You can build a working AI agent in about 45 minutes, with no code, inside an automation platform you may already pay for. Make, Zapier and n8n all ship agent features on standard plans from $9 to $20 a month — you do not need a dedicated agent builder to start.
This walkthrough builds one specific agent: an enquiry triage assistant that reads incoming email, decides what each message is, pulls the relevant context, and drafts a reply for you to approve. It never sends anything on its own. That constraint is deliberate, and it is what makes this safe to run on a real inbox on day one.
Key takeaways
- Build it where you already are. Make Core $9/mo, Zapier Professional $19.99/mo annual, n8n Starter €20/mo annual all include agent capability.
- Start with draft-for-approval, never send. One setting removes almost all the risk.
- An agent needs three things: a goal, tools it may use, and a stopping rule. Miss the stopping rule and it loops.
- Expect the cost to vary run to run — the agent chooses its own step count, unlike a fixed workflow.
- Shadow it for two weeks before you trust a single output.
What Is an AI Agent, Exactly?
An AI agent is an automation that decides its own steps. A normal workflow follows a path you drew in advance; an agent is given a goal, a set of tools, and permission to choose which tool to use next based on what it finds.
That single difference explains everything else about them — why they handle messy inputs better, why their cost is unpredictable, and why they need containment. If you can write the steps down in advance, you don’t want an agent, and rules will beat it on cost and reliability.
| Fixed workflow | AI agent | |
|---|---|---|
| Steps | You define them | It selects them at runtime |
| Handles unexpected input | Breaks or skips | Adapts, sometimes wrongly |
| Cost per run | Fixed and countable | Varies with steps taken |
| Debugging | Find the broken step | Read the reasoning trace |
| Right for | Repeating, known processes | Judgement over messy input |
What Will This Cost to Run?
Two meters run at once: your automation platform’s usage and the AI model’s tokens. Most beginners budget for the first and get surprised by the second.
| Platform | Entry plan | Billing unit | Effect on an agent |
|---|---|---|---|
| Make | Core $9/mo · 10,000 credits | One module run | Every step the agent takes bills separately |
| Zapier | Professional $19.99/mo annual · 750 tasks | One action reaching another app | Filters and formatters are free; app calls bill |
| n8n | Starter €20/mo annual · 2,500 executions | One whole workflow run | Step count doesn’t change the price |
For a triage agent handling 200 emails a month, expect the platform side to sit comfortably inside an entry plan and the model side to be a few dollars. The variance is the thing to watch, not the average — we break the arithmetic down further in what running an AI agent actually costs.
Step 1: Decide the One Job
Pick a single, narrow, repetitive job with a clear finish line. “Handle my inbox” is not a job — it’s a wish. “Read new enquiries, classify them as quote request, support issue or spam, and draft a reply” is a job.
Write the goal in one sentence before you open any tool. If you can’t, the agent won’t be able to either, because that sentence becomes its instructions.
Good first agents: enquiry triage, meeting-note follow-ups, invoice chasing drafts, lead research summaries. All read-heavy, all draft-only, all easy to check.
Step 2: Give It Tools and Take Away Permissions
An agent’s tools are the actions it may take. Grant only what the job needs, and prefer read-only versions of everything.
| Tool | Grant? | Why |
|---|---|---|
| Read new email | Yes | The input. No risk on its own |
| Search past emails or CRM | Yes, read-only | Context is what makes the draft good |
| Create a draft | Yes | The output, held for you |
| Send email | No | The one permission that turns a mistake into damage |
| Update CRM records | Not yet | Add later, restricted to one field |
| Delete anything | Never | No upside worth the downside |
Use a dedicated account or a filtered view rather than your whole mailbox where the platform allows it. An agent can be manipulated by the content it reads, so the less it can reach, the smaller the worst case — the reasoning we set out in AI agent security.
Step 3: Write the Instructions
Instructions are where beginners under-invest. Three paragraphs beats three sentences, and being explicit about failure beats being clever about success.
Cover four things: what the agent is, what counts as done, what it must never do, and what to do when unsure. That last one matters most. Without an explicit escape route, an agent facing an ambiguous email will invent something plausible rather than stop.
A workable skeleton, in your own words: “You triage incoming enquiries for a small business. For each message, classify it as quote request, support issue, or spam. Search past correspondence for relevant context. Draft a reply in a warm, brief, professional tone. Never invent prices, dates or commitments. If the message is ambiguous or mentions a complaint, refund or legal matter, label it FOR HUMAN and draft nothing.”
