What happened.

OpenAI announces Presence, a platform for deploying voice and chat agents in enterprise contexts, for both customer-facing and internal workflows. The announcement emphasizes “trusted” agents and brings together voice and text capabilities in a single product.

Why it matters.

OpenAI is formalizing the shift from model to agentic platform. The signal has been clear for weeks: operational documentation for teams on ChatGPT Work and the small business program. Presence is the layer where it all comes together: the agent becomes a product that answers the phone and manages workflows, backed by an enterprise contract and a single vendor.

For those using AI in their field, the skill that matters now is designing agentic workflows: defining what the agent can do, where it stops, how you verify it. OpenAI provides the infrastructure. Responsibility for objectives, limits, and controls stays with the user.

Key point to watch: the announcement is thin on technical details. No public pricing, no architecture. The word “trusted” carries most of the weight here, and it’s what has least to verify right now.

In detail

Until yesterday, if you wanted an enterprise agent with OpenAI you assembled the pieces yourself: the model API, an orchestration framework, a voice layer (Twilio, LiveKit or similar), and you wrote the security controls. Each piece needed to be selected, configured, tested separately. If something broke in the middle, debugging was on you. Presence bundles everything under a single product with the OpenAI brand, and that changes the conversation a team has when deciding whether to build an agent.

The move follows a clear trajectory. Mid-July, OpenAI documented how teams use ChatGPT Work on data science and sales, with concrete operational guides. Days later it published a scorecard to measure AI ROI in business flows. Presence formalizes the agentic platform underneath that design. The message for those evaluating tools is consistent: first guides on how to use AI in processes, then a platform to deploy it in production.

The specific novelty is the combination of voice and chat in the same product. Voice agents in enterprise have a huge market: customer service, internal support, scheduling. Until now they required artisanal integration between ASR (speech recognition), LLM (the model generating the response), and TTS (synthesis reading it aloud), with latencies stacking at each step. If a user speaks and the agent takes three seconds to respond because each component waits for the previous one to finish, experience degrades quickly. If Presence unifies these steps in a single deploy, the integration savings are real for those managing call centers or help desks. It’s also a skills saving: a team that needed to know telephony, NLP, and cloud infrastructure now deals with a single vendor.

The main limitation is that the announcement, for now, is almost entirely marketing. No public pricing, no technical architecture, no benchmarks on latency or voice quality. “Trusted” appears twice in the excerpt and promises guardrails, monitoring, reliability. How much of this is engineering and how much is branding will become clear when the product is accessible and testable. For now, those deciding whether to invest have little to base a comparison on with what they already have.

For those building with AI, the reading is two-fold. On one hand, an official enterprise offering from OpenAI simplifies the conversation with corporate procurement: there’s a name, a contract, a vendor. On the other, those who’ve already built agents on Bedrock or with Anthropic tools face another player to evaluate, with a competitive advantage that’s promised more than proven. The operational question is whether migrating justifies transition costs, and that answer isn’t in the public materials.

Here’s what’s solid: an agent’s quality in production depends more on how you wire and verify it than on model power. Presence provides the infrastructure. Workflow design, limits, quality controls remain your responsibility. The best platform doesn’t solve the problem of deciding when the agent should transfer the call to a person, or how to verify it didn’t promise the wrong discount.