共有:

Enterprise Voice Agents

OpenAI just entered the voice agent war.

The "build layer" for voice agents that ElevenLabs and Vapi pioneered is now something OpenAI wants to sell as a managed service of its own. Here's a close read of "OpenAI Presence," the company's third voice-related move this month alone.

AI Navigate Editorial2026.07.236 min read

Customer inquiry OpenAI Presence Policy Access Escalate Test → go live Resolved 75% To a human 25%
01

What Happened

From selling models
to running agents

OpenAI unveiled a new enterprise platform called "OpenAI Presence."

In its official announcement, OpenAI introduced "OpenAI Presence," a new platform for building and operating realtime voice and chat AI agents for enterprises. Company policies, standard operating procedures, permission controls, the scope of actions an agent may take, and the conditions for escalating to a human are bundled into a single configuration, then run through pre-deployment simulations covering common requests and high-risk edge cases before going live — a workflow OpenAI itself lays out in its help-center article.

This is not an isolated feature drop. In July alone, OpenAI shipped a lower-latency voice model, "GPT-Realtime-2.1," on the 6th (cutting p95 latency by more than 25%), the full-duplex "GPT-Live" on the 8th, and now Presence on the 22nd — three voice-related announcements in three consecutive weeks. The pattern suggests OpenAI's ambition has shifted up a level, from competing on raw model quality to absorbing the entire "build-and-operate" layer around voice agents.

02

By the Numbers

Presence, by the numbers

75%
Auto-resolution rate on OpenAI's own support line
+15pt
Cut in human handoff rate within 10 days via the improvement loop
~150
Deployment engineers added via the Tomoro acquisition

Presence already runs on OpenAI's own English-language support line, 1-888-GPT-0090, where it handles identity verification, account lookups, and approved actions such as refunds — and, per OpenAI, resolves 75% of inbound issues without human involvement. A "Codex-powered improvement loop" reportedly cut the human handoff rate by 15 percentage points in just 10 days. Those figures are company-reported, though, and no independent verification is available yet — worth keeping in mind.

Deployment is not self-serve. It's led by OpenAI's own Forward Deployed Engineers together with staff from Tomoro, the consulting firm OpenAI acquired in May — roughly 150 people in total. Tomoro previously built an in-game AI support system for the game company Supercell that served 110 million users within 12 weeks of launch. Presence looks like that playbook repackaged as an OpenAI-branded product.

03

How It Rolls Out

Rollout runs through three sales-led stages

This isn't a matter of grabbing an API key. An OpenAI team works alongside each enterprise.

Pick a workflow FDEs and the customer identify high-value tasks Connect access Wire up internal systems and policy rules Test, then go live Run high-risk cases before production
FIG. Forward Deployed Engineers and the customer walk through three stages together before go-live
01

Pick a workflow

OpenAI's FDEs and the customer identify a high-value workflow, such as call-center support or internal help desk.

02

Wire up access and policy

They configure what internal systems the agent can reach, what actions it may take, and when it must escalate to a person.

03

Test before shipping

Common requests and high-risk scenarios are run through pre-deployment simulations; only after passing does it go into production and keep improving.

04

Who It Hits

Who it affects, and how

Engineers

If you build your own voice pipeline, calling low-latency GPT-Realtime-2.1 directly is still an option. Presence isn't an API product, though — it's an FDE-led engagement. You hand off permission and escalation design entirely, but give up fine-grained control in return.

Business leaders

Pricing isn't public — deployments are "scoped individually." Unlike metered SaaS from ElevenLabs or Vapi, this looks closer to a consulting engagement to procure. Early customers named so far — BBVA Mexico, SoftBank Corp., and Australian insurer IAG — are the reference cases worth weighing for ROI.

PMs weighing in-housing

If you're evaluating whether to build call-center automation in-house, this is one more option on the table. If you're just chatting with ChatGPT by voice as an individual, none of this changes anything for you — Presence targets enterprise business channels specifically.

Build it yourselfOpenAI Presence
Contract STT/TTS separately — GPT-Realtime family, ElevenLabs, etc.GPT-Live and the Realtime family bundled into one contract
Your own engineers design permissions and escalation~150 FDEs design and run it alongside you
Self-serve API — start todayLimited general availability, sales-led, enterprise only
Metered pricing you can estimateUndisclosed, scoped per deal

What's being sold isn't the model — it's operating it.


05

What's Next

What to watch next, and what to discount

Three things are worth tracking in the near term. (1) Whether results from early customers — BBVA Mexico, SoftBank, and IAG — become independently verifiable, (2) whether Presence eventually goes self-serve and competes on price directly with ElevenLabs and Vapi, and (3) whether wider API access to GPT-Live and GPT-Realtime-2.1 lowers the bar for building this in-house without Presence at all.

A few things are worth discounting, too. The 75% resolution rate and the 15-point improvement are both OpenAI's own figures, and no third-party replication exists yet. The FDE-led, non-self-serve rollout also means cost structure stays opaque as adoption scales, and it's worth noting that industry commentary has grown more pointed about how far this kind of automation displaces human jobs. How much scaled automated response replaces human labor is a question that can't be separated from how Presence should be judged.