- OpenAI launched Presence on July 22, 2026 as a managed enterprise product for deploying AI agents in production workflows, pairing Forward Deployed Engineers and global systems integrators with each engagement rather than leaving implementation to customer teams.
- Each Presence deployment is scoped to a specific job, provisioned with least-privilege system access, and governed by explicit policies, guardrails and escalation rules, addressing the governance gaps that cause most enterprise agent projects to stall before production.
- A Codex-powered improvement loop analyses production sessions and escalations to propose updates, which enterprise teams review and approve before rollout, keeping agents current as policies and product conditions change.
OpenAI’s Presence, launched July 22, 2026, is the company’s clearest signal yet that selling API access is not enough for the enterprise market. The managed product deploys voice and chat agents directly into production workflows, with OpenAI‘s own Forward Deployed Engineers and global systems integrators running the implementation. BBVA Mexico, SoftBank and Insurance Australia Group are reportedly among the early testers.
Trust Built In From the Start
Every Presence engagement starts with a specific job: resolve billing disputes, handle insurance claims, manage IT service requests. The agent is provisioned with only the knowledge and system access that job requires, least-privilege by design. From there, the enterprise defines what the agent can do autonomously, what needs human approval, and when a human operator takes over entirely.
Before go-live, agents run through detailed simulations and evaluation suites covering both routine queries and edge cases. This is where Presence differs most sharply from generic LLM deployments, which typically bolt governance on after the fact. Policies, standard operating procedures, guardrails and approved action frameworks are built into the deployment from the first engagement, not retrofitted once problems surface in production. For teams that have hit the wall on agent reliability issues in live environments that ordering matters.
What “Managed” Actually Means Here
The computational burden for scalable inference sits with OpenAI’s cloud infrastructure. Enterprises avoid direct capital expenditure on AI accelerators and the operational overhead of managing inference clusters, though OpenAI has not publicly disclosed the specific infrastructure details behind Presence. What this means in practice: no fragmented API integrations, no in-house latency tuning, no separate hardware contracts. OpenAI owns the stack from model to deployment.
That is a direct challenge to the system integrator model that has dominated enterprise AI rollouts. Rather than an enterprise assembling its own toolchain from LLM APIs, orchestration frameworks and third-party evaluation tools, Presence packages all of it under one vendor relationship. Whether that consolidation is a feature or a lock-in risk depends on how much operational control a given enterprise wants to retain. The hidden costs of fragmented AI automation workflows are real, but so is vendor dependency.
The Codex Feedback Loop
Once an agent is live, production sessions and escalations feed into a continuous improvement process. Codex analyses those interactions, identifies performance gaps and proposes updates. Enterprise teams review and approve changes before anything rolls out. Agents stay current as product lines change, policies update and customer behaviour shifts.
Traditional software release cycles are too slow for this. A customer support agent trained on last quarter’s return policy is actively harmful when that policy changes. The human-approval gate in the Codex loop is the part worth watching closely: it preserves oversight while removing the engineering bottleneck of manual retraining cycles. How well it scales across dozens of concurrent agent deployments is an open question, but the architecture is sensible. Teams exploring RAG-augmented agent knowledge management will recognise the underlying challenge Codex is trying to solve here.
API-First No More
OpenAI built its enterprise business on API access and SDKs. Presence is a different motion entirely. OpenAI’s engineers identify workflows, connect systems, establish permissions, test agents and push them into production alongside the customer. That is a professional services model, not a developer tools model.
The competitive target is obvious: traditional enterprise software vendors and the large systems integrators that have historically owned complex deployments. Gartner has warned that over 40% of enterprise AI agent projects fail before reaching production, and Presence is explicitly positioned at that failure rate. BBVA Mexico, SoftBank and Insurance Australia Group are reportedly exploring the product, though OpenAI has not confirmed deal terms. If the model works, the monetisation logic is straightforward: enterprises pay a premium for outcomes, not tokens. That shifts OpenAI’s revenue from consumption pricing toward something closer to enterprise software contracts, with all the stickiness and margin that implies.
