OpenAI Support Model: Transforming User Support into an AI Operating Model
OpenAI has transitioned its user support from a traditional ticket-based system to an AI operating model designed for hypergrowth. This approach shifts the focus from processing transactional volume to creating a continuous learning loop where every user interaction improves the system's overall performance.
An AI Operating Model for Support
OpenAI views support not as a volume challenge, but as an engineering and operational design challenge. The organization has moved beyond using chatbots for simple deflection to a system where interactions across all channels—chat, email, phone, and embedded product help—continuously inform and refine the support experience.
This model is built on three core building blocks:
- Surfaces: The interaction points, including chat, email, phone, and help features embedded directly within the product.
- Knowledge: A dynamic system of guidance derived from real conversations, policies, and context, rather than static documentation.
- Evals and Classifiers: Shared definitions of quality created by humans and software to measure, improve, and highlight feedback.
Because these primitives power every surface, improvements made to one channel (such as a pattern identified in an enterprise conversation) automatically scale across all other support channels.
The Evolving Role of Support Representatives
In this new model, support representatives are transitioned from responders to "systems thinkers" and builders. Rather than focusing solely on resolving individual tickets, they are empowered to contribute to the architecture of the support system:
- Contributing to Evals: Reps flag interactions that serve as test cases to codify what "great" support looks like.
- Shipping Classifiers: Reps propose and implement classifiers when they identify new user patterns.
- Prototyping Automation: Reps build lightweight automations to close workflow gaps.
As Engineering Manager Shimul Sachdeva notes, "Agents aren’t just responding to tickets. They’re informing our knowledge base and our policies."
Technical Implementation via OpenAI Stack
OpenAI's internal support system is built using its own suite of developer tools, allowing for rapid iteration and integration of new capabilities:
- Agents SDK: Provides step-level traces and observability, enabling the team to replay runs, inspect tool calls, and debug root causes.
- Responses API: Powers classifiers used to monitor tone, correctness, and policy adherence.
- Realtime API: Enables the implementation of voice support.
- Evals Dashboard: Allows quality to be measured and visualized over time.
Starting with a simple Q&A system, the team utilized the Agents SDK to expand into dynamic actions such as processing refunds, managing invoices, and performing incident lookups.
Compounding Learning and Feedback Loops
The system is designed so that learning compounds over time. Evals turn everyday conversations into production tests, ensuring that responses are polite, clear, and consistent. This process allows the system to identify when a model should not answer a question, reducing errors.
This feedback loop extends beyond the AI to the organization itself. Specialists identify where models fall short and contribute datasets for fine-tuning, while observability dashboards provide a measurable way to track performance improvements. This integration ensures that patterns from user interactions feed directly back into knowledge bases, automation, and product design.