Your first AI automation without coding: from a repeated task to a flow that runs on its own

Imagine that every morning you open your email, copy quote requests into a spreadsheet, classify them by type of service and write an initial reply to each one. That is thirty minutes repeated the same way every day. It is exactly the kind of task that can become an automation.

You do not need to code. You need to understand how a flow is structured and at what point AI adds something.

How an automation works

Every automation, with or without AI, has the same anatomy:

• Trigger. The event that sets it off: an email arrives with a certain subject, someone fills in a form, a row is added to a spreadsheet, it is Monday at 9:00.
• Steps. What is done with the information: filtering, transforming, querying another tool and, this is where AI comes in, classifying, extracting or drafting.
• Action. The final result: saving to a spreadsheet, creating a task, sending a notification, leaving an email draft.

No-code platforms connect these pieces with visual blocks. You choose the trigger, add steps and configure each one with forms, without writing code.

Where AI adds value in the flow

Before generative AI, automations could only work with well-structured data. If the email said "I need a price for 50 T-shirts with a logo", no traditional block understood that this was a quote request for printing. A language model does. That is why AI steps are useful for:

• Classifying. Is it an enquiry, a complaint, an order or spam? Which service is it about?
• Extracting data. Pulling the company name, quantity and deadline out of free text.
• Summarising. Turning a long email or a transcript into three lines.
• Drafting. Preparing an initial reply in your tone, which someone reviews before sending.

Fixed flow or agent: start with the predictable

In its guide to building effective agents, Anthropic distinguishes between two kinds of systems. In workflows, the steps are defined in advance and the model only acts within each one. In agents, the model decides for itself which steps to take and which tools to use. Its advice is to look for the simplest possible solution and add complexity only when needed.

For your first automation, that means a fixed flow: always the same steps, in the same order. It is cheaper, easier to review and, when it fails, you know where. Agents have their place, but also more risks (we explain them in AI agents in production).

No-code tools

Three platforms cover most cases, and all have built-in AI steps or let you connect your OpenAI, Anthropic or Google account:

• Zapier. The easiest to start with, with thousands of connectable apps and a linear interface: one step after another.
• Make. More visual and flexible, with branching flows drawn like a diagram. A good option when you need conditions ("if it is a complaint, do this; if it is an order, do that").
• n8n. Open source, with the option to install it on your own server, which matters if you care where your data goes. It requires a bit more technical confidence.

All three have free plans with monthly limits that are enough for testing. Before committing, check how many runs the plan includes and what each AI step costs.

Step by step: classify requests and prepare replies

A complete example you can adapt to almost any business.

1. Choose the trigger

"When an email arrives at quotes@yourcompany.com" or "when someone submits the contact form on the website". Start with a single source.

2. Add an AI step to classify and extract

Configure the step with a clear instruction and a fixed output format, so the following steps know what to expect:

Analyse this message from a potential customer and reply ONLY with
these fields, one per line:

type: [quote / enquiry / complaint / other]
service: [service name or "not stated"]
quantity: [number or "not stated"]
deadline: [date or "not stated"]
urgent: [yes / no]
summary: [one sentence]

If a detail does not appear in the message, write "not stated". Do not infer it.

Message:
{{email body}}

3. Save the result

Add a row to a spreadsheet with the extracted fields, the date and the sender. You now have a tidy record without copying anything by hand.

4. Prepare a draft, not a send

Add a second AI step that writes an initial reply, and configure it to save it as a draft in your email, not to send it. You review it, adjust what is needed and send it. For inspiration on tone, the Difficult emails template is a good base.

5. Flag what is urgent

If the "urgent" field is "yes", send yourself a notification on your phone or chat tool.

Common mistakes

• Automating something you have not mastered by hand. If you do not know what the result should look like, you will not be able to tell whether the automation is failing.
• Letting AI send directly. A reviewed draft takes seconds; a mistake sent to a customer costs much more.
• Instructions with no output format. If it answers differently each time, the next step breaks.
• Not reviewing the first few days. During the first week, check every run. That is when the cases you had not foreseen appear.
• Connecting everything at once. Each connection is another permission over your data. Add only what the flow needs (and read Prompt injection if the flow processes third-party emails or web pages).

Keep reading

• AI for small businesses: how to choose which task to automate first.
• AI agents in production: when to move from a fixed flow to an agent.
• Prompt injection: the risk of AI obeying someone else: why it matters in flows that read email.
• How to write a prompt that works the first time: how to write the instructions for each step.
• In the glossary: workflow, autonomous agent and API.

Sources

• Anthropic (2024). Building effective agents. Anthropic Engineering.
• n8n. n8n documentation.