OpenAI launched Presence today, a purpose-built enterprise platform that lets companies deploy AI agents across voice and chat with policy controls, identity verification, and a built-in governance layer. The launch signals OpenAI’s most direct move yet into the operational AI stack that large enterprises actually use to run their customer-facing workflows.

What Presence Actually Does

Presence is not a raw API product. It sits above the model layer and gives enterprise teams a structured environment to define how agents behave, enforce policies, evaluate performance, and improve over time without requiring constant engineering intervention.

The platform brings six core capabilities together:

Policies and standard operating procedures — Teams can codify their existing rules, escalation paths, and compliance requirements directly into how agents behave, not just in a system prompt that might get ignored.

Guardrails — Hard limits on what agents can say and do, enforced at runtime. Agents cannot take actions outside the approved set regardless of what a customer says.

Approved actions — Agents can verify identity, retrieve account data, apply billing resolutions, and issue refunds, but only within explicitly approved boundaries. The action list is defined and locked before deployment.

Simulations — Before going live, teams can test agents against scenarios to catch failure modes. This closes one of the biggest reliability gaps in enterprise AI deployments: agents that look fine in demos but break on edge cases in production.

Evaluation tools — Ongoing performance measurement that tracks accuracy, policy adherence, and escalation rates after launch.

Codex-powered improvement — After launch, a Codex-based process analyzes real conversations to surface patterns, identify gaps, and suggest rule updates. Agents get better without needing manual review of thousands of transcripts.

Together these capabilities address the reliability gap that has held back serious enterprise deployment. Most companies have run AI agent pilots. Far fewer have moved them into production workflows where failures actually cost money or damage customer relationships.

Who Is Already Using It

OpenAI announced three anchor customers at launch.

BBVA Mexico is deploying Presence for customer interactions, with the platform connecting to account systems so agents can resolve issues without escalation to human agents.

SoftBank Corp. has deployed Japanese-language voice agents. SoftBank says the conversations are natural and accurate, which matters considerably more in Japanese than in English given the precision the language demands and the expectations Japanese customers bring to service interactions.

Retail Insurance Australia, part of IAG, is using Presence for customer-facing support, one of the highest-stakes environments for voice AI given the compliance requirements in financial services.

Not a Self-Service Product

This is worth understanding before anyone tries to sign up. Presence is not available through the standard OpenAI API dashboard. Deployments are led by OpenAI Forward Deployed Engineers or select global systems integrators.

That deployment model connects directly to the OpenAI Deployment Company, a subsidiary formed in May 2026 with $4 billion in backing at a $10 billion valuation. OpenAI built that entity specifically to handle enterprise implementation work at scale. Presence is the product that entity exists to deploy.

The practical implication is that Presence is not a tool for a team of three trying to add an AI agent to their help desk. It is an enterprise contract product aimed at organizations with meaningful volume in customer voice and chat workflows.

What This Means for Business

Presence is market validation for a thesis that has been building across the AI vendor landscape: enterprises need more than a capable model. They need a layer that sits between the model and their operations, enforcing the rules that make AI trustworthy enough to hand customer conversations to.

For companies evaluating AI agents for their customer-facing workflows, a few things follow from today’s launch:

Governance is now table stakes. Every serious enterprise AI platform is adding policy controls, guardrails, and audit trails. If a vendor cannot answer how an agent gets constrained when it starts behaving incorrectly, that is a gap that will cause problems in production.

Voice AI is entering the enterprise mainstream. The SoftBank and IAG deployments are not experiments. They are production systems. The industry has crossed from proof-of-concept into deployment at scale, and the pace of that shift is accelerating.

Implementation model matters as much as technology. OpenAI choosing to deploy Presence through Forward Deployed Engineers rather than self-service reflects the reality that the hardest part of enterprise AI is not the model. It is connecting the model to the right data, defining the right policies, and building the evaluation loop that keeps quality high over time.

For businesses that are not yet at enterprise scale but are watching where this goes, the pattern holds at smaller sizes. Any AI voice or chat deployment that relies purely on the model to figure out what it should and should not do will eventually produce an outcome that reflects the lack of explicit controls.

The gap between “the model is impressive” and “the model is reliably useful in production” is exactly what platforms like Presence are built to close.

---

Enterprise DNA builds AI agent systems for businesses across operations, voice, and applications. If you are working through how to put AI agents into your workflows in a way that is actually reliable, book a discovery call to talk through what that looks like in practice.