How to Create Your First AI Workflow

Map the Manual Process

Before building anything, write down exactly what happens in the current manual process. For email triage, this looks like: (1) Email arrives in shared inbox, (2) Agent reads the email content, (3) Agent identifies the topic and urgency, (4) Agent checks the customer account for context, (5) Agent assigns the email to the appropriate team or queue, (6) Agent adds internal notes about the classification reasoning.

Document every decision point and the criteria used. What makes an email "urgent" versus "normal"? Which keywords or situations trigger escalation? What information does the agent look up in the customer account? Write these criteria explicitly because they become the instructions for the AI model. Interview the people who currently do this work, as they know edge cases and patterns that are not documented anywhere.

Identify AI Insertion Points

Look at each step in the manual process and determine whether it requires: (a) structured data movement (traditional automation handles this), (b) reading and understanding unstructured content (AI handles this), or (c) human judgment on novel or high-stakes situations (keep this manual).

In the email triage example, steps 2 and 3 (reading the email and identifying topic/urgency) are clear AI insertion points. Step 4 (checking the customer account) is traditional automation via an API call. Step 5 (assigning to a team) is conditional routing based on the AI classification. Step 6 (adding notes) is AI-generated content summarizing the reasoning. The trigger (email arriving) is a standard email trigger.

Build the Trigger

In your workflow platform, create a new workflow and configure the trigger. For email triage, use an email trigger connected to the shared inbox. Configure it to capture the sender address, subject line, body text, and any attachments. Test the trigger by sending a test email and verifying the data appears correctly in the workflow execution log.

If your trigger is a webhook (from a web form, CRM, or other system), configure the webhook URL in the source system and define the expected payload structure. Send a test event and confirm the data arrives in the correct format. Getting the trigger right before building the rest of the workflow prevents debugging issues later.

Create the AI Processing Step

Add an AI node after the trigger. This is where the core intelligence happens. Write a prompt that gives the model clear instructions, the input data, and the expected output format. For email classification:

"You are classifying customer support emails. Read the email below and provide your analysis in the following JSON format: {category: one of [billing, technical, feature-request, complaint, general], urgency: one of [critical, high, normal, low], summary: a one-sentence summary of the issue}. Email subject: [subject]. Email body: [body]."

Select the appropriate model. For classification tasks, Claude Haiku or GPT-4o-mini provides good accuracy at low cost. For tasks requiring deeper reasoning or content generation, use Claude Sonnet or GPT-4o. Test the AI node with 10-15 sample emails and verify the classifications match what a human would decide. Adjust the prompt based on any misclassifications.

Add Routing and Actions

After the AI node, add conditional branches based on the classification output. Create a branch for each category: billing emails route to the billing team queue, technical emails route to the engineering support queue, complaints route to the customer success team, and so on.

Within each branch, add the output actions: create a ticket in the ticketing system with the AI-generated summary and classification, send a Slack notification to the relevant team channel, and update the customer record with the interaction. For critical urgency items, add an immediate alert via SMS or phone call to the on-call person.

Add a fallback branch for cases where the AI output does not match the expected format or where the classification confidence is low. Route these to a manual review queue with the AI assessment included so the reviewer can quickly validate or correct the classification.

Test with Real Data

Collect 30-50 emails from the actual inbox that represent the full range of categories, urgency levels, and edge cases. Run each email through the workflow manually (using the platform test/debug mode) and record the AI classification alongside the classification a human would assign.

Calculate the agreement rate. If the AI matches the human classification on 85% or more of the test cases, the workflow is ready for refinement. If the agreement rate is below 80%, the prompt needs significant revision. Look at the misclassified cases for patterns: are certain categories confused with each other? Are multi-topic emails handled poorly? Are certain writing styles misinterpreted?

Refine and Deploy

Revise the AI prompt to address the specific misclassification patterns found during testing. Add examples of tricky cases directly in the prompt (few-shot prompting) to guide the model. Re-test with the same sample data plus any new edge cases you discovered. Repeat this cycle until the agreement rate meets your target.

Add error handling for API failures, malformed responses, and timeout scenarios. Configure retry logic with a fallback model if the primary model is unavailable. Set up monitoring to track classification accuracy over time by periodically sampling automated decisions for human review.

Enable the trigger in production. For the first week, have the team review every automated classification to catch any issues that testing missed. After confirming production accuracy matches test accuracy, reduce human review to a daily sample of 10-20 classifications. Continue monitoring metrics and refining prompts as new patterns emerge.