Field noteNov 15, 20265 min read1 source

Databricks and OpenAI publish three patterns for governed enterprise AI, including Codex via Unity Gateway

In one minute

Databricks describes three ways enterprises can use OpenAI tools with company data: business questions through ChatGPT and Genie, Codex routed through Unity Gateway, and custom agents deployed as Databricks Apps. For builders, the clearest operational detail is a governed route for Codex that keeps the OpenAI key in the gateway and logs requests for budget and spend oversight.

What changed: The companies frame their existing tools as three governed enterprise adoption paths: employee work, coding through Codex and Unity Gateway, and custom agents for business processes.

What happened

Databricks’ announcement presents three enterprise adoption patterns for its partnership with OpenAI. The partnership began in September 2025 as a multi-year agreement, the company says. OpenAI supplies GPT models, Codex, ChatGPT, and the OpenAI Agents SDK; Databricks supplies data, business context, and governance through products including Unity Catalog, Unity Gateway, Genie, Agent Bricks, and Databricks Apps.

The first pattern is business-user access to data through ChatGPT and the Databricks Genie plugin, which runs on Genie MCP. A user can ask a question in natural language, such as how weekly active-user adoption compares with a target by plan tier. ChatGPT uses Genie to interpret the company’s definitions, find relevant data, run a Databricks query, and return results as graphs or charts. Genie Ontology stores approved metrics, definitions, and relationships among company data. Databricks says users see only information they already have permission to access and that answers link to their sources in Databricks. Users may also ask ChatGPT Work for recommended next steps and carry out actions they approve through connected tools. The suggested starting point is installing the Genie plugin in ChatGPT.

The second pattern is coding with Codex under enterprise controls. Developers sign in using a Databricks credential, while the OpenAI key remains in Unity Gateway. Requests are logged, and engineering leaders can set budgets, review spending by team, and select which GPT model handles a task. Databricks says developers use Codex as they do today; its setup path is to install the Unity Gateway CLI and run ug codex.

The third pattern is custom agents for business processes. Databricks says OpenAI models can run next to company data under Unity Catalog governance. When Databricks serves the model, teams do not need to manage an external API key. They can build and evaluate agents in Agent Bricks or use the OpenAI Agents SDK and deploy agents as Databricks Apps. The company points to an Agent - OpenAI Agents SDK template in the Databricks workspace as a starting point.

The post also describes a broader architecture: Unity Gateway routes model requests, applies team budgets, and keeps an audit log; Unity Catalog governs Lakehouse data; and Lakebase, a managed Postgres database, stores agent memory and state. Databricks says its Model Serving can run GPT models near company data for use from SQL, notebooks, and agents.

What the archive adds

New Runtime’s earlier coverage described how cloud data platforms and AI infrastructure vendors are pairing governed data access with operational paths for applications and controlled deployment (Cloud data platforms and AI infrastructure vendors increasingly offer deployments that keep workloads within customer-controlled accounts, October 22, 2026). This is a new, more specific account of that operating model, with Codex routing and usage controls as a concrete integration pattern.

Trend signal

This supports the pattern that AI-native organizations are moving toward review and orchestration: the story describes budgets, team-level spend review, request logs, permissions, and user-approved actions as part of deployment rather than relying on model access alone. See AI-native organizations move toward review and orchestration.

Why now

Databricks says the partnership started in September 2025 and describes the current integration paths as practical starting points for teams already using ChatGPT and Codex, or building in Databricks. It also says separate posts with architecture details, setup steps, and customer lessons are planned over the following weeks.

Connections to other events

New Runtime’s earlier coverage described Factory and Databricks presenting enterprise modernization as one engineering workflow across data and dependent applications (Factory is partnering with Databricks to bring autonomous software engineering to enterprise data and AI, October 22, 2026). The current announcement addresses a related operational layer: how a coding agent can be routed through a governed gateway, with credentials, logging, and budgets managed centrally.

Why it matters

Enterprise adoption often stalls between giving engineers access to an agent and making that access governable at scale. The Codex-through-Gateway setup offers a concrete boundary for that problem: developers retain their normal coding workflow, while the organization can keep the provider key out of individual environments, record requests, assign budgets, and inspect spend by team. That can make rollout more legible to security, finance, and engineering leaders.

The other two patterns show that governance is not only a matter of restricting access. Business users can query data using company-defined metrics and existing permissions, while application teams can deploy agents near governed data. These are useful architectural options, but the announcement does not demonstrate that every organization should adopt all three. Teams should evaluate which boundary addresses their actual bottleneck: data context, coding-agent oversight, or production deployment.

Limits and uncertainty

Databricks is describing its own partnership and products, so the post is not an independent assessment of effectiveness. It reports no customer adoption figures, measured productivity gains, cost reductions, latency, or comparative security results. The post does not explain the details of Gateway’s logging retention, audit access, budget enforcement behavior, or how model selection is governed. Actual availability and configuration may also depend on workspace setup and product documentation; teams should validate their specific environment before relying on these controls.

Skill to build now

Map one coding-agent request path this week: identify where credentials are held, which gateway or proxy sees requests, what is logged, who can review usage, and whether budgets can be applied by team. Compare that map with Databricks’ Unity Gateway setup for Codex. Even if you use another platform, the exercise turns “enterprise-ready” into testable controls rather than a general promise.

Reader takeaway

The notable implementation detail is not simply that Codex works alongside Databricks. It is the proposed control path: Databricks credentials for sign-in, an OpenAI key retained in Unity Gateway, logged requests, and team-level budgets and spend review. Treat that as a deployment pattern to validate, not proof of outcomes.

Evidence / sources

  1. 1databricks.com/blog/openai-and-databricks-enterprise-ai-your-governed-data
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicCodexExplore the codex topic hub.
  2. 02topicEnterprise AiExplore the enterprise-ai topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract