OpenAI Presence: own the agent config, not the black box

OpenAI shipped Presence on July 22, and the pitch is not another chatbot API. It's a managed system for standing up production voice and chat agents — with OpenAI's own engineers embedded in your ops to wire it up. OpenAI says it already resolves 75% of inbound issues without a human. That's a real capability. It's also a control plane you rent, and the difference matters more than the demo.

What actually happened

Per OpenAI's announcement and VentureBeat, Presence bundles the parts you'd otherwise assemble yourself: company knowledge, standard operating procedures, approved actions, permissions, simulations, evaluation tools, guardrails, and escalation rules. Each agent is scoped to only the systems a given workflow needs, and admins keep control over operational boundaries, approval thresholds, and human-handoff criteria.

The delivery model is the tell. This runs through the OpenAI Deployment Company — the same forward-deployed arm behind the Northslope acquisition — so you get engineers on-site, not a self-serve dashboard. Launch customers are enterprise-scale: BBVA testing voice banking in Mexico, SoftBank running Japanese-language support, and IAG's Retail Insurance Australia on customer-facing service. OpenAI reports its Codex-powered improvement loop cut human handoffs by 15 percentage points in 10 days.

It's not alone. The same buyer is being courted by Google's Gemini Enterprise, Meta's business agent platform, and the NVIDIA–ServiceNow tie-up. Four vendors, one control plane.

Why the agent platform war matters for your business

Notice what the platform actually contains. The model is a commodity input. The value — and the lock-in — is the config: your SOPs encoded as approved actions, your permission scopes, your guardrails, your evaluation sets, your escalation rules. That's the asset. Whoever holds it holds your customer-facing operations.

For a small business, the high-touch package isn't for you anyway; forward-deployed engineers are priced above your line. But the thing they're building — a governed agent that answers the phone, knows your process, and hands off cleanly — is not exotic. It's a well-scoped prompt library, a permissions map, a handoff rule, and an eval set, sitting on components you can point at whichever model clears your bar.

So take the lesson, not the invoice. Buy the outcome where it makes sense, but own the config. If your agent's SOPs, guardrails, and escalation logic live in a vendor's platform, moving is a rebuild. If they live in your repo behind a swappable model, moving is a Tuesday. The platform war is a fight over who owns your control plane — make sure the answer is you.

Sources: OpenAI, VentureBeat.