Turn a business process into an agentic workflow

Most AI automation projects fail for the same reason: they start with the tool instead of the process. This article gives you a method to go from “we spend hours on this every week” to a reliable workflow, where AI handles the judgement and plain automation handles the rest. It is for business owners and team leads. No coding needed for the thinking part.

Marrallisa covered this method in the DataCamp webinar How to turn a business process into an agentic workflow.

Step 1: Pick the right process

Good first candidates are:

• Frequent: it happens daily or weekly.
• Bounded: it has a clear start and a clear end result.
• Annoying but low risk: a mistake is easy to spot and fix.

Examples: sorting incoming requests, processing invoices, qualifying leads, preparing weekly reports. Avoid processes where one wrong action is costly, such as payments, until you have experience.

Step 2: Map the process as it is today

Sit down with the person who does the work and write every step down, in order. For each step, note:

1. What starts it.
2. What information they look at.
3. Which systems they use.
4. What they decide, and on what basis.
5. What the output is, and who receives it.
6. What goes wrong regularly.

A numbered list or a simple flowchart is enough. Write down the exceptions too. They are where automations break.

Step 3: Find the decision points

Go through your map and mark each step as one of two types:

Type	Recognise it by	Build it with
Rule-based	The same input always gives the same output. You can write the rule down.	Regular n8n nodes: IF, Switch, Edit Fields, app nodes
Judgement	Reading free text, recognising intent, summarising, writing a draft	An AI Agent or LLM node

Here is a quote request process, mapped this way:

Step	Type	Built with
New email arrives in quotes@	Rule	Email trigger
Is it a quote request, a question or spam?	Judgement	AI classification
Extract company, service and budget	Judgement	AI with structured output
Is the company already a customer?	Rule	CRM lookup
Draft a reply with the right package	Judgement	AI Agent
Send the reply	Rule, high risk	Human approval, then send
Log the request	Rule	Data Table or CRM

Step 4: Keep as much as possible deterministic

Rule-based steps are fast, free and always behave the same way. Each judgement step adds cost, delay and uncertainty. So:

1. Use AI only where you marked judgement.
2. Give each AI step one clear job, with a structured output (fields, not prose).
3. Route on that output with normal nodes, such as a Switch on category.
4. Give an agent only the tools it needs for its job. See The AI Agent node explained.

Tip: If you find yourself writing a very long system message full of rules, some of those rules probably belong in IF and Switch nodes instead.

Step 5: Put a human in the loop

Decide in advance where a person must check the work. Typical places:

• Before anything leaves the company: emails to customers, published content.
• Before anything is changed or deleted in a core system.
• When the AI is unsure. Ask it for a confidence field and route low scores to a person.

In n8n you can add human review to specific agent tools, or use a send-and-wait message in Slack, email or Telegram that pauses the workflow until someone approves.

Step 6: Add guardrails

1. Input and output checks: the Guardrails node can catch jailbreak attempts, personal data and keywords before text reaches the model or leaves the workflow.
2. Limits: set Max Iterations on agents and a timeout on long workflows.
3. Logging: store each AI decision and its input in a Data Table, so you can review mistakes.
4. Alerts: add an error workflow. See Error handling and alerts.

Step 7: Roll out in stages

1. Shadow mode: the workflow runs alongside the person and only produces drafts. Compare results for a week or two.
2. Assisted: the person approves every action with one click.
3. Automated: remove approval for the cases that proved reliable. Keep it for the risky ones.

Measure time saved and error rate at each stage. That tells you when to move on, and gives you the numbers to justify the next process.

If you would like help mapping your first process, book a strategic call: https://calendly.com/marrallisa/30min

Next steps

• The AI Agent node explained
• Build workflows with Claude
• Process incoming email with the IMAP trigger
• Store data with Data Tables