OpenAI introduces Presence, a product designed so companies can deploy reliable AI agents that do high-value work without losing control. Sound familiar — powerful models in production that you don't fully trust? Presence aims to solve exactly that with systems, rules and a continuous improvement loop.
What is OpenAI Presence
Presence isn't just another model: it's a packaged solution that combines model reasoning with policies, guardrails and escalation rules. It's built to answer questions, solve problems, use internal systems, execute approved actions and ask for human help when needed.
Each deployment starts with a specific job: handling billing, processing an insurance claim, or managing internal IT requests. The agent only gets the knowledge and access needed for that task. Your company defines what the agent can do, when it needs approval and when it should hand off to a person.
Presence and control: Presence lets the AI act in your workflow, but within clear and verifiable limits.
How it works in practice
- Targeted integration: OpenAI and its teams identify high-value workflows and connect the necessary systems, data and permissions.
- Policies and guardrails: your company decides the operational and security rules; Presence enforces those rules on every interaction.
- Simulations and testing: before launch it’s tested against common requests, edge cases and risk scenarios.
- Improvement cycle: production signals and escalations reveal gaps. A component driven by
Codexproposes changes that teams can test and approve for gradual rollout.
Result? An agent that doesn't stay static: it learns alongside your products, policies and customers, without you losing control.
Use cases and real results
Presence is already used in real-time voice and chat experiences: customer support, outbound sales and high-risk internal flows. OpenAI uses it on their English phone line (1-888-GPT-0090), where the agent understands open requests, verifies the caller, checks account context and performs approved actions.
In a few weeks it reached or exceeded human-quality benchmarks for support and currently resolves 75% of incoming incidents without human intervention. Thanks to the improvement loop with Codex, the escalation rate to humans dropped 15 percentage points in just 10 days.
Some partners and early trials:
- BBVA is exploring voice support for banking needs in Mexico.
- SoftBank is testing natural conversations in Japanese.
- IAG is studying support during high-demand events like weather-related disruptions.
Control, security and continuous evaluation
Before exposing the agent to real users, Presence lets you test whether the agent reached the correct outcome, followed policy, used tools properly and escalated when necessary. Guardrails kick in when an interaction goes beyond the defined limits.
After launch, production sessions and quality signals show where the agent works and where it needs tweaks. The Codex plugin investigates those signals and proposes updates your team can test against the production version and approve in controlled rollouts.
Availability and what you should consider
Presence is available to eligible enterprise customers through a limited general availability program. Deployments are led by OpenAI's Forward Deployed Engineers (FDEs) and certain global integrators; it's not a self-serve product yet.
If your organization needs capabilities Presence doesn't cover today, the FDEs and partners can collaborate to bring that case to production. OpenAI will also continue offering access to frontier models via the API for voice customers.
Presence looks like a toolkit to bring AI agents into production responsibly and at scale. Worth exploring? If you have repeatable workflows, data and a need for control, it can significantly speed up the path from proof-of-concept to reliable operations.