OpenAI has today launched OpenAI Presence, an enterprise product designed to help organizations deploy reliable AI agents.
As agents take on more consequential customer and employee workflows, CX leaders will need to move beyond relying on controlled demonstrations and human review alone to judge whether systems are ready for production.
This launch comes as OpenAI and Hugging Face investigate a security incident in which advanced models exploited vulnerabilities to access information on Hugging Face’s production infrastructure.
The Cost of Getting Deployment Wrong
Today, enterprise AI agents are expected to carry out consequential tasks within established boundaries, consistently completing end-to-end workflows across customer service and internal operations.
When AI agents deployed to perform complex tasks or access sensitive information, they can introduce operational and regulatory risks that are not apparent in controlled AI prototypes.
This causes caution for organizations expanding AI into high-value workflows, as the supporting infrastructure required for autonomous agent safety has not always kept pace.
Businesses must therefore establish clear permissions and governance frameworks, as without these safeguards, the risk of inconsistent customer experiences and unintended actions increases, likely damaging trust.
In CX environments particularly, organizations face increasing pressure to improve service availability while controlling operating costs.
In conversation with CX Today, Sushil Kumar, CEO of Cyara, argued that reducing AI hallucinations and maintaining reliable CX requires governance that operates continuously alongside AI systems.
“Human oversight still matters, but it should function as a checkpoint rather than the entire CX assurance system,” he argues.
“Effective AI governance has to move at machine-speed, with automated validation, guardrails, and real-time testing.”
Voice interactions also add another layer of complexity because agents must understand requests in natural language to make decisions and respond in real time, particularly in highly regulated industries where interactions may involve sensitive data, financial consequences, or timely decisions.
Building for Controlled Scale
The launch of OpenAI Presence enables organizations to determine exactly what knowledge an AI agent can access, the enterprise systems it may interact with, and what actions it is authorized to perform, ensuring when conversations or tasks should be escalated to a human.
This approach aims to ensure that AI agents operate within defined boundaries, working alongside OpenAI’s enterprise customers through its Forward Deployed Engineers and selected systems integration partners to connect the agent with relevant business systems and governance policies preparing the deployment for production.
Presence includes tools that simulate real-world scenarios before deployment and continuously monitor production performance to identify weaknesses and opportunities for improvement.
Furthermore, the platform involves Codex-powered improvement process, designed to analyze production data, investigate issues, and recommend updates to refine agent performance.
The performance of Presence has reportedly shown strong results for OpenAI’s English-language phone support service, where the AI agent resolves approximately 75% of inbound issues without human assistance, with its Codex-powered improvement process reducing handoffs by 15% over a 10-day period.
The Wider Safety Debate – Hugging Face
OpenAI Presence highlights the recent challenges surrounding the safety of deploying capable AI agents.
Last Thursday, the AI company, Hugging Face, disclosed an incident during an internal cyber capability evaluation in which advanced models exploited a chain of vulnerabilities to gain internet access and access information on Hugging Face’s production infrastructure.
The advanced models, operating with reduced cyber safety restrictions, were detected and contained before broader disruption, with both OpenAI and Hugging Face continuing to investigate the incident while strengthening security controls and evaluation practices.
This incident highlights that overly advanced AI agents can discover unexpected paths through connected systems when pursuing multi-step objectives, requiring organizations to input stronger governance and oversight to ensure capabilities remain within defined boundaries.
For CX, as enterprise AI agents are increasingly expected to authenticate customers, access sensitive information, and complete approved actions across business systems, these capabilities increase the importance of restricting system access.
Presence aims to solve this dilemma by helping organizations deploy AI agents safely while maintaining oversight of their behavior and decision-making.
Clem Delangue, Co-founder and CEO at Hugging Face, explained that advancing AI capabilities must be accompanied by collaborative security research, transparent evaluation practices, and robust governance.
“We’re grateful for the collaboration with OpenAI on this and other topics,” he said.
“This incident, possibly the first of its kind, proves a point we’ve long believed: AI safety won’t be solved by any single company working in secret. It will be solved in the open, collaboratively, with broad access to AI for every defender, everywhere.”
