OpenAI Presence: when infrastructure becomes a service

On 22 July OpenAI announced Presence, an enterprise platform built to create and govern AI agents operating across voice and chat channels. The stated aim is to provide “a shared foundation for company context, policies, permissions, guardrails, actions and evaluations”, keeping behaviour consistent no matter which channel a customer uses. This is a meaningful shift: instead of supplying only the engine through an API, the company now supplies the chassis that engine is bolted to and controlled from, with built-in functions such as parsing open-ended requests, verifying a caller's identity and executing approved actions — refunds, billing fixes, account changes.

The performance figures released are striking, but they need to be read for what they are: self-reported data drawn exclusively from OpenAI's own customer service operation. On its English-language support line, OpenAI says Presence resolves 75% of incoming requests without human intervention, and that an improvement loop built on Codex cut handoffs to human agents by 15 percentage points in ten days. The process works like this: a Presence plugin analyses sessions and escalations to propose updates, which then have to be tested against the production version and approved by staff before release. There is no independent verification of these numbers so far, and no detail on the exact scope of the requests being counted.

Commercially, Presence is not a self-service product. The platform is in “limited general availability” for eligible enterprise customers, with deployments handled by OpenAI's Forward Deployed Engineers and selected system integrators. OpenAI has published no price list, saying that “during this limited GA phase, deployments are scoped individually based on each customer's use case and implementation needs. Broader pricing details to come as availability expands”. That approach moves OpenAI away from selling catalogue software and towards something much closer to high-value consulting, putting it in direct competition with many of the customer-experience vendors that were, until now, its own customers.

Among the first to try the platform are BBVA Mexico, SoftBank Corp. and IAG / Retail Insurance Australia — all described in the announcement as exploring or testing, not necessarily running it live. OpenAI has not said how many customers are actually in production. Tadahisa Murakami, VP at SoftBank, noted that “frontline teams have rated the agent's Japanese-language conversations highly for their natural and accurate quality”. The spread of AI into customer service, though, is not obstacle-free: as The Register pointed out, Gartner predicts that by 2027 half of the organisations planning AI-led customer service will abandon those plans — context that turns OpenAI's “high-touch” move into a bet on human supervision being the thing that makes automation genuinely dependable.

Selling picks and shovels was the easy part; now OpenAI is going down the mine on someone else's behalf, hoping the ground is solid enough to justify the consulting fee. The numbers look promising, but until we see reports from the people actually swinging the picks, the line between hype and substance remains a matter of trust.

— Olya