OpenAI Presence: even OpenAI won’t sell its AI agents self-service — what it means for your SME

On July 22, 2026, OpenAI launched Presence, its platform for enterprise voice and chat AI agents — customer support, sales, sensitive internal workflows. The detail that stands out is not the technology itself: unlike ChatGPT Enterprise, Presence is not available self-service. Every deployment is run by OpenAI’s Forward Deployed Engineers or by selected systems integrators, who work with the customer to define the scope of access to internal systems (CRM, ERP), the rules for escalating to a human, and a Codex-driven continuous improvement loop. Three early customers are named: BBVA (voice banking support in Mexico), SoftBank (Japanese-language conversations), and insurer IAG (handling demand spikes during severe-weather events).

The number that justifies the hand-holding

OpenAI runs its own English-language support line on Presence, resolving 75% of inbound calls without human involvement — and the Codex-driven improvement loop cut human handoffs by 15 percentage points in ten days. OpenAI’s own figures have been independently confirmed by several trade outlets. But The Register points out the flip side: with no public pricing and rationed by the number of engineers available, Presence stays, for now, reserved for organizations the size of BBVA or SoftBank.

What the research on oversight confirms

OpenAI’s choice tracks what academic research on AI governance documents. A 2025 study by Kitty Kioskli and co-authors, trustSense: Measuring Human Oversight Maturity for Trustworthy AI, published in the journal Computers, proposes a framework for measuring how mature an AI system’s human oversight actually is: explicit review checkpoints, documented escalation channels, and a real ability to interrupt or correct the agent while it runs. The authors find that this oversight maturity, more than the underlying model’s raw capability, is what determines whether an AI deployment stays reliable once it reaches production.

What it means for your SME

An SME does not need to reach BBVA’s size to apply the same principles at its own scale. Whether the project is an AI support desk, a sales agent, or automated email sorting, the question to ask any provider is not "can your agent do X?" but "what exactly can it access, and what happens when it doesn’t know the answer?"

• →Before signing, ask for the precise list of systems and actions the agent will have access to — not a generic description of its capabilities.
• →Get the human escalation rules in writing: which cases the agent hands off, to whom, and within what timeframe.
• →Track an oversight metric (escalation rate, correction rate after go-live) before expanding the agent’s scope to new cases.
• →Don’t mistake a generic no-code tool for a supervised production agent — the difference doesn’t show up in a demo, it shows up in the correction loop after launch.

The most useful signal from this launch isn’t technical — OpenAI, which has every incentive to sell volume, still chose hands-on deployment over self-service for its own enterprise agents. That’s confirmation, from the largest player in the market, that an AI agent that holds up in production is built with a defined scope and documented oversight, not with one more subscription.