OpenAI Presence — Vendor-Managed Agents, Not Software

OpenAI announced Presence on July 22, 2026 — a managed platform for deploying enterprise AI agents across voice and chat channels. The headline feature is not the model powering it (GPT-5.6) or the Codex-driven improvement loop. It is the delivery model: Presence is not self-serve. Every deployment is led by OpenAI Forward Deployed Engineers and select global systems integrators, with no announced path to self-managed access. You are not buying software. You are buying a deployment engagement with your model vendor — and that distinction changes every calculation in your procurement process.

TL;DR

• What: OpenAI launched Presence, a managed enterprise agent platform for voice and chat workflows — no self-serve access
• How: Every deployment led by OpenAI Forward Deployed Engineers; Codex proposes behavioral updates post-launch, staff approve before rollout
• Signal: OpenAI is building a professional services business with a product wrapper — governance infrastructure runs inside their infrastructure, not yours
• Action: If you are evaluating enterprise agent platforms, treat Presence as a managed service engagement, not a software purchase — and price your exit accordingly

What Happened

Presence targets large organizations deploying AI agents for customer support, sales, and internal operations. Each deployment starts with a single, narrowly scoped task — resolving billing issues, handling insurance claims, triaging IT service requests. The agent receives only the knowledge and system access required for that specific job. After launch, production session data and escalations feed into a Codex-powered loop that proposes behavioral updates. Teams test each proposed change against the version currently in production, then approve a controlled rollout.

OpenAI is its own first customer. Presence powers the company’s English-language phone support at 1-888-GPT-0090, handling open-ended inbound requests, verifying callers, pulling account context, and taking approved actions. OpenAI reports that the system resolves 75% of inbound issues without human assistance and that its Codex improvement loop reduced human handoffs by 15 percentage points in just 10 days.

The 75% resolution rate and 15-percentage-point improvement in handoffs are company-reported figures — not independently verified. Treat them as indicative, not as benchmarks you can hold OpenAI to in a contract.

The platform runs on GPT-5.6, the model family OpenAI released last month. Early customers exploring Presence include BBVA (voice support for routine banking in Mexico), SoftBank (Japanese-language customer conversations), and IAG, the Australian insurer testing support during high-demand periods like severe weather events.

None of this is available for you to sign up for today. OpenAI is rolling it out through what it calls a “limited general availability program.” The delivery vehicle is the OpenAI Deployment Company, a separate entity established on May 11, 2026 with more than $4 billion in initial investment. TPG leads, with Advent, Bain Capital, and Brookfield as co-lead founding partners. Bain & Company, Capgemini, and McKinsey are among 19 global partners. OpenAI acquired Tomoro to staff the Deployment Company with approximately 150 experienced Forward Deployed Engineers and Deployment Specialists from day one.

Why This Matters

I have watched every major AI lab launch “enterprise agent governance” in 2026 — AWS AgentCore, Anthropic Managed Agents, Microsoft Agent 365, Google Gemini Enterprise. Every one claimed to be a platform. OpenAI Presence is the most honest about what it actually is: a professional services business with a product wrapper.

That honesty matters because the FDE-led delivery model means OpenAI owns the implementation relationship. Not on day one while you are getting started, but permanently. There is no announced graduation path where your team takes over operations and runs Presence independently. The Codex improvement loop — the mechanism that makes agents get better over time — runs inside OpenAI’s infrastructure. Your staff approves the updates Codex proposes, but Codex does the investigating, the pattern recognition across sessions, and the generation of proposed changes. You are a reviewer in someone else’s CI pipeline.

Compare this to the competing approaches. AWS AgentCore gives you managed infrastructure but lets your team control agent logic and deployment. Anthropic’s managed agents run on Anthropic’s models but provide API-level access for teams that want to own orchestration. Microsoft’s Agent 365 sits inside a broader enterprise stack where governance distributes across existing IT controls. Presence is the only one where the vendor’s professional services team is a structural dependency, not a bootstrap accelerator.

The narrow task scoping reinforces this dynamic. Starting with one specific job per deployment is a legitimate governance pattern — it limits blast radius and makes behavior predictable. But it also means expanding scope requires re-engaging OpenAI. Want to extend your billing-issue agent to handle refund approvals? That is a new deployment engagement, not a configuration change.

If you are evaluating Presence, the key procurement question is not “can it handle our use case?” It is “what does it cost to leave after 18 months?” Map every governance component — policies, guardrails, escalation rules, the Codex improvement loop — to where it runs. If the answer is “inside OpenAI’s infrastructure” for all of them, your exit cost is rebuilding the entire governance layer.

The $4 billion Deployment Company backing tells you where OpenAI thinks the margin is. Not in API calls. Not in model access. In owning the implementation relationship with enterprises that cannot easily switch once governance, improvement loops, and institutional knowledge accumulate inside OpenAI’s systems. TPG and Bain Capital did not invest $4 billion in a software product. They invested in a recurring professional services revenue stream with software-grade margins.

This is the Accenture playbook with a model vendor’s structural advantages. Accenture locks in through institutional knowledge and switching costs. OpenAI locks in through institutional knowledge, switching costs, and the fact that the improvement engine runs on their proprietary model inside their infrastructure. The governance layer that makes Presence useful is the same layer that binds you to OpenAI’s roadmap and pricing decisions.

The Take

Presence is not a bad product. The 75% autonomous resolution rate on OpenAI’s own support line — if accurate — suggests the technology works for narrowly scoped, high-volume customer interactions. The Codex improvement loop is a genuinely clever feedback mechanism. And the narrow-task-scoping approach is more honest than competitors who promise general-purpose agents and deliver brittle demos.

But the procurement decision here is fundamentally different from buying software or even buying API access. You are entering a managed relationship where your vendor controls the improvement cycle, the governance infrastructure, and the deployment methodology. Your engineering team approves changes but does not drive them. That is the trade-off, stated plainly.

If your organization has high-volume, narrowly scoped customer interactions and your procurement team is comfortable with a managed vendor relationship that has no announced self-serve path, Presence is worth evaluating. If your engineering leadership wants to own the agent stack — the logic, the governance, the improvement loop — this is the wrong product category entirely. Look at self-hosted options or platforms that give you API-level control over agent behavior.

The question I would ask in any Presence evaluation meeting: “Show me the exit plan.” Not because you plan to leave, but because the answer tells you exactly how much control you are giving up. If OpenAI cannot describe what a migration away from Presence looks like, that is your answer about who is really in the driver’s seat.

Related

• Anthropic Managed Agents — Runtime Lock-In: Anthropic’s competing approach uses managed runtime but preserves more API-level control for engineering teams
• AWS AgentCore — Managed Harness Lock-In: AWS takes the infrastructure-managed approach while letting teams own agent logic — the structural inverse of Presence
• Microsoft Agent 365 — Enterprise Control Plane: Microsoft distributes governance across existing enterprise IT controls rather than centralizing in a vendor-managed engagement