OpenAI has introduced Presence, a managed platform for deploying enterprise voice and chat agents that can use company systems, take approved actions, and escalate interactions to people when necessary. Companies considering the platform must decide what these agents can access and do, when human approval is required, and which changes are allowed to reach production.

Those decisions become more difficult after launch. An agent’s answers and actions can become outdated as products, policies, and user behavior change. Companies therefore need a way to update the agent while retaining control over which proposed changes reach production.

Presence addresses that challenge by combining the AI model with company-defined permissions, policies, guardrails, simulations, evaluations, and escalation rules. Teams monitor the agent’s performance after launch, test suggested improvements against the version already in use, and decide which updates move into production.

The platform affects enterprises considering AI agents for customer support, outbound sales, billing, insurance claims, employee IT requests, and other workflows involving company systems or approved actions. It also affects the customers and employees who may rely on those agents for help. Presence is currently available to eligible enterprise customers through a limited, OpenAI-led program rather than as a self-serve product.

In short, Presence gives enterprises a managed way to deploy AI agents that can perform approved work and adapt after launch while the company retains control over their access, actions, escalation rules, and updates.

OpenAI Presence is a managed enterprise system for deploying, testing, monitoring, and improving production AI voice and chat agents within company-defined boundaries.

A managed enterprise agent platform is a system for deploying AI voice and chat agents with company-defined access, actions, testing, and human oversight.

OpenAI has introduced Presence, an enterprise platform for putting real-time voice and chat agents into production across customer-facing and internal workflows. OpenAI says enterprises have already shown that AI agents can work, and the challenge now is making them reliable enough to perform high-value work in production.

Presence agents can answer questions, resolve issues, use company systems, take approved actions, and transfer work to people when necessary. OpenAI lists customer support and outbound sales among the supported uses, along with higher-risk work involving billing issues, insurance claims, and employee IT service requests.

The answers and actions that work today may no longer be correct when a company changes its products or policies or when users begin interacting with the agent differently. The agent’s behavior must adapt to those changes if it is going to remain reliable. OpenAI says that requires systems that help operate and update the agent, evaluations that test whether it continues to perform correctly, and deployment expertise to manage those changes in a real business setting.

Presence provides that support around the AI model. OpenAI says the platform combines model reasoning with company policies, guardrails, and rules for when an interaction must be escalated to verify the agent’s accuracy and performance. It also includes standard operating procedures, a defined set of approved actions, simulations, evaluation tools, and a Codex-powered process for improving the agent after launch.

Giving an agent access to company documents does not make it ready to handle real business work. Each Presence rollout also requires the agent’s job to be clearly defined, the necessary company systems to be connected, and the agent to pass testing, review, and approval. An agent that can use those systems and take action raises an immediate question for businesses: How do they control exactly what it is allowed to do?

Every Presence deployment begins with a specific job. The agent receives only the company knowledge and system access required to perform that work, with connections to business systems made through APIs and other tools that limit its permissions to the approved workflow. The exact capabilities, system connections, and supported actions are determined during technical planning for each deployment.

The company then sets the boundaries for the agent’s work. It decides which actions the agent may take, which require approval, and when the interaction must be transferred to a person.

In a billing-support workflow, for example, the agent can understand the customer’s request and verify the customer before accessing or acting on the account. It can then retrieve the relevant account information from connected company systems, apply the company’s billing policy, and complete only an action that has been approved as part of the workflow.

Those boundaries define what the agent is supposed to do, but companies still need to know whether it will follow them when an interaction becomes difficult or unexpected. So, how can enterprises test the agent before it begins working with real customers or employees?

Before a Presence agent begins working with customers or employees, the company and its deployment team complete security, privacy, and legal reviews. They also test how the agent responds to common requests, edge cases, which are less common or unexpected situations at the boundaries of its work, and scenarios carrying greater potential risk.

Simulations recreate situations the agent may encounter without first exposing real users to them. Graders then evaluate whether the agent reached the intended result, followed company policy, used its connected tools correctly, and transferred the interaction to a person when required.

Guardrails provide another layer of control by intervening when an interaction crosses the boundaries established by the company. The deployment moves toward a controlled production rollout only after simulations, evaluations, and acceptance testing have been completed.

