OpenAI on OpenAI: Applying AI to Our Own Workflows

OpenAI’s internal push for “agent-driven workflows” is being framed less as a quest for efficiency and more as a way to amplify company expertise—turning the best practices of top performers into systems that scale across entire teams. The core claim is straightforward: automation matters, but the bigger payoff comes from capturing the craft of practitioners and distributing it so every employee can operate like the organization’s best operator.

Three internal deployments illustrate how that idea works in practice: a Go-to-Market assistant for sales, Openhouse for HR and onboarding knowledge, and a support system that improves itself as ticket patterns change. Each system follows a similar architecture—connectors to pull in relevant data, an orchestration layer that routes tasks to specialized “skills,” and services that push actions back into tools employees already use. The difference is the center of gravity: sales skills are grounded in meeting prep, product knowledge, demos, and customer research; HR skills are grounded in people and policy context; support skills are grounded in standardized operating procedures (SOPs) derived from expert handling of real conversations.

In sales, the Go-to-Market assistant was built after teams hit a breaking point and needed faster customer research and technical answers without sacrificing customer experience. Developers sat with a top rep, Sophie, and extracted her “formula” for winning—prepping for meetings, generating product champions, planning demos, and following up quickly. That craft was encoded into an agent system with a semantic data layer over customer and strategy documents, plus specialized skills (starting with four and growing to about 10) such as meetings, product knowledge, custom demos, and customer research. The assistant then operates inside existing workflows—primarily ChatGPT and Slack—producing meeting briefs, demo scripts, and follow-up answers. It also closes the loop: when a rep flags a missed takeaway, the system regenerates the output and triggers prompt optimization that flows to developers for approval, effectively distributing improvements across the org. Reps exchange about 20 messages per week with the assistant and report saving roughly one full day weekly.

For HR, Openhouse targets the knowledge gap created by rapid global hiring. It connects HR systems like Workday with internal announcements and role expectations, then surfaces answers through Chatbot and Slack-based Q&A. A real example: during a New York trip, an employee uses Openhouse to get travel policy and office access details with citations, then extends the same question into a directory-driven search for a coworker who can help with a customer-facing use case. The result is a path from policy knowledge to human expertise without silo-hopping.

Support tackles a different kind of scale problem: hypergrowth and rapid product change can spike ticket volume by orders of magnitude, as seen around the launch of image gen. The support system defines expert “gold standards” for when to answer, when to escalate, and when to tag for audit, then codifies them into SOP knowledge. That SOP knowledge is tied to knowledge and eval so interactions update the system over time—especially when new ticket patterns cause automation failures. Reported outcomes are concrete: about 70% of tickets are deflected or handled autonomously, the system outperforms legacy by about 30%, and roughly 80% of manually reviewed tickets earn highly positive QA ratings.

The takeaway is a blueprint for internal builders: find the top operator to model, embed agents into familiar tools rather than creating new software, and use scaled platforms (OpenAI’s agent kit, including agent builder, Chatkit, and eval) to accelerate deployment. The pitch ends with a sprint-style challenge: build something teams can’t live without, starting this week.