Step 4: Set the Stopping Rules
Agents loop. Left uncapped, one that can’t complete its goal will keep trying, and each attempt costs money. Every platform lets you cap this, and the setting is easy to skip past.
| Limit | Sensible starting value | What it stops |
|---|---|---|
| Maximum steps per run | 8–10 | Infinite reasoning loops |
| Maximum runs per hour | 20 | A mail flood draining your plan |
| Timeout per run | 2 minutes | Silent hangs |
| Monthly spend cap | Whatever you’d shrug at | The bill you didn’t expect |
Step 5: Shadow It Before You Trust It
Run it in draft-only mode for two weeks and read every output alongside what you would have written. You’re grading two things: how often it’s right, and how badly it’s wrong when it’s wrong.
The dangerous failure is not the obvious error. It’s the reply that reads well, sounds confident and is subtly incorrect — a slightly wrong lead time, a price from an old quote. Those ship, because nobody reads a good-looking draft carefully.
Keep a tally. If it drafts something you’d have sent unaltered eight times in ten, it’s earning its place. Below six, the instructions need work rather than the model.
Resist the urge to fix a weak agent by switching to a more expensive model. In our experience the cause is almost always an instruction that never said what “done” looks like, or a missing rule about what to do with an edge case. Rewriting three sentences is free; upgrading the model is a permanent line on the bill and usually fixes less than you hoped.
When Should You Move to a Dedicated Agent Builder?
When you’ve confirmed a specific capability your platform lacks — persistent memory across sessions, computer use, or multi-agent handoff. Not before. Dedicated builders start around $37 a month and reach $200, and several don’t publish what consumes a credit.
Build the first one where you already are. You’ll learn what you actually need from running it, and that turns tool selection into a decision with evidence behind it.
Frequently Asked Questions
How do I build an AI agent without coding?
Use the agent feature inside an automation platform you already have — Make from $9/month, Zapier Professional from $19.99/month annual, or n8n Starter from €20/month annual. Give it a narrow goal, read-only tools, a draft-only output and step limits.
How long does it take to build a first AI agent?
About 45 minutes for a triage agent like the one above, most of it spent writing instructions rather than clicking. Budget a further two weeks of shadow running before you rely on its output.
What is the easiest platform for a first AI agent?
Whichever you already use. If starting fresh, Make is cheapest at $9/month, and n8n’s per-execution billing is kindest to agents because step count doesn’t change the price. Zapier has the widest app coverage.
Should my first AI agent be able to send email?
No. Have it create drafts and approve them yourself. Send permission is what converts a manipulated or mistaken agent into a customer-facing problem, and you lose very little convenience by keeping a human in the loop initially.
Why does my AI agent keep looping?
It cannot reach its stopping condition, usually because the goal is vague or the tools can’t satisfy it. Set a maximum step count of 8–10, then tighten the instructions to state exactly what “done” means and what to do when unsure.
How much does it cost to run an AI agent?
Two meters: platform usage and model tokens. For roughly 200 runs a month an entry plan usually covers the platform side with a few dollars of model cost. The variance matters more than the average, because the agent chooses its own step count.
The Bottom Line
The build is the easy part. Choosing a job narrow enough to succeed, and withholding the permissions that make failure expensive, is the actual skill.
Pick one repetitive reading-and-drafting task. Grant read-only tools plus a draft action, nothing else. Write instructions that say what to do when unsure. Cap the steps. Then shadow it for a fortnight and count how often you’d have sent what it wrote.
If it earns trust, widen it one permission at a time. If it doesn’t, you’ve spent 45 minutes and learned exactly where the judgement in that job lives. If the fixed-workflow version turns out to be what you actually wanted, the same beginner path runs through your first Zapier workflow and your first Make scenario. Next, see AI agents for business for where they fit overall, or browse the tutorials hub for the fixed-workflow versions.
Want one practical automation you can set up in 15 minutes, twice a month? Join the free newsletter.
Sources
- Make — Pricing plans
- Zapier — Pricing plans
- n8n — Pricing plans
- Zapier Help — What counts as a task
- OWASP — Top 10 for Large Language Model Applications
All prices read off each vendor’s own pricing page, last verified 11 August 2026. The step limits, shadow-running period and permission ordering are our own recommendations based on the failure modes described, not vendor guidance.