Pre-launch testing can prepare the agent for known situations and anticipated risks. Once company products, policies, and user behavior begin changing, however, how can teams identify new problems and safely update the agent once it’s already serving users?

Testing and evaluation continue after a Presence agent enters production. Teams monitor real interactions, including production sessions, escalations, and other quality signals, to see where the agent is performing well and where it needs attention. Escalations can reveal work the agent could not complete or situations in which company rules required a person to take over. As products, policies, and user behavior change, those signals can also expose new gaps in the agent’s performance.

Codex uses the Presence plugin to investigate the production signals and suggest updates; it does not make those changes on its own. Company teams retain control by testing each proposed change against the version currently serving users and deciding which changes get made through a controlled rollout. OpenAI says this process allows Presence to improve as the business, its customers, and its employees change.

Companies can also carry policies, evaluations, and escalation rules from one Presence deployment to another. They can then change the parts specific to a particular workflow or communication channel without rebuilding every shared control from the beginning.

OpenAI developed Presence in close collaboration with its Research team. The company says general lessons from individual deployments inform its continuing research and product development, allowing what it learns from production use to improve Presence for customers over time.

This controlled improvement process explains how Presence is intended to maintain reliability after launch, but what evidence has OpenAI provided that the process works in an operating support service?

Presence powers OpenAI’s English-language phone-support service at 1-888-GPT-0090. The agent can respond beyond a fixed menu of predetermined questions by handling open-ended support requests, verifying callers, using relevant account information, and taking approved support actions.

OpenAI says the agent met or exceeded the company’s benchmarks for frontline human-support quality within weeks and now resolves 75% of inbound issues without human assistance. The company also reports that the Codex-powered improvement process reduced human handoffs by 15 percentage points over 10 days. These results are based on OpenAI’s own benchmarks, and the supplied sources do not provide the underlying data or measurement methods. The results have also not been independently verified.

Other companies are working with Presence at earlier stages. BBVA is a design partner helping OpenAI shape and refine voice experiences for financial customer service while exploring AI-powered voice support for everyday banking needs in Mexico. BBVA says it wants to provide a faster, more seamless, and more personalized customer experience, but it has not reported deployment results.

SoftBank is exploring agents that can communicate naturally in Japanese, connect to the processes needed to resolve customer requests, and represent the company consistently. SoftBank says its frontline teams rated the agent’s Japanese-language conversations highly for naturalness and accuracy. That assessment is qualitative, with no measurements or independent testing provided.

IAG is exploring customer agents that could help people navigate insurance claims and make support more readily available during high-demand periods, including severe weather and natural disasters. The company describes this as an opportunity it is pursuing and has not reported results from a completed deployment.

Presence has been used in OpenAI’s support operation and is being explored by other enterprises, but which companies can deploy it today?

Presence is available to eligible enterprise customers through a limited general availability program. Access depends on whether the proposed workflow fits the platform, whether the customer is ready to implement it, and whether OpenAI or its partners have the capacity to deliver the project. Presence is offered as a managed deployment and is not currently available as a self-serve product.

Deployments are led by OpenAI Forward Deployed Engineers, technical teams that work directly with customers to put the system into production, or by selected global systems integrators. OpenAI and the customer begin by defining the desired business outcomes, success criteria, and initial high-value workflows. Deployment teams then connect the required company knowledge and business systems, establish policies and permissions, set rules for when work must be escalated, and test the agent before bringing it into production.

If a proposed workflow goes beyond what Presence currently supports, OpenAI’s engineers and partners may work with the customer to bring that workflow into production. OpenAI and its selected systems integrators can then continue supporting the deployment as it expands.

OpenAI will continue giving voice customers access to its frontier models through the OpenAI API alongside Presence. The company describes that API access separately from a Presence deployment and does not present it as a substitute for Presence’s managed controls or deployment support.

During limited availability, the supported communication channel, contact-center connection, routing, customer authentication, and human-handoff process are determined separately for each deployment. Pricing and project scope are also specific to each customer rather than based on one publicly stated price or standard implementation package.

