OpenAI is selling its newest enterprise agent offering as a managed deployment rather than a self-service product. Presence combines models, forward-deployed engineers and selected systems integrators to implement one tightly scoped workflow at a time.
That delivery model acknowledges what many agent projects have discovered: production failures usually emerge from permissions, integration and operating policy, not from an inability to generate convincing language.
Presence starts with one job and explicit boundaries
An engagement may focus on resolving a billing dispute, handling an insurance claim or completing an employee IT request. The agent receives only the knowledge and system access needed for that job. Customers define its permitted actions, approval points and conditions for human takeover.
OpenAI describes a six-stage process covering business outcomes, security, privacy and legal review, simulation, acceptance testing, staged rollout and post-launch improvement. The documentation states directly that ingesting company documents does not make an agent ready for production.
Simulations and graders assess whether the system reached the correct outcome, followed policy, used tools appropriately and escalated at the right moment. Session histories and action records support auditing, while controlled releases and rollback reduce the impact of a bad version.
After launch, Codex reviews production sessions and escalations and proposes changes. A customer team must test and approve those changes before deployment. That creates an improvement loop without allowing an agent to rewrite its own operating rules unchecked.
The approach aligns with Gartner’s warning that more than 40% of agentic projects may be cancelled by the end of 2027 because of weak governance, unclear value and poor operational discipline. Presence bundles implementation work around exactly those failure points.
Embedded engineers create both value and scarcity
OpenAI’s Forward Deployed Engineers work inside customer environments, following a model associated with long, operationally intensive enterprise projects. This can accelerate integration because the model provider sees actual constraints instead of handing a dashboard to the buyer.
It also limits scale. Software can serve another customer instantly; engineers approved to work with a bank’s core systems cannot. OpenAI says eligibility depends on workflow fit, readiness and available delivery capacity. Presence is therefore rationed partly by consulting resources.
The model provider is also the implementation partner, which can blur accountability. Contracts should specify who owns policy design, integration defects, model changes and production incidents. Customers need the ability to inspect logs, export evaluation data and preserve an exit path if they later change provider.
Pricing is not public and is set per engagement. The underlying model is also unspecified because OpenAI intends to adjust configuration as a workflow evolves. Flexibility can improve performance, but agreements should define notice, regression testing and acceptance criteria whenever the model changes.
Voice and chat are supported during limited general availability, while contact-centre integration, identity and handoff are designed case by case. The signed architecture and contract—not a universal product setting—govern data handling.
Early results are promising but remain vendor evidence
OpenAI points to its own English-language phone-support agent as the strongest example. It says the system resolves 75% of inbound issues without a person and reduced handoffs by 15 percentage points in ten days through the Codex improvement loop.
Those figures use OpenAI’s channel, grading and reporting and have not been independently verified. Named customers are earlier in the cycle: BBVA is exploring voice support in Mexico, SoftBank is testing Japanese conversations and IAG is examining high-demand events such as severe weather. None is described as operating Presence at broad scale.
Buyers should measure cost per correctly resolved case, customer satisfaction, escalation quality and the effort required to maintain policies. Autonomous resolution is valuable only if it does not hide repeat contacts, incorrect concessions or frustrated users who could not reach a person.
Presence sits alongside self-service Workspace Agents and API-based voice systems. The core capability may be similar; the difference is who performs the difficult deployment work.
That makes Presence an honest enterprise proposition. OpenAI is not pretending an API key completes organisational change. Its challenge is proving that a labour-intensive managed service can deliver repeatable economics while leaving customers with genuine governance—not simply a successful project that remains dependent on the engineers who installed it.
