OpenAI Presence Brings Engineers Into the Enterprise AI Agent Deployment Process
OpenAI is introducing a new way for large organizations to deploy artificial intelligence agents, and it looks less like purchasing software and more like hiring a specialized implementation team.
Announced on July 22, OpenAI Presence is a managed enterprise service that combines AI agents with engineers who work directly with customers to design, test, integrate, and improve automated workflows.
The service is currently available only through a limited general availability program. It cannot yet be purchased or configured through a self-service platform.
Deployments are handled by OpenAI’s Forward Deployed Engineers, working alongside selected global systems integrators and the customer’s internal technology teams.
A service built around one specific business task
Instead of asking an enterprise to automate an entire department immediately, each Presence engagement begins with one clearly defined job.
This could involve resolving a billing dispute, processing an insurance claim, answering a customer support inquiry, or handling an employee’s IT service request.
The AI agent receives access only to the information and systems required to perform that task.
Customers also define the operational rules, including:
- What actions the agent is allowed to perform
- Which decisions require approval
- When a case must be escalated
- When a human employee should take control
This approach is designed to prevent an AI agent from receiving unnecessary access or acting beyond its assigned responsibilities.
After deployment, OpenAI’s Codex analyzes production sessions, failed interactions, and human escalations. It may then recommend modifications to the agent’s instructions, tools, or workflow.
However, these proposed changes must still be tested and approved by the customer before being released into production.
Production agents require more than company documents
OpenAI’s documentation emphasizes that an enterprise agent does not become ready for real-world use simply by uploading policies, manuals, and internal documents.
The company describes a six-stage deployment process covering:
- Business outcome definition
- Security and system architecture
- Privacy and legal review
- Simulation and acceptance testing
- Controlled production rollout
- Continuous post-launch improvement
This process reflects one of the major lessons enterprises have learned from early AI agent projects.
The language model is only one component. Successful implementation also depends on system integration, identity verification, permissions, data governance, escalation procedures, and organizational change management.
Addressing the weaknesses of early agentic AI projects
OpenAI Presence arrives as many businesses struggle to move AI agents from experimental demonstrations into reliable production systems.
Gartner has predicted that more than 40 percent of agentic AI projects could be cancelled by the end of 2027. The research firm identified unclear business value, weak governance, and poor operational discipline as major causes of failure.
Presence appears to be designed around those concerns.
Before an agent interacts with customers or employees, simulations and automated evaluation systems can test whether it:
- Reached the correct outcome
- Followed company policies
- Used connected tools properly
- Requested approval when required
- Escalated the interaction at the correct time
Guardrails are intended to stop or redirect conversations when they move outside approved boundaries.
Session records and action histories can also be reviewed for audit and compliance purposes.
When human intervention is required, the system can transfer structured information about the case instead of handing an employee only a long, unorganized conversation transcript.
New versions of an agent may also be introduced through staged deployments, with rollback procedures available when problems appear.
OpenAI is entering the implementation business
The managed nature of Presence represents a significant shift for OpenAI.
The company has primarily sold access through API consumption, ChatGPT subscriptions, and enterprise software licenses. Presence is closer to a consulting and systems implementation engagement.
This model may help enterprises that lack the internal expertise required to build production-grade agents. However, it also creates a major scaling challenge.
Software can be distributed to thousands of customers quickly. Experienced engineers who are authorized to work inside a bank, insurance company, airline, or telecommunications provider cannot be scaled as easily.
OpenAI says participation will depend on three factors:
- Whether the workflow is suitable for the service
- Whether the organization is ready for implementation
- Whether deployment personnel are available
This means delivery capacity, not only model capability, may determine which customers receive access.
The term Forward Deployed Engineer is closely associated with Palantir, where engineers may work inside customer environments for extended periods to solve operational problems.
By adopting a similar structure, OpenAI is moving into a market traditionally occupied by consulting firms, systems integrators, and enterprise technology service providers.
It will still depend on those partners to increase deployment capacity, but it may also compete with them for control over the most important parts of an AI implementation.
Accountability must be defined clearly
OpenAI’s expanded role also raises questions about responsibility.
When the model provider also helps design the workflow, connect enterprise systems, and configure operating policies, customers must clearly define who is accountable when an agent makes an incorrect decision.
Contracts may need to specify responsibility for:
- Policy configuration errors
- Incorrect system permissions
- Model behavior
- Failed escalations
- Unauthorized actions
- Data handling incidents
- Changes to the underlying model configuration
These responsibilities cannot simply be assumed after the agent is already operating in production.
