AI Integration

Flow lets you use AI models (like ChatGPT or Claude) directly in your workflows. You give the AI instructions in plain English, and it sends back a response that you can use in your logic.

This is useful for tasks like:

• Classifying support tickets ("Is this urgent or not?")
• Summarizing long text
• Generating email responses
• Making decisions that require understanding context

Setting up an AI service

Declare an AI service in the services: block, just like you would for a regular web service:

services:
    Analyst is an AI using "openai/gpt-4o"
    Writer is an AI using "anthropic/claude-sonnet-4-20250514"

The format is provider/model-name:

• openai — ChatGPT models (gpt-4o, gpt-4o-mini)
• anthropic — Claude models (claude-sonnet-4-20250514, claude-haiku-4-5-20251001)

You can declare multiple AI services with different models for different tasks.

Asking the AI to do something

Use the ask statement followed by the service name and your instruction:

ask Analyst to summarize the quarterly sales data
    save the result as summary

log summary

This sends the instruction to the AI model and saves the response. You can write any instruction in plain English — just like you would ask a colleague.

Working with AI responses

When the AI responds, it includes two pieces of information:

• result — the actual text response from the AI
• confidence — a number from 0 to 1 showing how confident the AI is (1 = very confident)

ask Analyst to classify this support ticket as urgent or normal
    save the result as classification

if classification.confidence is above 0.8:
    log "High confidence: {classification.result}"
otherwise:
    log "Low confidence — needs a human to review this"

Including data in your instructions

You can include variables and data in your instruction using curly braces:

set report to order.summary

ask Writer to write a friendly confirmation email for this order: {report}
    save the result as email-text

log email-text

This lets the AI see the actual data and respond accordingly.

Setting up API keys

AI services require an API key to work. These are like passwords that identify you to the AI provider.

Getting your API key:

• OpenAI (ChatGPT): Sign up at platform.openai.com and create an API key
• Anthropic (Claude): Sign up at console.anthropic.com and create an API key

Adding the key to your project:

Create a file called .env in your project folder (next to your .flow files):

OPENAI_API_KEY=sk-your-key-here
ANTHROPIC_API_KEY=sk-ant-your-key-here

Flow automatically picks up the correct key based on which AI provider you're using. Never share this file or upload it to the internet.

Testing without making real AI calls

AI calls cost money (a small amount per request). When you're developing and testing, use mock mode to avoid charges:

flow test my-workflow.flow --dry-run

In mock mode, AI calls return simulated responses. This lets you test that your workflow logic is correct without spending money on real AI calls.

Complete example

Here's a workflow that automatically sorts incoming support tickets:

config:
    name: "Support Ticket Classifier"
    version: 1

services:
    Classifier is an AI using "openai/gpt-4o"

workflow:
    trigger: when a support ticket is received

    set message to ticket.body

    ask Classifier to classify this support message as "billing", "technical", or "general": {message}
        save the result as category

    if category.result contains "billing":
        log "Routing to billing team"
        complete with department "billing" and ticket ticket.id
    otherwise if category.result contains "technical":
        log "Routing to engineering"
        complete with department "engineering" and ticket ticket.id
    otherwise:
        log "Routing to general support"
        complete with department "general" and ticket ticket.id

This reads the ticket message, asks the AI to classify it, then routes it to the appropriate team based on the AI's response.

Next steps

• Services — Connect to regular web services
• Webhook Server — Deploy workflows as web endpoints
• Examples — See more real-world examples