OpenAI Launches Presence, a Consulting-Style Product for Running AI Agents in Production
Not self-serve. OpenAI's own engineers run a six-stage process to get a voice or chat agent into production, then a system called Codex watches it run.
OpenAI's new AI agent product for enterprises is not self-serve. It's closer to consulting.
OpenAI announced Presence on July 22, 2026, a managed product for putting AI agents into production for real businesses, across voice and chat. It is not self-serve.
How it actually works
Delivery runs through OpenAI's own Forward Deployed Engineers plus a small number of selected systems integrators, not a signup form. The process runs in six stages: scoping the business outcome, a security and legal review, simulation testing, acceptance testing, a staged rollout, then post-launch iteration. After launch, a system called Codex reads production sessions and escalations, then proposes changes that the customer's own team tests and approves before anything ships.
Who's using it
Named early customers include BBVA, exploring voice support for everyday banking in Mexico, SoftBank, testing Japanese-language conversations that a SoftBank VP says the team rated highly for natural, accurate quality, and IAG, exploring support during high-demand events like severe weather.
The bigger pattern
Presence extends what OpenAI calls Workspace Agents, its existing customizable GPTs mostly built for internal use, into something meant for real customer-facing traffic. It follows OpenAI's Deployment Company, a separate consulting arm that launched in May 2026 with backing from 19 investors and delivery partners including Bain, McKinsey, and Capgemini. Presence isn't described anywhere as literally built on top of the Deployment Company, but the two share the same underlying bet: that getting an AI agent to work reliably in a real business is still mostly a services problem, not a product one.
Why a build studio cares
This is OpenAI admitting, with its own pricing model, that shipping an agent into a real business isn't a weekend integration. The six-stage process it built for its own enterprise customers is close to the same shape of work we do for clients: scope the job narrowly, review it for security and legal exposure, test it against real scenarios, then roll it out in stages instead of all at once.
Next step: read The Decoder's coverage or The Register's writeup. If you're weighing how to get an agent into production safely, write to us at hello@gattyworks.com.