Build an AI workflow from plain English
You do not need to start with code. Start by describing what happens now, what should happen next, where AI helps, and where a human should stay in control.
The workflow we are building
Imagine someone fills out a form on your website asking for information. Instead of manually reading every submission, deciding what it means, writing a reply, and recording what happened, we will map that repeated process into a safe automation.
Form submitted → validate → AI summarizes + classifies → human approval → send follow-up → log result
Step zero
Say the job in normal language.
Before opening an automation tool, write one sentence describing the result you want. Avoid tool names at first. Describe the business process.
TRY THIS PROMPT
“When someone submits my contact form, check that their email and message are present, summarize what they want, label the request as sales, support, partnership, or other, show me the proposed reply for approval, then send it and record the result.”
That single paragraph already contains most of a workflow: a trigger, validation rules, an AI task, decision logic, a human checkpoint, an action, and a record of what happened.
Break it into pieces
Turn the sentence into six blocks.
Trigger
Something happens that starts the workflow — for example, a person submits your website form.
Validate the input
Check that the information you need is actually there before spending AI tokens or sending anything.
Let AI do one clear job
Summarize, classify, extract, rewrite, or decide between a small set of options. Keep the job narrow and measurable.
Add a human checkpoint
If the workflow can email customers, publish content, spend money, or change important data, require approval while you are learning.
Take action
Send the email, create the task, update the record, publish the draft, or hand the result to the next tool.
Log the result
Save what happened so you can debug failures and improve the workflow instead of guessing.
Choose a builder
Use AI to help build the automation itself.
Modern automation platforms can now take natural-language instructions and help turn them into workflows. Two strong options are n8n and Make. The important part is not picking a “winner” yet — it is learning how to describe the process clearly enough that the tool can help you build it.
n8n
n8n's AI Workflow Builder can create, refine, and debug workflows from natural-language descriptions. Paste in the workflow sentence above, inspect what it creates, then refine one section at a time.
Read the n8n guideMake
Make's Maia builder is designed around creating automations through natural-language conversation. Describe the outcome, review the proposed scenario, and keep refining until the flow matches the real process.
Read Make's overviewDo not automate the risky part first.
If a mistake can send the wrong message, spend money, delete data, publish publicly, or affect a customer, keep a human approval step while you test. Automation should remove repetitive work without removing your ability to stop a bad outcome.
Test before turning it loose
Run boring tests on purpose.
A workflow is not finished when the happy path works once. Test the cases that usually create trouble.
- A normal submission with every field filled in
- A submission with a missing email address
- A duplicate submission
- A weird or unclear message
- A case the AI should not handle automatically
The lesson
Plain English is the blueprint, not the magic.
AI can help turn your description into nodes, modules, mappings, and logic. You still own the important part: deciding what the process should do, what counts as a good result, where mistakes matter, and when a person needs to approve the next step.
Once you can describe a repeated process as trigger → inputs → AI job → rules → approval → action → log, you can automate a surprising amount of real work without starting as a programmer.
Your turn
Pick one repetitive task you did this week.
Write it in one paragraph, break it into the seven-part blueprint, and build the smallest safe version first.