AI Is Moving Fast—but Your Business Should Set the Limits
AI agents are becoming more capable of handling tasks, making decisions and connecting with business systems. For a Malaysian SME, that may sound like the answer to staff shortages, repetitive administration and slow customer response times. However, the latest enterprise experience points to an important lesson: the businesses getting the best results are not necessarily giving AI unlimited freedom.
They are giving AI a narrow job, clear instructions and human approval at important decision points. This matters whether you run a Klang Valley services company, a Penang manufacturer, a Johor logistics business or a family-owned retailer in Sabah. An AI agent that sends the wrong quotation, changes an inventory record or exposes customer information can create more work than it removes.
What Happened
According to the source article, Gartner forecasts that more than 40% of current agentic AI projects may not continue until 2028 because of rising costs, unclear business value and weak risk controls. The article also cites McKinsey’s 2026 AI Trust Maturity Survey, which reports that average responsible-AI maturity is only 2.3 out of 4, while approximately 30% of organisations have reached level three or higher in governance and agentic AI controls. Source: VentureBeat
The problem is not simply that AI models make mistakes. It is that an autonomous agent can make several connected decisions before anyone notices. If something goes wrong in a multi-step process, it may be difficult to identify which instruction, data source or decision caused the failure. The article also reports that security and risk issues are the greatest obstacle to wider agentic AI adoption for nearly two-thirds of organisations. Source: VentureBeat
In practice, this means the market is shifting from “who can launch the most autonomous AI?” to “who can operate AI safely and consistently?” The strongest approach is governed orchestration: assigning separate agents to specific responsibilities, recording their actions and requiring people to approve high-impact decisions.
“The goal isn’t maximum control. It’s calibrated control, concentrated where the cost of an error is actually high.”
Why This Matters for Malaysian SMEs
Many Malaysian SMEs do not need an AI agent that controls an entire business process. You may get better results from a small assistant that classifies enquiries, drafts replies, checks stock levels or prepares a daily sales summary. These tasks are useful, repeatable and relatively easy for a staff member to review.
Consider a local wholesaler receiving orders through WhatsApp, email and an online marketplace. One AI agent can extract product names and quantities from incoming messages. A separate workflow can check whether the items appear in the inventory system. However, the agent should not automatically confirm a large order, alter prices or promise a delivery date without a staff member checking the result.
A Malaysian accounting or professional-services firm can apply the same idea to document handling. AI may sort invoices, identify missing fields and prepare a draft reconciliation list. It should not independently approve payments, change accounting records or send sensitive client documents externally. Those actions should remain behind a human checkpoint.
For restaurants, clinics, workshops and retail businesses, customer communication is another practical use case. AI can answer common questions about operating hours, services, delivery areas and appointment availability. When a customer complains, requests a refund, raises a safety issue or shares sensitive personal information, the conversation should move to a trained employee.
Practical control points for your business
| Business activity | AI can assist with | Human approval should remain for |
|---|---|---|
| Sales enquiries | Classifying leads and drafting replies | Discounts, quotations and contract commitments |
| Inventory | Flagging low-stock items and summarising movement | Purchase orders and supplier changes |
| Finance administration | Extracting invoice details and spotting duplicates | Payments, refunds and bank-related actions |
| Customer service | Answering routine questions | Complaints, disputes and sensitive cases |
| Human resources | Organising applications and preparing interview questions | Hiring, rejection and disciplinary decisions |
This structure also helps you meet your responsibilities under Malaysia’s data-protection environment. The Personal Data Protection Department explains that the Personal Data Protection Act 2010 regulates the processing of personal data in commercial transactions in Malaysia. Source: Personal Data Protection Department If an AI tool handles customer names, identification details, phone numbers, health information or employee records, you should know what data it receives, where it is processed and who can access it.
How to Limit an AI Agent Without Making It Useless
The first step is to define one responsibility. Avoid instructions such as “manage customer service” or “handle operations”. A better instruction is “categorise incoming support messages and recommend a response from the approved knowledge base”. The narrower version is easier to test and less likely to create unexpected actions.
Next, separate reading permissions from writing permissions. An agent may be allowed to read product availability but not edit the inventory system. It may prepare a payment list but not submit the payment. It may draft an email but not send it. This separation reduces the damage caused by an incorrect recommendation or compromised account.
Every important action should also have a traceable record. At minimum, keep the original request, the information used, the agent’s recommendation, the employee’s decision and the final action. If a customer asks why a quotation was issued or a staff member questions an automated update, you should be able to reconstruct what happened.
Use approval thresholds that match the risk. A low-risk reply about business hours may be sent automatically. A reply involving a refund, legal promise, personal-data request or unusual discount should wait for approval. You can adjust the threshold as you learn how reliable the workflow is.
The Bigger Picture
The long-term opportunity is not to remove people from every workflow. It is to redesign work so that AI handles speed and repetition while people retain judgment, responsibility and relationships. This is especially relevant for SMEs, where one person may manage sales, purchasing, customer service and administration in the same day.
Start with a process that has clear inputs and outputs. Document how your staff currently handle it, including exceptions and approval steps. Then introduce one limited AI function and measure whether it reduces response time, incomplete records or repetitive work. Do not expand the agent’s permissions simply because the first trial appears successful.
Before deployment, ask yourself four questions:
- Can you explain why the agent made a particular recommendation?
- Does the agent have one clearly defined responsibility?
- Which actions require a human before they happen?
- If the account were misused, which systems and data could the agent reach?
For a Malaysian SME, responsible AI does not need to begin with a large transformation programme. It can begin with a single workflow, limited access, clear approval rules and reliable records. The businesses that benefit most from AI agents will be those that make them useful without allowing them to become unaccountable.
Ready to Streamline Your Operations?
Technology moves fast. Your operations should keep up. AutoRunBiz builds AI systems that run your daily workflows — from WhatsApp order capture to accounting. Book a free 15-min ops audit →