AI
4 OpenAI operations a decision flow can call directly, with the credentials your own contract issued. The response is data the rest of the flow reads, branches on, and keeps in the trace.
Who they are
OpenAI is an AI research company that sells access to its GPT models through an API, alongside ChatGPT. For decisioning the relevant surfaces are text generation, structured extraction, and embeddings — turning text into vectors, which is what makes de-duplicating free-text fields practical.
What a flow can call
/v1/responsesThe current text endpoint — Responses, not Chat Completions.
/v1/responsesSame endpoint constrained to a JSON schema, so the reply parses instead of needing a regex.
/v1/embeddingsTurns text into a vector for similarity work — matching a free-text employer or address against known values, or spotting near-duplicate applications.
/v1/modelsModel ids the key can actually reach. Cheapest way to confirm a key works, and better than hardcoding a list that ages out.
Where it sits in the decision
A model call is for the judgment a table cannot express — reading a document, summarising a file, weighing a narrative. Because the output is not deterministic, a flow constrains it with a schema and keeps the prompt and the response in the trace, so the reasoning behind a referral stays as inspectable as a rule.
Whatever OpenAI returns is part of the run, so it is part of the record. When someone asks months later why an applicant was declined, the answer cites what came back at the time rather than re-fetching from a service whose answer has since changed.
Common questions
Add OpenAI as a connection in your workspace authenticating with an API key, with the credentials your own contract issued — ArboRule calls the provider as you, and never holds a contract on your behalf. Once the connection exists, any flow in the workspace can place a Connection node and choose one of its operations. The credentials live on the connection, not in the flow, so a policy owner can use OpenAI in a decision without ever seeing the secret.
OpenAI exposes one host for both environments, so there is no separate sandbox to point at. Test runs still execute in Sandbox and are recorded separately in decision history, but the call goes to the same place as production — so guard it with your own test credentials, rate limits, or data.
4 operations, including “Ask a model (Responses)”, “Extract structured JSON”, “Embed text (vector)”. Each one is a step you place on the canvas and map into the fields your rules read, and most flows start with “Ask a model (Responses)”. The list comes from the same manifest the engine uses to make the call, so this page cannot describe an operation the product does not have.
A model call is for the judgment a table cannot express — reading a document, summarising a file, weighing a narrative. Because the output is not deterministic, a flow constrains it with a schema and keeps the prompt and the response in the trace, so the reasoning behind a referral stays as inspectable as a rule.
Also in AI
The AI lab behind the Claude models.
anthropic.comSaved Claude Code automations that run on demand.
claude.ai
Google Gemini4 operationsGoogle's family of multimodal large language models.
ai.google.dev
Vertex AI3 operationsGoogle Cloud's machine learning platform, including the Gemini models.
cloud.google.comReady when you are
Build the flow in Sandbox, connect your account, and watch the decision pull what it needs before it answers.