AI & Agents

Combine AI agents, structured prompts, deterministic rules, and human approval points in one automation. Managed Gemini models run without your own API key and bill through Glow. Configure the Step’s inputs, tools, and output before testing it.

At a glanceVerified in app · Oct 30, 2026

Use it to: Build automations with Workflow Assistant and run managed or BYOK language models inside AI Prompt, AI Data Transform, AI Agent, and AI Conditions
Find it: Bottom dock → AI (AI Prompt, AI Agent), Tools → Data → Change (AI Data Transform), or Assistant panel
Cost: Managed Gemini models use workspace Glow AI tokens (no API key needed) · External AI app connectors use Step credits plus provider billing
Limits: AI Steps see only explicitly mapped Workflow data references (plus enabled tools, memory, and reference files on AI Agent)
Output: Text, JSON, or schema-validated Separate fields (AI Prompt and AI Agent) · Up to 3 positional values (AI Data Transform)
Instead: Summarize (for deterministic counts, totals, and averages) or typed Conditions / Switch (for fixed routing rules)

Which one to reach for

Draft or edit an automation on the canvas from a plain-English request.

Workflow Assistant

Classify, summarize, or generate one structured response from supplied data.

AI Prompt

Extract up to three ordered values from unstructured input in parallel.

AI Data Transform

Reason across multiple turns, call permitted tools, and use session memory.

AI Agent

Choosing an AI Step compares inputs, outputs and operating controls. For counts, totals and averages, use Summarize rather than a model. To build or debug from ChatGPT, Claude, Gemini, Cursor, or VS Code while watching the canvas update live, click Connect AI tool in the Workflow Assistant header or follow Connecting AI Apps & Editors.

AI inside a routing step

Conditions and Filter offer AI Condition for rules written in words. Use it when a typed comparison cannot express the test. Each evaluation costs a model call, once per item for a Filter rule, and results can vary between runs. Switch is typed-only.

Two ways to run AI in an automation

AI Prompt, AI Data Transform and AI Agent use managed models. AI Agent also supports connected OpenRouter. Provider connectors use your own accounts. You can combine these in one automation; Workflow Assistant builds the automation rather than running as a Step.

Built-in Glow AIGlow's AI Steps

Managed Gemini models run without your own provider API key.
AI Agent also supports a workspace OpenRouter connection.
Metered through Glow AI tokens (or OpenRouter for OpenRouter models).
Configure inputs, output format, and any permitted tools directly on the Step.

External App ActionsAI App Connectors

Connect your own OpenAI, Anthropic, or other provider account.
Provider bills model usage directly to your connected account.
Glow meters each action execution using Step credits.
Connect the provider account first, then configure the connector action.

The providers you can connect

Find provider connectors in the AI group in the Apps menu. Check the selected connector’s account and authentication requirements in Credentials. Provider charges follow your agreement with that provider; Glow execution usage still applies.

Choosing a model

Use AI Agent’s Brain (model) dropdown to choose a model. AI Prompt and AI Data Transform use managed defaults. Compare answer quality, tool choices, latency and recorded usage on representative inputs. See Models for model options.

What a run actually uses

Plan for both execution usage and AI usage. Call counts help estimate volume, but model choice, input and output size, repair and tool use also matter.

AI Prompt makes one answer call, with an optional additional call for shape repair.
AI Data Transform makes one per instruction, so three instructions is three calls, made at the same time.
An AI Condition rule makes one call each time it is evaluated. Filter caps list length multiplied by configured AI rules at 100 item-rule pairs.
The AI Agent can make multiple calls within its configured turn, model-call, tool-call, active-time and estimated-credit limits. Counts can vary between runs on the same input.

Measure representative runs before going Live and check System Limits.

Try it before you trust it

Choose a real test

For AI Agent and AI Prompt, use Run test (or Test & Debug) with explicit input. For other eligible actions, use an authorized Workflow Assistant action test. Run with test data replays saved Testing data, not a fresh model response. Real tests can incur usage and external side effects; see Testing & Debugging before running them.

Include difficult inputs

Test representative inputs, missing information, ambiguous requests and cases that should take a review or error route.

Check downstream fields

Check the values and shape later Steps need before connecting them.

Successful execution or a valid shape does not establish factual accuracy.

AI Prompt and AI Agent offer Text, JSON, and Separate fields under Output. Use Separate fields to define required fields through Fields or Advanced (JSON Schema). JSON parses the answer without validating which fields come back.

The wider AI catalogue

For document extraction, search, speech or other provider services, find your service in the Integrations Directory . Check the selected action’s fields, account requirements and charges before building.

Retrieval over your own documents

If you build retrieval with an embedding provider and a vector database, plan both parts:

Part	Work to configure
Index documents	Read the selected documents, generate embeddings and store them with retrievable passages.
Answer a question	Embed the question, query the database and supply retrieved passages to AI Prompt or AI Agent.

Review source access controls and provider charges. For public web sources, test the URL with the agent’s enabled reading tool and selected model; public reachability does not guarantee readable content.

What AI steps cannot see

Map earlier Step outputs into the AI Step’s instructions using references selected from Workflow data. Connecting Steps does not by itself supply all earlier outputs as model context.

AI Agent can also receive context through enabled tools, configured memory and reference sources. Check the supplied context and tool results when reviewing an answer.

Verify it works

Run: Add an AI Prompt or AI Agent Step to a draft, set its instructions, and click Run test
You should see: The Step returns a structured response and logs token usage without needing a provider API key.

What’s Next?

👉 AI Prompt →: classify, summarize and transform supplied text.
AI Agent: deploy autonomous, tool-using agents with session memory and reference sources.
Models: compare managed models and connected OpenRouter options.