Build Your First AI Agent Without Code: A Small-Business Guide
What an AI agent actually is, how to pick a problem worth automating, and the four-part pattern we teach in our Singapore workshops: capture, instructions, schedule, output.
An AI agent is an AI tool that does one job on its own, on a schedule. It reads something, makes a decision, writes the result somewhere and reports back, without you prompting it each time. That's the difference from a chatbot, which only answers when you ask.
A simple way to see the difference:
• Chat: you ask "what category is this expense?" and it answers. Nothing happens unless you act on it.
• AI agent: a bot reads your expense messages each evening, categorises them, logs them to a spreadsheet and sends you a two-line summary.
• Agentic AI: several agents, or one agent making many kinds of decisions, coordinate across a whole workflow.
For a small business, the middle rung is where the value is today. You don't need a multi-agent system. You need one agent that reliably does one annoying job.
Pick a problem worth building an agent for
A good first agent moves one of two numbers: it grows the top line (more sales, more work done) or protects the bottom line (less time, less cost, fewer errors from doing the same thing by hand).
To find one, look at your own day. What did you redo today because nobody automated it? What takes twenty minutes and shouldn't? Be precise. "Admin" tells you nothing; "retyping the same form into three systems" tells you exactly what to build.
Ideas attendees have built or asked for in our workshops:
• A news and reading digest that summarises what's worth reading each morning.
• Turning scattered marketing ideas into draft social posts once a week.
• A first draft of tender or bid proposals from a folder of past submissions.
• Expense capture: message a receipt photo, get a categorised spreadsheet row.
If you're not sure the task needs AI at all, check it against Does your business actually need AI? first.
The four-part pattern: capture, instructions, schedule, output
Every simple agent we build has the same shape. Once you've built one, the next is mostly renaming things.
• Capture: where the input arrives. A Telegram chat, an email label, a shared folder or a spreadsheet you add to during the day.
• Instructions: a plain-English brief telling the agent what to read, what to decide, what to write and what never to do.
• Schedule: when it runs, for example every evening at 9pm or every Monday at 8am.
• Output: where the result lands and how you hear about it, such as a spreadsheet row plus a short message.
This split between fast capture during the day and batched processing on a schedule isn't a workaround. For most admin work it's the better design, because you get one tidy summary instead of a stream of notifications.
The tools you need (no coding)
You can build this kind of agent without writing code. In our workshops we use:
• Claude Cowork as the agent. It can read your files and connected apps, follow standing instructions, and run as a scheduled task. A paid Claude plan is needed to complete the setup. See what Claude Cowork is.
• Telegram as the capture and notification layer, because creating a bot is free and takes a couple of minutes through Telegram's official BotFather.
• A spreadsheet as the memory, so you can see and correct everything the agent writes.
The full, tested setup, including every error message you're likely to hit, is in Connect Telegram to Claude Cowork. The materials from our Build Your Own AI Agent session at Zendesk Singapore are also public.
Write instructions that hold up
The instructions are where most first agents succeed or fail. Write them like a brief for a careful new hire, not a one-line prompt.
Good agent instructions include:
• Exactly what to read, and how to know what's new since the last run.
• The fields to fill, with a fixed list of allowed values (for example, a closed list of expense categories).
• What to do when unsure. For example: "If you cannot work out the amount, leave it blank and mark the row unsure. Never guess a number."
• What never to touch, such as "do not edit or delete existing rows".
• What to send when there's nothing to do, so silence always means something broke rather than "no news".
That last pair of rules turns a clever demo into something you can trust. For reusable patterns, see recurring tasks on Claude.
Test it once, then break it on purpose
Run the agent once by hand before you schedule it. Confirm the output lands where you expect and looks right. Only then put it on a daily or weekly schedule.
Then break it deliberately. Send it something ambiguous, such as a blurry receipt, a message with two amounts, or a note in a different language. A well-instructed agent should flag what it isn't sure about, not guess. If it guesses, tighten the instructions and try again.
This habit of writing the rules and guardrails so an agent works without you checking every output is what people call agentic engineering. You're doing a small version of it the first time you test an edge case.
What not to hand an AI agent yet
Start with low-risk, reversible work. Keep these out of your first agents:
• Anything that sends money or commits you to a contract.
• Messages to customers that go out without you seeing them first. Have the agent draft; you approve.
• Confidential client or personal data you haven't cleared for use with an AI tool. Singapore's PDPA still applies. See our PDPA and AI compliance guide and whether it's safe to connect Claude to work Gmail.
Once an agent has run cleanly for a few weeks, widen what it's allowed to do one step at a time. If you'd like to build your first one with help, our events page lists upcoming hands-on sessions, and AI Foundations covers the same ground for teams.
Frequently asked questions
What is an AI agent in simple terms?
An AI agent is an AI tool that does one job on its own, usually on a schedule. It reads an input, makes a decision, writes the result somewhere and reports back, without you prompting it each time. A chatbot only answers when asked; an agent keeps working in the background.
Can I build an AI agent without coding?
Yes. Tools like Claude Cowork let you describe an agent's job in plain English, connect it to your files or apps, and run it on a schedule. Pair it with a free Telegram bot for capture and notifications, and a spreadsheet as its memory, and you can build a working agent without writing code.
What is a good first AI agent for a small business?
Pick one repetitive, low-risk task you do weekly: capturing expenses from receipt photos, summarising industry news each morning, turning scattered ideas into draft social posts, or drafting proposals from past examples. Choose something where a wrong first draft is easy to spot and fix, and where the time saved is obvious.
How do I stop an AI agent from making mistakes?
Write detailed instructions with fixed allowed values, tell it to flag rather than guess when unsure, forbid editing existing records, and make it report even when there is nothing to do. Run it once by hand, then feed it ambiguous inputs on purpose before scheduling it. Keep a human approving anything customer-facing.
Is it safe to give an AI agent access to business data?
It can be, if you limit it. Start with data you would be comfortable sharing with a careful new hire, avoid confidential client or personal data until you have checked PDPA obligations, and never let a first agent send money or unreviewed customer messages. Widen its access gradually once it has run reliably.
