What Is Glow?
Build and govern automations that combine AI agents, deterministic rules, and human approval on one shared canvas. Your team and the Workflow Assistant edit the graph together in real time. Each automation pairs agentic Steps (AI that reasons with permitted tools) and deterministic Steps (fixed rules).
- Use it to
- Combine governed AI agents, deterministic routing rules, and Human Review on one shared multiplayer canvas
- Find it
- Automations → Create new → Workflow (or Connect AI tool over MCP)
- Cost
- Free for triggers, Conditions, Switch, Filter, Wait, and Repeater · 1 Step credit per paid action · Glow AI tokens for AI Steps
- Instead
- Choose the Right Step (/build/which-step) when deciding between an AI Step and a deterministic rule
Give an agent selected tools to investigate a request, use business rules to route its findings, and require approval before consequential actions.
The Hybrid Execution Architecture
An automation can combine agent reasoning, deterministic execution and human decisions in one visual graph. You define which part of the process each one controls.
| Part of the architecture | Its role in the process |
|---|---|
| Governed AI agents | An AI Agent reasons over a task and calls tools. Configure its connected accounts, permitted actions and approval settings. |
| Deterministic business logic | Conditions and Switch apply explicit routing rules. Code and data Steps handle exact calculations and transformations. |
| Human governance | Place Human Review at a decision point. Connect the approved outcome to the action that needs sign-off, and handle other outcomes separately. |
| Real-time multiplayer collaboration | Your team, the Workflow Assistant, and connected AI apps (Claude, ChatGPT, Gemini, or your IDE) build on the shared canvas. Review the graph and test results before enabling automatic runs. |
Governance is configured in the automation: tool permissions determine access, approval settings gate selected tool calls, and Human Review gates the downstream route you connect. Instructions tell an agent what to aim for; permissions and review controls define what it may do.
Build together, with AI beside you
An example of working together
One canvas for your team
Open an automation with a colleague and edit it together. Changes appear in real time, so you can build a branch or explain a process without passing files around. See how collaboration works.
The Workflow Assistant and your AI tools build where you build
Describe what you need in the Workflow Assistant, or click Connect AI tool to build from Claude, ChatGPT, Gemini, Cursor, or VS Code. Steps appear on your live canvas as you chat. Review the draft and test it before switching to Live.
Choose how to build
Use the canvas directly, ask the Workflow Assistant for help, or work from an external AI client:
- On the canvas: add and connect Steps yourself, configure their fields and review changes with a colleague. Choose the right Step.
- With the Workflow Assistant: describe the automation or the change you need, then inspect the Steps it creates or edits. Start with a worked prompt.
- From ChatGPT, Claude, Gemini, or your IDE: connect an AI app or coding editor through MCP to inspect, build, and test automations while watching your canvas update in real time. Connect an AI tool.
The Workflow Assistant builds the automation; an AI Agent does work when it runs.
Glow operates its own native integration framework, where app Steps and agent tools share the same action implementation and scoped permissions. See How Glow Integrations Work.
Follow one request through Glow
Imagine a customer sends: “I was charged twice. Can you help?” Here is how a five-Step automation investigates the request, applies routing rules, and asks a person to approve the reply before sending it:
| Part of the process | What you configure |
|---|---|
| Receive the request | A webhook starts the run with the customer’s message. |
| Investigate and draft | Give the AI Agent selected lookup actions. Ask it to return findings, a category and a draft reply in defined output fields. |
| Apply a business rule | Switch routes the category to the relevant review. Send missing or unexpected categories to a fallback route. |
| Get a person’s decision | Human Review presents the findings and draft to an authorized reviewer. |
| Send only after approval | Connect the approved outcome to a separate app action that sends the reply. Rejection and expiry routes must not reach that action. |
Each block on the canvas is a Step. You connect Steps and choose which earlier results they use. After a test, inspect the results and errors to check what happened.
AI helps where a fixed rule is not enough
- Investigate unstructured requests with selected read-only lookup tools.
- Classify messages or draft proposed replies into structured fields.
- Use AI Prompt for single-turn tasks and AI Agent for multi-step tool use.
- Route categories with Conditions or Switch using explicit rules.
- Gate external writes behind Human Review so a person approves first.
- Give rejection and expiry explicit non-write routes.
A downstream Human Review gates steps after it; it does not undo tools an agent already called. See AI Prompt for single-call classifications.
From an idea to a running automation
Draft is not a sandbox: Run sends real messages. Pin write Steps or use test destinations; see How testing works.
Building for clients? Keep their work separate
Consultants, agencies and MSPs can manage separate client workspaces through delegated access. Build with the client’s accounts instead of mixing several companies’ automations in one workspace. You can also deploy independent copies of an automation to client workspaces for their own setup and review.
Start with Working With Clients, or follow the first-client guide.
What you can build
- Interactive AI Web Assistant: launch a live chat assistant with session memory in two steps.
- Gmail AI Triage & Inbox Labeling: classify incoming messages, prepare drafts for review, and label or archive mail.
- Lead routing: send each lead down the right branch using a business rule.
- Support ticket triage: classify requests with AI and alert a person when needed.
- Invoice review: turn extracted invoice text into structured data and request approval before writing it to another app.
Evaluate Glow for Your Organization
Follow this path to assess the platform before choosing a pilot:
| Review | Start with | Decide |
|---|---|---|
| Automation architecture | Agents, rules, and human review | Which decisions need reasoning, exact rules, or a person’s approval? |
| Integration framework | How Glow Integrations Work | Which app operations, accounts, events, and agent tools does your process need? |
| Access and data handling | Security & Compliance and MCP Administration | Which access, per-client MCP permissions, retention, and contractual requirements apply? |
| Operations and usage | Operating Automations and Billing | How will you test, activate, monitor, and fund the automation? |
| Client delivery | Working With Clients | How will you separate client workspaces, delegate access, and hand over the work? |
For a technical pilot, choose one representative process and test its success, fallback, and approval routes with controlled data. Review the recorded results and external changes. Agree any required service levels and evidence with your account team.
- Run
- Open Automations → Create new → Workflow on app.getglow.ai, or follow Tutorial 1 to place a Scheduler and AI Prompt on the canvas
- You should see
- Steps connect visually on the shared canvas, and clicking Run in Draft records each Step's output in Executions.
What’s Next?
Build your first automation → Start with a Scheduler and an AI Prompt, then run them manually in Draft and read the result. No external app account is needed.
Already have a process in mind? Choose the right Step, or browse Use Cases & Examples.