By Rakibul Hasan, Moonlab
You can build a simple AI agent without writing code using a visual automation tool such as n8n, Make or Zapier. The pattern is the same in each: a trigger starts it, an AI step reads or decides, tool steps act, and a person approves anything risky.
This guide walks through the steps with a hypothetical example: an agent that reads new enquiries, drafts a reply and logs the lead. Tool features and pricing change often, so use this as a method and check each tool's current documentation for exact menus and limits.
Step 1: choose a small, safe task
Pick a job that happens often, follows a pattern and is easy to check. Good first choices are sorting enquiries, drafting replies, summarising meetings or extracting details from emails. Avoid anything that moves money or makes promises.
Write the task as a sentence: "When a new enquiry arrives, classify it, draft a reply from our price list and service area, and save it to the CRM for approval." If you cannot write it like that, narrow it down.
Step 2: pick the tool
All three tools can do this. The choice depends on your team and data.
- Zapier is the simplest for short linear flows and works well if no technical person will maintain it.
- Make has a visual canvas that handles branches and data reshaping well.
- n8n offers more control, code steps and the option to self-host, with a steeper learning curve.
Step 3: set the trigger and gather the data
Choose what starts the agent: a new form entry, a new email in a label, a new row in a sheet. Then collect everything the AI will need in one place: the message, the customer's history from your CRM, and your reference information such as services, areas and prices.
The more relevant context the AI step gets, the better the output. Do not paste your whole knowledge base into every call. Fetch only what is relevant, for example the three most relevant FAQ entries.
Step 4: add the AI step with clear instructions
Add the AI model step and write instructions that cover the role, the task, the rules and the output format. Ask for a structured output such as a category, a confidence level and a draft reply, so later steps can read it reliably.
- Role: "You write replies for a small cleaning company."
- Rules: use only the information provided, never quote prices not in the list, and say a team member will confirm if unsure.
- Format: return the category, a one-line summary and the draft text in separate fields.
- Fallback: if the request does not fit any category, return "needs human".
Step 5: add tools and branches
Next, connect the steps that act: look up the contact in the CRM, create a record if none exists, save the draft, and notify the right person. Use a branch on the AI's category field. Enquiries labelled "quote" go one way, "complaint" another, "needs human" straight to a person.
Keep the number of tools small. Each extra capability adds ways to fail. For a deeper look at what agents can reliably do, see moonlab.space/guides/ai-agents-for-small-business.
Step 6: add guardrails and a human check
Before the agent sends anything to a customer, add an approval step: a message to you or your team with the draft and Approve and Edit buttons or links. Add an error path that alerts a named person when a run fails. Set usage limits in your AI provider account so a bug cannot create a surprise bill.
Never put passwords or card data into prompts. Mask personal details you do not need. If you handle EU or UK data, read moonlab.space/guides/gdpr-ai-automation first.
A worked example: an enquiry reply agent in n8n, Make or Zapier
Say you run a small cleaning company and want new website enquiries handled faster. In the tool of your choice, the workflow has six blocks. A trigger fires when the website form is submitted. A lookup step searches your CRM or sheet for the email address. An AI step receives the message, your price guide and your service area, and returns three fields: category, a confidence note and a draft reply. A branch sends anything labelled complaint or needs human to the owner directly. Everything else goes to an approval message in chat with the draft. When approved, the email is sent and the lead is saved with a status.
Build it one block at a time and test after each block, using the tool's built-in run history to see what each step received and returned. Look at where data goes missing. Most first-time problems are an empty field passed into the AI step, a trigger that fires twice, or an AI answer that comes back in the wrong format. Fix the instruction text first, for example by asking for plain labelled fields only.
- Mistake: starting with a big, ambitious agent. Build the smallest useful version first.
- Mistake: putting your API keys in a shared screenshot or workflow notes. Use the tool's credential store.
- Mistake: no error path. Add a step that messages a named person when a run fails.
- Mistake: testing only with clean examples. Include spam, short messages and other languages.
- Mistake: building in a personal account. Use accounts the business owns so the workflow survives staff changes.
- Decision rule: if a plain rule can do the step, use a rule and keep the AI for reading or writing text.
Step 7: test, then go live in stages
Test with at least twenty real past enquiries, including strange ones: very short, in another language, angry, spam. Check each draft. Fix the instructions and rules, not just single outputs.
Then run in shadow mode for a week or two: the agent drafts, but your team still replies the usual way and compares. When the drafts are reliably good, switch to approve-and-send. Review a sample of runs every week and keep a log of changes. Once the basic version is stable, you can extend it, for example with the follow-ups in moonlab.space/guides/lead-follow-up-sequence.
If you would rather have someone build and monitor it, our free audit at moonlab.space/contact is a good place to start.
