How to Build an AI Agent for a Business: 8 Steps
An 8-step method to build an AI agent for a Malaysian business: pick one job, limit tools and permissions, test on real cases, add approval and logs.
To build a business AI agent, pick one narrow job, list the tools it may use, give it the least permission it needs, write clear instructions, test it on 50 or more real past cases, add human approval for risky actions, log everything and launch in shadow mode first. The work is mostly job definition and testing.
Part of the guide: AI Automation in Malaysia: The 2026 Guide for SMEs

On this page
- Step 1: What is the one job?
- Step 2: What does a person do today?
- Step 3: What tools does the agent get?
- Step 4: What permissions does it have?
- Step 5: How do you write the instructions?
- Step 6: How do you test it?
- Step 7: Where does a person approve?
- Step 8: How do you launch?
- What should you do next?
- Sources
- Frequently asked questions
An AI agent is a language model that picks its own tools to finish a task. The method below works with n8n, Make, code or an agent platform, because the hard decisions are the same.
Part of the guide: AI automation in Malaysia. The steps reflect our own build practice and public guidance on LLM application security.
Step 1: What is the one job?
Write the job as a sentence a new hire could follow: "Read each incoming supplier email, classify it as invoice, delivery query or other, and create the matching task in the finance board." If you cannot write it in a sentence, it is two agents or no agent.

Step 2: What does a person do today?
Watch the person do it three times and write the steps, the exceptions and the systems touched. The exceptions become your test cases.
Step 3: What tools does the agent get?
List each action as a tool with a plain description: "search_orders(customer)", "create_task(title, owner)". Mark each as read or write. Start with read-only tools and add write tools later.
| Tool | Type | Needs approval? |
|---|---|---|
| Look up order | Read | No |
| Draft reply | Write (draft only) | Yes, person sends |
| Create task | Write (internal) | No, but logged |
| Refund or payment | Write (money) | Always |

Step 4: What permissions does it have?
Give it its own service account, never a person's login. Limit it to the folders, tables and fields it needs. If it only reads orders, it cannot edit them. This is the single best defence against both mistakes and attacks. Details in AI agent governance checklist.
Step 5: How do you write the instructions?
Short, specific, with examples. Say what to do, what not to do, and when to stop and ask a person. Include two or three real examples of correct output. Tell it to answer only from the data supplied, and to say "I do not know" when the data is missing.
Step 6: How do you test it?
Collect 50 to 100 real past cases, including the awkward ones, and write the correct result for each. Run the agent on all of them and count. Set the pass mark before you run. Re-run the whole set whenever you change the prompt or model. For how a typical n8n agent is wired, see How an n8n AI agent works.
Step 7: Where does a person approve?
Put approval in front of anything that spends money, deletes, sends to customers in bulk or changes records you cannot easily restore. Approval can be a WhatsApp or Slack button. As accuracy proves out, you can widen what runs without approval, one category at a time.
Step 8: How do you launch?
- Shadow mode for one to two weeks: the agent proposes, the person acts, you compare.
- Limited live: one team or one product line.
- Full live with alerts on errors, cost per day and unusual volume.
Log the input, the output, the tool calls and who approved. You will need it for debugging and, if personal data is involved, for answering questions under the PDPA. See Self-hosting n8n and PDPA and Prompt injection: the risk for business agents.
What should you do next?
Choose one job and run steps 1 to 3 this week. If you want help, our AI Opportunity Audit ranks your candidate jobs by value, and we build and run agents. To train your team to do this themselves, see our HRD Corp Claimable courses. Book a free call.
These eight steps are the backbone of the capstone in our AI automation and agents course, done on a real process.
Sources
- Anthropic, Building effective agents (start simple, add complexity only when needed) (checked November 2026).
- OWASP GenAI Security Project, LLM06 Excessive Agency and LLM01 Prompt Injection (checked November 2026).
- OpenAI, Working with evals (test sets before launch) (checked November 2026).
- Model Context Protocol, What is MCP? (checked November 2026).
Frequently asked questions
Which AI model should I use for an agent?
Use a current model from a major provider and test two on your own cases. Choose by accuracy on your test set, cost per run and where the provider processes data, not by headlines. Keep the model name in one setting so you can swap it.
How many tools should an agent have?
As few as the job needs. Every extra tool is more ways to go wrong. Three to five well described tools is a good first agent.
Can a non-developer build one?
A simple agent in n8n can be built by a trained non-developer. Giving it write access to customer or financial systems needs a developer or partner because of permissions, logging and testing.
How do I know the agent is good enough?
Set the pass mark before you test, for example 95 percent correct on 100 past cases and zero wrong actions on the risky category. Then measure, rather than judging by a few good demos.