OpenAI presents its own support line as proof
OpenAI describes Presence as “battle-tested,” although the product itself is new to the market.
The company says the service was developed from several years of experience building and operating agents with enterprise customers before the offering received an official name.
Its strongest example is OpenAI’s English-language telephone support line, 1-888-GPT-0090.
According to OpenAI, the agent reached or exceeded its internal benchmarks for frontline human support within several weeks.
The company also reports that the system:
- Resolves 75 percent of incoming issues without human assistance
- Reduced human handoffs by 15 percentage points within ten days
- Continues to improve through the Codex evaluation and recommendation process
These figures were reported by OpenAI using its own benchmarks and evaluation methods. They have not been independently verified.
The results are still significant because they provide an early indication of how OpenAI intends to measure the operational performance of its enterprise agents.
BBVA, SoftBank, and IAG join early testing
OpenAI has identified three organizations working with Presence during its early deployment stage.
BBVA Mexico is exploring voice-based support for common banking concerns.
SoftBank is testing Japanese-language customer interactions.
International Airlines Group, or IAG, is exploring the use of AI support during periods of unusually high demand, including severe weather disruptions.
BBVA Mexico’s head of AI transformation, Daniel Ordaz, described the bank as a design partner helping develop and refine voice experiences for financial customer service.
The partnerships demonstrate interest from major companies, but OpenAI has not said that any of the three organizations are already operating Presence at full enterprise scale.
This distinction is important. Early design partnerships provide valuable testing environments, but they are different from broad production deployments serving millions of users.
Pricing and model details remain undisclosed
OpenAI has not published standard pricing for Presence.
Implementation scope and cost will be determined separately for each customer and workflow.
This is common for large enterprise technology projects, but it makes direct cost comparisons difficult. Potential customers do not yet have a public benchmark for measuring the cost per resolved support request against traditional contact-center providers.
OpenAI has also not disclosed which specific models power Presence.
The company says model selection will depend on the requirements of each workflow and may change as the deployment evolves.
This flexibility allows OpenAI to upgrade an agent when better models become available. However, enterprises will need contractual protections and evaluation requirements to ensure that future model changes do not reduce accuracy, compliance, or reliability.
Companies that have built internal testing systems around specific model versions may also require advance notification and formal acceptance testing before any configuration is changed.
Data governance will be defined per deployment
During the limited general availability period, Presence can support voice or chat interactions.
Contact-center integration, identity verification, routing, authentication, and human handoff procedures will be determined for each deployment.
Data handling arrangements will also depend on the signed architecture and customer contract.
For regulated industries such as banking, healthcare, insurance, and telecommunications, these documents will be more important than general product marketing materials.
Organizations will need precise answers about data storage, retention, regional processing, access controls, audit records, and the use of production data for system improvement.
Three different paths for enterprise AI agents
OpenAI now offers organizations several ways to build and deploy AI agents.
ChatGPT Workspace Agents provide a more self-service option for teams working inside environments such as ChatGPT and Slack.
The OpenAI API allows developers to build custom applications using the company’s models.
OpenAI Presence provides a managed implementation model in which OpenAI engineers and partners help design and operate the agent.
The underlying AI capabilities may overlap, but the main difference is who performs the integration and deployment work.
For some enterprises, the deciding factor may no longer be which company has the most powerful model.
The more practical question may be which provider has enough qualified engineers, integration partners, governance processes, and operational experience to make the agent work reliably inside the organization.
What this means for Philippine enterprises
OpenAI Presence could be relevant to large Philippine banks, telecommunications companies, business process outsourcing firms, airlines, insurers, hospitals, and government agencies seeking to automate high-volume service workflows.
However, local deployment would require more than translating an English-language agent into Filipino.
Agents operating in the Philippines may need to understand Tagalog, Taglish, regional accents, local customer behavior, Philippine regulations, and culturally specific communication patterns.
Organizations must also consider the Data Privacy Act of 2012, industry-specific regulations, cybersecurity requirements, and rules governing automated decision-making.
For the Philippine BPO industry, Presence represents both a risk and an opportunity.
Routine customer service tasks may increasingly be handled by AI agents. At the same time, demand may grow for Filipinos who can train, test, evaluate, localize, supervise, and improve those systems.
The future of enterprise AI may not be about replacing every human worker.
It may be about building a new operational structure where AI handles repeatable tasks, humans manage complex and sensitive cases, and specialized engineers continuously improve the system connecting them.
OpenAI Presence is an early example of that model.
Its success will depend not only on the intelligence of its agents, but also on governance, implementation discipline, local adaptation, and the availability of skilled people capable of deploying AI responsibly.
