How OpenAI turns enterprise data into trusted context for agentic AI with Fivetran

A look under the hood at how OpenAI built an in-house data agent on trusted data, using Fivetran and dbt. The agent details come from OpenAI’s write-up, “Inside OpenAI’s in-house data agent.” The infrastructure details come from our work with OpenAI.

Operating at that scale requires a flexible and future-proof open data infrastructure capable of keeping pace with the speed of AI innovation and changing market landscape. OpenAI’s teams need timely access to governed and trustworthy data at their fingertips. However, as the company grows, so does the scale and complexity of its data environment. OpenAI’s data ecosystem spans a diverse and constantly changing set of data sources and tools, requiring a mix of in-house pipelines, third-party connectors, and manual processes to bring that data together. Previously, engineers spent hours each week keeping information flowing across API and schema changes, work that required specialized technical expertise and diverted some of OpenAI’s most valuable resources toward maintaining existing infrastructure. In a hyper-competitive AI market where the pace of innovation can define market leadership, every hour spent keeping data infrastructure running was an hour not spent advancing the data and AI capabilities that could shape what OpenAI builds next. Even with that effort, some data could only be refreshed weekly — a model that couldn't keep pace with the speed of the business.

OpenAI’s data platform needs to serve more than 3,500 internal users with more than 600 petabytes of data spanning over 70,000 different data sets (source: OpenAI). At that scale, answering a business or technical question was becoming increasingly time-consuming, requiring analysts to identify the correct tables and metrics, understand how they differed, assess data freshness, and apply the appropriate business logic and context.

OpenAI needed to make high-quality, conversational analytics accessible across engineering, data science, go-to-market, finance, and research teams without forcing every employee to become a context or data engineer. They needed an internal data agent that could answer business questions through natural language.

Open Data Infrastructure: Standardized enterprise data ready for AI agents

For OpenAI’s data agent to deliver trustworthy answers, it first needed a reliable data foundation — one capable of unifying data at scale while preserving the context required to interpret it correctly.

Open Data Infrastructure (ODI) is an architectural approach that enables organizations to store data once in open table formats (i.e. Apache Iceberg, Delta Lake) and use it across multiple compute engines, analytics tools, and LLMs. Unlike architectures centered on a single vendor, ODI avoids lock-in without requiring companies to maintain duplicate copies of their data. This reduces cost, complexity, and the governance risks created when those copies fall out of sync.

Fivetran + dbt bring this architecture to life. Fivetran’s 750+ pre-built connectors and Connector SDK, which lets teams build custom connectors for specialized sources, reliably move data from OpenAI’s source systems into Azure Data Lake Storage. Fivetran Managed Data Lake Service automatically converts that data into open table formats and integrates it with catalogs, making it available to any compatible compute engine or LLM.

dbt turns this standardized source data into trusted, usable data products. Documented models, unit tests, column-level lineage, and consistent metric definitions capture the business logic behind the data — providing both people and AI systems with the meaning needed to use it correctly.

Together, Fivetran + dbt give OpenAI the flexibility to adopt the best tools for each workload as the market evolves. Data remains open, current, governed, and consistent without requiring engineers to repeatedly move it, duplicate it, or manage the underlying infrastructure.

The solution: Giving an in-house data agent trusted context

With an ODI managing and maintaining their data, OpenAI built a bespoke data agent grounded in its internal data sets, business knowledge, user permissions, and automated workflows. Employees can ask complex questions in natural language, and the agent identifies relevant data, executes queries, validates its work, and synthesizes findings in minutes.

Fivetran + dbt supply the trusted data and analytical context at the heart of this system. Fivetran continuously delivers current operational data, while dbt provides modeled data sets, lineage, and documented logic. OpenAI then enriches this foundation with institutional knowledge, memory, and agent-specific validation.

Available wherever employees work — including Slack, the web, IDEs, the Codex CLI and OpenAI’s internal ChatGPT — the agent provides consistent access to trusted insights while enforcing OpenAI’s existing data permissions (source: OpenAI).

The result: Turning questions into decisions in real-time

Giving OpenAI’s employees real-time, self-service access to information reduces ad hoc reporting work for engineers and analysts, redirecting that capacity toward higher-value innovation. At OpenAI’s scale, those productivity gains help the company operate more effectively and improve the AI products used by 1 billion people and 2.5 million business customers.

The agent handles much of the end-to-end analytical workflow, including discovering data, writing and executing queries, investigating unexpected results, and producing notebooks or reports. It also learns from corrections and recurring patterns, giving future analyses a stronger starting point (source: OpenAI).

An ODI design that delivers fresh, governed data and context is changing the economics of analytics: Instead of every new question becoming another ticket for the data team to action over days and weeks, employees gain always-available, trusted insights. Data specialists can spend more time improving the quality of shared data and supporting high-value decisions, while the broader business can learn and act faster.

Fivetran + dbt provide the enterprise standard for OpenAI’s agentic AI and analytics

OpenAI’s experience demonstrates that an AI agent is only as trustworthy as the data and context behind it. Agents you can trust require complete and current data, an open architecture that can scale, and governed models that communicate what the data means. Furthermore, a strong open data foundation not only serves agentic AI but all other analytics use cases. Choosing Fivetran + dbt has given OpenAI an extensible underlying data infrastructure that not only feeds data agents, but also powered one of OpenAI’s largest marketing initiatives to date: a high-profile ad placement during the Super Bowl, the biggest night in U.S. sports and advertising, maximizing their multi-million dollar investment. 

With Fivetran + dbt as its open data foundation, OpenAI can transform data at extraordinary scale into trusted intelligence — accelerating innovation today and creating the foundation for breakthroughs yet to be imagined.

‍

[CTA_MODULE]