Let AI handle 18% of the workflow. Build the rest with rules.

When a workflow includes AI, it is tempting to draw the whole chain as AI steps. That usually makes the workflow more expensive and harder to debug than it needs to be. A rule can check whether an amount exceeds a threshold. Use AI when a step requires interpretation, language, or judgment.

On August 14, Zapier published a simple model for separating those steps. It works even if you build in Make, Power Automate, n8n, or your own code.

Source: Zapier: The 4 primary roles of AI in automated workflows.

Eighteen percent is an observation, not a target

In Zapier's AI Workflow Index, AI accounted for 18 percent of the steps in the analyzed AI workflows. The rest included rules, branches, formatting, app connections, and data movement. The dataset covered the 375 highest AI adopters in a panel of 1,500 mid-market and enterprise Zapier customers from April 15 through June 14, 2026.

The 18 percent figure is not a target for every organization. Even experienced adopters place AI inside a larger, predictable chain. Zapier's modeled comparison also estimated 71 percent lower running costs when AI was reserved for reasoning. That figure is a model estimate, not a promise about your own savings.

Source: Zapier AI Workflow Index, Q2 2026.

Fill in a decision card before choosing a model

Take one recurring workflow and complete the card below. Document the work before choosing the tool.

• Workflow: What must happen from start to finish?
• Trigger: Which event starts the workflow?
• Rule steps: Which steps have one clear answer that can be expressed as conditions, formats, or lookups?
• AI step: Which single step requires interpreting free text, context, or a judgment?
• Receiver: Will a person read the output, will it update a record, steer a decision, or create work?
• Output contract: Which fields, labels, or text must AI return?
• Control: What must be validated, approved, or logged before the next step runs?

The receiver helps you choose AI's role:

• Communicator: AI writes a draft for a person to read.
• Clerk: AI extracts details from unstructured material and fills defined fields.
• Analyst: AI returns a judgment that a rule uses to choose a route.
• Coordinator: AI creates a task or ticket that someone owns.

Zapier found that companies usually started with the Communicator or Clerk. Their output can be reviewed or corrected before another system acts. The Analyst and Coordinator need clearer thresholds, ownership, and logs because their output puts work in motion.

Source: Zapier AI Workflow Index: where AI belongs and how controls change by role.

Example: invoice intake with one AI step

Imagine a shared inbox where supplier invoices arrive as PDF files. A sensible first workflow could look like this:

1. A rule checks that the email has a PDF attachment and saves the original.
2. AI acts as the Clerk and extracts the supplier, invoice number, invoice date, due date, currency, and total amount.
3. A schema check stops the run if a required field is missing or malformed.
4. A conventional duplicate check looks for the same supplier and invoice number.
5. A rule sends new suppliers, large amounts, and uncertain fields to a named reviewer.
6. Approved data is written to the accounting system's queue. The original file and AI output receive the same run ID.

AI handles the part that rules struggle with: interpreting varied documents. File checks, formats, duplicates, amount thresholds, and routing remain normal rules. If AI misses a date, you can see which step failed instead of debugging an agent that owns the entire chain.

Test the card on a representative set of invoices before allowing automatic writes. Measure how often every required field is correct, which documents should stop, and how corrections are logged. Once the Clerk meets the acceptance criteria, you can decide whether another AI role is actually needed.

Build the next workflow from its handoffs

Choose a routine where someone currently reads free text and then fills in a system. Draw the trigger, AI step, rule steps, and review point on one page. That becomes a concrete Tool Forge build brief and lets you change the model or automation platform without reinventing the workflow.