OpenAI’s New Voice AI Is a Leap Forward. It Still Skips Small Service Businesses

4 min read

OpenAI released GPT-Realtime-2 in its API in May 2026, then shipped a faster follow-up version in July. Separately, the company introduced Presence, a platform for deploying voice and chat agents at large enterprises. Neither is something a five-person HVAC company can sign up for and start using this week, and that gap matters more than the announcement itself.

What OpenAI actually shipped

GPT-Realtime-2 is described by OpenAI as its most capable voice model yet, built to hold a 128K-token context window and score 96.6% on the Big Bench Audio benchmark. In plain terms: it can track a longer, more complicated conversation without losing the thread, and reason about what a caller is actually asking for rather than matching keywords. A July update, gpt-realtime-2.1, cut response latency further, which matters a lot on a phone call where a half-second pause reads as dead air.

Presence, OpenAI’s enterprise deployment platform, is a separate thing: a way for large organizations to launch and manage voice and chat agents with OpenAI’s own deployment engineers involved. Coverage from VentureBeat notes it is already used internally by OpenAI for English-language phone support and by a small set of large companies including BBVA and SoftBank.

Why this is happening now

The timing lines up with where customer service leadership already is. A Gartner survey of 321 customer service and support leaders, fielded in October 2025, found 91% reported pressure from executive leadership to implement AI in 2026. That pressure is aimed at large service organizations first, the kind with a Gartner subscription and a dedicated ops team, not a two-truck plumbing outfit. But the underlying reason for the pressure, callers expecting an instant, competent response, applies just as much to a small business as a large one.

The limitation that matters most

Presence is not self-serve. OpenAI’s own coverage describes deployments as led by OpenAI’s Forward Deployed Engineers or select systems integrators, aimed at enterprise accounts. GPT-Realtime-2 by itself is an API, not a phone system: it has no way to answer a real call, check a calendar, or book an appointment on its own. Someone still has to build the qualifying questions, connect it to a calendar and CRM, define what it should never do without a human, and handle the actual telephony.

That build gap is exactly where a smaller service business gets stuck. The model getting smarter does not remove the work of turning a model into a working front desk.

Norrsyn’s view

Better underlying voice models are good news for the industry: less awkward pauses, fewer misheard requests, more natural handoffs to a human. But the model is the engine, not the car. The default position we build from is that AI should make a business easier to reach, run, and trust, automating delays and repetitive administration while keeping judgment, accountability, and the sensitive parts of customer relationships with people. A more capable model doesn’t change that division of labor, it just makes automating the reachable part work better.

If you’re evaluating a voice AI vendor

• Ask which underlying model they use and how often they update it, since this space is moving fast.
• Ask what happens when the AI hits something it can’t or shouldn’t handle: does it escalate cleanly to a human, or does the caller get stuck?
• Ask what data it needs from your calendar and CRM to actually book something, not just take a message.
• Get a real number on missed-call recovery from their existing clients, not just a capability demo.