Presence could improve the customer experience. Anyone who has struggled through an automated phone menu or repeatedly asked for a human representative knows how frustrating those systems can be. If Presence works as intended, callers could explain their problem in their own words and either have it resolved or be transferred to a person when needed.

Interested organizations must contact their OpenAI account team to determine whether Presence is suitable and available for their proposed work. Presence gives enterprises a managed route to production voice and chat agents, but that route currently depends on eligibility and direct support from OpenAI or one of its selected deployment partners.

Q: What is OpenAI Presence?
A: OpenAI Presence is a managed enterprise platform for deploying production AI voice and chat agents. The agents can answer questions, use connected company systems, take approved actions, and escalate interactions to people when required.

Q: How is OpenAI Presence different from using the OpenAI API?
A: The OpenAI API gives voice customers access to the company’s frontier models. Presence adds the managed controls and deployment support surrounding those models, including company policies, permissions, evaluations, guardrails, escalation rules, and a process for improving agents after launch.

Q: How do companies control what Presence agents can access and do?
A: Each company defines the agent’s job and gives it only the knowledge, system access, and permissions required for that work. The company also determines which actions the agent can complete, which require approval, and when an interaction must be transferred to a person.

Q: How are OpenAI Presence agents tested before launch?
A: Presence deployments use simulations to recreate common requests, unexpected situations, and higher-risk scenarios before the agent serves real users. Graders evaluate whether the agent reaches the intended result, follows company policy, uses connected tools correctly, and escalates when required. Security, privacy, legal, and acceptance reviews are also completed before a controlled production rollout.

Q: How are Presence agents updated after they enter production?
A: Teams monitor production sessions, escalations, and other quality signals to identify performance gaps. Codex investigates those signals and suggests improvements, but it does not make changes independently. Company teams test each proposed update against the version currently serving users and decide which changes reach production.

Q: What results has OpenAI reported for Presence?
A: OpenAI reports that Presence resolves 75% of inbound issues for its English-language phone-support service and that its Codex-powered improvement process reduced human handoffs by 15 percentage points over 10 days. The supplied sources do not include the underlying data or measurement methods, and the results have not been independently verified.

Q: Can any company use OpenAI Presence?
A: No. Presence is available to eligible enterprise customers through a limited general availability program. Organizations must contact their OpenAI account team to determine whether their proposed workflow is suitable and whether deployment support is available. Pricing, project scope, system connections, and supported communication channels are determined separately for each customer.

OpenAI Presence gives enterprises a managed system for deploying production voice and chat agents that can perform approved work and adapt after launch. Companies retain authority over what those agents can access and do, when people must intervene, and which changes reach production.

OpenAI’s clearest evidence comes from its English-language phone-support service. The company reports that Presence resolves 75% of inbound issues without human intervention and that the Codex-powered improvement process reduced human handoffs by 15 percentage points over 10 days. The supplied sources contain no underlying measurement methods or independent verification, so those figures remain company-reported results.

Business leaders and teams responsible for customer support, outbound sales, billing, insurance claims, employee IT services, and other workflows involving company systems are most directly affected. They must decide which work an agent can handle, what authority it receives, and when a person must take over. Customers and employees are also affected when they rely on these agents to resolve problems or escalate their requests appropriately.

Voice and chat agents need to remain reliable in real-world interactions even as products, company policies, and user behavior change. Presence gives enterprises a controlled way to identify performance gaps, test suggested improvements, and update an agent while preserving company approval over which changes reach users.

Organizations evaluating Presence need to define the proposed workflow, determine which systems and information the agent may access, establish permitted actions and escalation rules, and maintain the ability to test updates after launch. They must also determine whether their project qualifies for the limited program and whether OpenAI or one of its selected partners can support the deployment.

In short, as enterprise voice and chat agents take on work involving company systems and real customer or employee needs, reliability will depend on the controls surrounding them. An enterprise voice or chat agent is only as dependable as the rules governing what it can do and how it can change.

For enterprises, Presence offers a controlled route to production voice and chat agents; for customers and employees, its value will be measured more simply: Did the agent solve the problem, or get them to a person who could?

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.