OpenAI puts support agents into enterprise service
OpenAI launched Presence, an AI agent platform built to run customer and employee support for organizations. It works across realtime voice and chat channels, and it is not sold self-service: deployments run through a limited access program, led by OpenAI engineers or selected integrators.
Support agents become a packaged product
OpenAI introduced Presence, its enterprise support platform, with the details reported by industry press on July 23, 2026. The platform runs agents that answer questions, resolve issues, connect to company systems, take approved actions and escalate to a person when human judgement is needed, across both voice and chat. OpenAI says the platform also powers its own English-language phone support channel, where 75% of inbound issues are resolved without human assistance. Deployments are built around a single task at a time, covering scenarios that range from billing questions to insurance claims.
Limited access and a human-approved repair loop
Presence cannot be bought on a self-service basis. Access is granted to eligible organizations through a limited general availability program, with deployments led by OpenAI's forward deployed engineers or selected global systems integrators. Guardrails are defined before launch and edge cases are rehearsed through simulation; once live, a feedback loop that depends on human approval keeps running.
That loop is the distinguishing part: OpenAI's coding model looks at the points where the support agent fails, writes a fix, and a person signs it off. According to figures the company shared, the method cut handoffs to humans by 15 percentage points within ten days. Early customers include BBVA, which is trying the platform for banking support in Mexico, SoftBank, testing Japanese-language conversations, and Australian insurer IAG, assessing support load during high-demand periods.
From chatbot to a layer that takes action
The support scenario discussed for years was a chatbot that produced text answers. Presence describes something beyond that: a layer that connects to systems, verifies identity and carries out approved actions. Two cautions stand out in the commentary. The first is reliability at scale: the real question is not how many tasks AI can take on, but whether it handles them consistently at volume. The second is cost: integration and governance expenses often exceed model usage costs and can erode early savings. On employment, the expectation is slower hiring rather than sudden layoffs, with tier-1 support seen as the most exposed area.
What it means for businesses in Türkiye
Presence is a limited access product positioned for enterprise scale, so it is not something small and mid-sized businesses in Türkiye will buy directly in the short term. The model it establishes, however, travels well. An assistant layer that handles frequent questions, runs routine tasks such as appointment and order status, and hands over to a person the moment it cannot resolve something can be built today. Three things decide the outcome in Türkiye: the quality of Turkish conversation, data processing that complies with KVKK, and the permission boundaries that define which actions an agent may take on its own. Keeping hard-to-reverse steps such as payments, cancellations and price commitments under human approval prevents the speed of automation from turning into risk. Measurement matters just as much after launch: without tracking which questions get resolved, where handovers happen and how long customers wait, the system goes blind.
The UNALSOFT take
Our reading is this: in support automation the difference comes from the system built around the model, not the model itself. Which questions go to the agent, which action stops at human approval, how the handover works when something breaks and how conversations are measured are what decide the result. That is what we build in our Agentic AI service: agents connected to a business's real workflows, with defined boundaries and a human kept in the loop on critical decisions. Models change; this skeleton does not.
Boundaries first, automation second.
Let's clarify together which questions belong to an agent and which stay with a person.