# OpenAI on OpenAI: Applying AI to Our Own Workflows

Source: https://www.youtube.com/watch?v=nKuXMDCtyQI
Recap page: https://rapidrecap.app/video/nKuXMDCtyQI
Generated: 2025-10-08T18:03:27.475+00:00

---
## Quick Overview

OpenAI is applying AI internally to enhance workflows across sales, HR, and support, leading to significant improvements like 70% ticket deflection and saving reps one full day per week, achieved by building internal tools like the GTM Assistant and OpenHouse, which leverage AI to amplify expertise and scale operations efficiently.

**Key Points:**
- OpenAI's internal application of AI resulted in 70% ticket deflection and saved sales reps approximately one full day per week for higher-leverage work.
- The GTM Assistant, built using an architecture of Surfaces (ChatGPT/Slack/OpenAI Platform), Agent Orchestration (SDKs, GPT-5), and Connectors/Knowledge (Databricks, Vector Store), supports sales workflows.
- OpenHouse, an internal HR/People tool, uses a similar layered architecture, connecting to HR systems and OpenAI Vector Store to onboard new employees and answer policy questions.
- The Support Agent framework, using Help Center/Articles and Support/SOPs data, demonstrated AI handling support tasks, achieving 80% positive evaluation.
- The process for creating these internal AI tools involved reviewing conversation logs, defining gold standards for actions, codifying standards into knowledge, and connecting knowledge for evals/classifiers.
- OpenAI is experiencing exponential growth in support tickets following product launches like ImageGen, highlighting the need for scalable support solutions.

![Screenshot at 07:05: The GTM Assistant architecture diagram showing three layers: Surfaces, Agent Orchestration \(using GPT-5\), and Connectors & Knowledge, illustrating the framework used to automate internal workflows.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-07-05.png)

**Context:** The presentation, delivered by Scotty Huh from GTM Innovation at OpenAI DevDay [2025], focuses on how OpenAI applies its own AI technology internally to improve efficiency and scale operations across different departments, specifically highlighting solutions for Go-to-Market (GTM) and HR/People functions, and later, scaling customer support.

## Detailed Analysis

The presentation details OpenAI's internal application of AI to solve scaling challenges across GTM, HR/People, and Support functions. For GTM, the GTM Assistant framework, which utilizes ChatGPT, Slack, and the OpenAI Platform across three layers (Surfaces, Agent Orchestration via SDK/GPT-5, and Connectors to Databricks/Vector Store), successfully amplified sales expertise. This resulted in reps exchanging over 20 messages weekly with the assistant and saving about one full day per week for higher-leverage work. The presentation also introduced OpenHouse, an internal tool for HR/People knowledge, built on a similar structure, which helps new hires quickly understand company operations and policies by connecting to HR systems and OpenAI Vector Stores. For customer support, the Support Agent model, which uses ChatKit and Realtime API surfaces, demonstrated strong results: 70% ticket deflection, a 30% increase in solution rate, and 80% positive evaluation, outperforming legacy systems by 30%. The speaker emphasized that the key to scaling support is building this AI into familiar tools and creating a self-improving loop driven by evals and guardrails. He concluded by challenging the audience to use these principles to build internal AI tools that their teams cannot live without.

### GTM Assistant Architecture

- Surfaces (ChatGPT (MCP), Slack, OpenAI Platform)
- Agent Orchestration (Meetings, Product Knowledge, Custom Demos, Customer Research via Agents SDK, GPT-5, Responses API)
- Connectors & Knowledge (Customer Data/Databricks, OpenAI Vector Store)

### OpenHouse Architecture

- Surfaces (ChatGPT (MCP), Slack, ChatKit)
- Agent Orchestration (Company Knowledge, People Connector, Career Growth)
- Connectors & Knowledge (People/HR Systems, OpenHouse CMS, Google Drive, Notion, OpenAI Vector Store)

### Support Agent Architecture

- Surfaces (Help center (ChatKit), Realtime API)
- Agent Orchestration (Ticket Classification, Actions via Agents SDK, Responses API)
- Connectors & Knowledge (Customer Tickets, Help Center + Articles, Support + SOPs, OpenAI Vector Store)

### Internal AI Impact Metrics

- 70% Ticket Deflection
- +30% Solution Rate
- 80% Positive Evaluation

### Internal AI Process Definition

- 01: Reviewed conversation logs with specialists
- 02: Defined gold standards for actions
- 03: Codified standards into knowledge
- 04: Connected knowledge to evals and classifiers

### OpenAI Internal Usage

- 20+ messages exchanged per rep weekly with GTM assistant
- ~1 full day saved for higher leverage work per rep weekly

![Screenshot at 00:07: Opening slide for Scotty Huh's presentation titled "OpenAI on OpenAI: Applying AI to Our Own Workflows".](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-00-07.png)
![Screenshot at 00:56: Slide showing the two key goals for GTM Assistant: 20+ messages exchanged weekly and ~1 full day saved for higher leverage work.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-00-56.png)
![Screenshot at 01:32: Slide posing the key question for internal AI application: "How do we use AI to drive efficiency?"](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-01-32.png)
![Screenshot at 02:44: Slide outlining the "Golden age of internal building" which promises 10x company agility and impact through internal AI.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-02-44.png)
![Screenshot at 05:57: The GTM Assistant architecture diagram showing Surfaces, Agent Orchestration \(including GPT-5\), and Connectors & Knowledge.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-05-57.png)
![Screenshot at 07:36: Slide illustrating the four-step process for defining the support agent workflow, starting with reviewing conversation logs.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-07-36.png)
![Screenshot at 12:13: Slide showing the "gtm-assistant-scotty" thread in Slack, where feedback is provided on missed takeaways.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-12-13.png)
![Screenshot at 12:24: Demonstration of providing feedback within the tool, selecting "Approve" for a suggested change related to prompt tuning.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-12-24.png)
![Screenshot at 16:17: Slide showcasing the OpenHouse architecture, focused on People/HR systems.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-16-17.png)
![Screenshot at 20:42: Slide illustrating the scaling challenge in support tickets following the Image Gen launch, showing spikes in volume that the team successfully managed.](https://ss.rapidrecap.app/screens/nKuXMDCtyQI/00-20-42.png)