Overview

Workflows let you chain together AI tasks into an automated sequence. Instead of running each step manually — extracting key terms, researching the relevant law, drafting a summary — you build a flow once and run it whenever you need it. Each step feeds its output into the next, so the whole process runs end to end without you switching between tools. They’re built on a visual canvas where you lay out steps, connect them, and configure what each one does. The result is a reusable pipeline that can handle complex, multi-step work in a fraction of the time it would take to do manually.

​

How workflows are structured

A workflow is a linear sequence of steps. Each step is a node that does one thing — generate text, search legal sources, translate a document, run a grid review. You connect nodes in the order you want them to run, and data flows from one step to the next automatically. There’s no branching or looping — every workflow is a straight line from start to finish. This keeps things predictable: you always know what runs when, and each step can reference the output of any earlier step.

​

What steps can do

Steps fall into two categories: inputs that collect information from you, and actions that do something with it.

​

Inputs

​

Actions

A Code interpreter step can return files, such as Excel and CSV files, charts, and HTML pages. The files appear in Final assets after the run. For previews and downloads, see “Files, charts and HTML pages” in the Chat overview.

​

How data flows between steps

Each step can reference the output of any step that came before it. When you’re configuring a step’s instructions, type @ to mention a previous step by name. At runtime, that mention is replaced with the actual output from that step. For example, a “Run prompt” step can @ a “Files upload” step to analyze the uploaded documents, and a “Drafting” step further down can @ both to produce a final report. The data threads through the workflow automatically.

​

Workflows in chat

You can run workflows directly from a chat without opening the builder. Ask the AI to handle a multi-step task and it looks through your library for a matching workflow and runs it inline — pausing for input the same way it would in the builder, and showing each step in the chat thread as it goes. If nothing in your library fits, the AI can sketch a new workflow on the fly. It proposes a draft — a sequence of steps designed for what you asked — and runs it. After the run, the draft lands in your library under Personal, where you can keep it, edit it in the builder, or discard it. This means you can ask for something like “translate these contracts to English, summarize each, and flag any unusual clauses” and let Saga build and run the right flow without you laying out the canvas yourself.