honored when migrating configurations.
Channel support during limited GA will cover voice or chat, and contact center integration, routing, authentication, and handoff design will be reviewed on a deployment-by-deployment basis. Data processing follows the same pattern, with signed architectures and contracts treated as administrative records rather than published policies.
Presence is separate from ChatGPT and ChatGPT Workspace Agents, which remains a self-service path for teams building within Slack, while voice customers maintain API access to OpenAI’s frontier model. The company currently offers roughly the same functionality in three ways. This is more about who does the work than what the technology can do.
Therefore, the buyer chooses not only the model’s capabilities but also the delivery capabilities, and OpenAI’s own responsibility is the distribution capability.
See: HP accelerates enterprise workflows with OpenAI Frontier
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber Security & Cloud Expo. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
