Classic automation runs on rules: “when a form arrives, write it to a sheet.” AI automation inserts an understanding step: “when a form arrives, summarise what they want, judge urgency, route to the right person.” The difference is being able to automate work that cannot be written as a rule — free text, photos, voice. This article keeps the concept plain and ends with three concrete examples.
Three parts
- 01Trigger and connections — form, email, WhatsApp, order system. This layer is built with a tool like n8n.
- 02Understanding step — a language model is told “put this message in one of these three classes” or “summarise this form in two sentences”. It returns a structured answer (for example JSON).
- 03Action — branching by class: urgent goes to the phone, routine goes to the sheet, an order goes to shipping.
Automation exists without the understanding step; for most businesses that is actually the first need. AI is only necessary when the input is free-form.
Where it works
- Classifying incoming messages. Splitting WhatsApp and Instagram DMs into “price question / order / complaint / spam” and treating each differently.
- Form summaries. Reducing a long brief to two sentences for the team.
- Order validation. Is the address incomplete, is the item in stock, has payment arrived — check, and ask the customer if something is missing.
- Content preparation. When a new product is entered, draft a description, social copy and tag suggestions. Draft — not publish.
- Market scans. Reading competitor pages on a schedule and reporting price or product changes.
Where it does not
- Alone where precision is required. Invoice amounts, medical information, legal answers; models work approximately, the final word must be human.
- Unbounded customer chat. A bot without limits makes false promises. What the bot answers and what it hands to a human is written down first.
- When the data is not yours. If customer messages go to a model, which provider they go to, whether they are stored, and whether your privacy notice says so all matter.
Example 1 — WhatsApp AI automation
A message from the WhatsApp Business API lands in n8n. The model classifies it and detects language. If it is a “price question”, a reply draft is produced from the price list and goes to approval; once approved, it is sent. If it is a “complaint”, a notification goes straight to the phone and the bot does not reply. Messages at night get a “we will reply in the morning” template. Rule: the bot makes no promises, the human does.
Example 2 — Order notification and validation
A new-order webhook arrives from the e-commerce system. The workflow checks the stock sheet and whether the address is incomplete. If so, a one-question message goes to the customer; if complete, a shipping label is created and a summary goes to the business. Detailed steps are in the order notification workflow.
Example 3 — Form follow-up
When a website form arrives, the model summarises it and tags sector and urgency; the team sees the summary in Telegram and finds the full record in the sheet. An unanswered form is re-flagged after 24 hours. Steps are on the form follow-up workflow page.
How to start
For a week, write down the tasks that repeat and interrupt your day. Pick the single task at the top of the list. Build it first without the understanding step (rules only); add the model only if that is not enough. The first workflow should be small and boring; big workflows are combinations of small ones.
If someone builds it for you, the criterion is: the workflow must be built in your accounts, documented and handed over. Our AI automations service works that way; write the first item on your list in the brief form and we will tell you which workflow solves it and how long it takes.