Most Malaysian business owners have now tried AI in some form. Fewer have turned it into something the business depends on. A recent AWS-commissioned study, Unlocking Malaysia's AI Potential 2026, puts numbers on that gap, and they are worth reading closely before you sign up for another tool.
What the AWS study found
According to the study, the share of Malaysian businesses using AI rose from 27% to 38% in a year, as reported by Ecommerce News Asia. That works out to more than 3.4 million firms, according to The Sun.
The less flattering finding sits underneath. Most of that use is basic: drafting text, answering simple questions, summarising documents. The Edge Malaysia reported that the majority of businesses remain at this level. Only 19% of firms using AI have a formal plan to scale it across more than one function, as The Sun and Bernama both noted.
The study also found that most firms obtain AI through outside vendors rather than building it, and that few have set clear accountability or escalation rules for what happens when the AI gets something wrong (Ecommerce News Asia).
Put those findings together and a familiar picture appears. Staff use a chat assistant here and there. A vendor has switched on an AI feature inside an existing system. Nobody owns the result, nobody measured the starting point, and nobody can say whether the business is better off.
Why "basic use" stalls
Basic use is not a failure. A sales executive who drafts quotations faster, or an admin officer who summarises a long tender document in minutes, is getting real value. The problem is that this value stays with the individual. It does not change how the business runs, and it rarely shows up in the numbers an owner watches.
Three things usually keep an SME stuck at this stage.
The unit of change is a tool, not a process
When AI arrives as a subscription, the natural question is "who should use this?" A better question is "which piece of work should change?" A tool can be used a little by everyone and fully by no one. A process either runs differently or it does not.
Nobody measured the before
If you do not know how many hours your team spends keying supplier invoices each month, you cannot tell whether an automation saved anything. Without that baseline, every AI project ends in opinion rather than evidence, and opinion is a weak case for spending more.
Nobody owns the exceptions
AI handles the routine cases well and the odd ones less well. A blurred invoice, a customer writing in a mix of Malay and English, an order with a non-standard discount. If no one is named to deal with these, staff quietly stop trusting the system and go back to the old way. This matches the accountability gap the study describes.
Count processes, not chatbots
The more useful measure for an SME is not how many AI tools you have, but how many business processes now run end to end with AI doing part of the work. One finished process teaches you more than five half-used tools.

Good first candidates share four traits: they repeat every day or week, they follow mostly predictable rules, they involve moving information from one place to another, and someone on the team already complains about them. Two common examples in Malaysian SMEs:
- Supplier invoices into SQL Account or AutoCount. Invoices arrive by email, WhatsApp and post. Someone reads each one and types the supplier, date, line items and tax into the accounting system. Document AI can read the invoice and prepare the entry, with a person checking anything the system is unsure of. With e-invoicing now live, the invoices you receive deserve the same attention as the ones you issue; we covered this in the hard part of e-invoicing.
- WhatsApp enquiries into bookings. Clinics, service centres and property agents receive the same questions all day: prices, availability, location. An AI agent can answer the routine questions, collect the details and place a booking in your calendar or system, then hand anything unusual to a staff member.
Neither is glamorous. Both are the kind of work that, once done properly, frees real hours and produces numbers you can track.
A one-page plan to scale your first process
Before you choose software or a vendor, fill in one page with the following. If you cannot complete it, the project is not ready yet.
1. The process
Name it precisely, with a start and an end. "Accounts" is too broad. "From a supplier invoice arriving to the entry being posted in AutoCount" is right.
2. The owner
One named person, not a department. They decide what counts as success, they handle exceptions or assign someone to, and they report results to you each month.
3. The baseline
Measure the current state for two to four weeks before you change anything. Useful measures include hours spent per week, number of items handled, error rate, and time from arrival to completion. Rough counts are fine; no counts are not.
4. The success measure
Choose one or two measures and a target you would be satisfied with. For example (an illustration, not a client figure), "invoice entry time per week falls by half, with no rise in posting errors". Write it down before the project starts, so the result cannot be redefined afterwards.
5. The escalation path
Decide what the system should do when it is unsure, who receives that case, and how quickly they must act. Decide also how staff report an error and who fixes it. This is the part most firms skip, and it is the part that keeps people trusting the system.
6. The data rules
Note what personal or customer data the process touches, where it is stored, and which vendor can see it. Malaysia is moving towards clearer guidance here; our note on the draft AI and personal data framework lists five checks worth running.
Then, and only then, expand
Once the first process has run for two or three months against its baseline, you have something most firms in the study do not: evidence. Use it to choose the next process. Often it is a neighbour of the first. Invoice capture leads naturally to purchase order matching, and enquiry handling leads to follow-up reminders.
Each new process gets its own one-page plan. Over a year, that sequence of small, measured changes becomes the scaling plan that only a minority of firms have today. It also gives you a clear record if you later apply for a grant or report to a board.
A note on vendors
The study found most Malaysian firms get AI through vendors, and for most SMEs that is sensible. Building in-house is rarely the best use of a small team. But buying from a vendor does not transfer responsibility for the outcome. Ask any vendor, including us, to help you complete the one-page plan before quoting. A vendor who will not discuss baselines, owners and escalation is selling a tool, not a change to your business.

Where to start this week
Pick one process that annoys your team, name an owner, and start counting. That costs nothing and tells you more than any product demonstration.
Codech Solutions builds AI agents, document automation and business insights for SMEs in Malaysia and Hong Kong. If you would like a second view on which process to start with, our AI audit walks through your current workflows and data with you and suggests a first candidate, with the baseline measures to collect. You can also contact us directly.
Frequently asked questions
How many Malaysian businesses use AI now?
The AWS-commissioned study Unlocking Malaysia's AI Potential 2026 reported that 38% of Malaysian businesses use AI, up from 27% a year earlier, which is more than 3.4 million firms. Most of that use is basic.
Why do so few firms scale AI beyond basic use?
Usually because AI is introduced as a tool rather than as a change to a specific process, the starting point is never measured, and nobody is named to own the results and the exceptions. The study found only 19% of adopters have a formal plan to scale.
What is a good first AI process for a Malaysian SME?
Something repetitive, rule-based and already resented by staff. Common examples are capturing supplier invoices into SQL Account or AutoCount, and turning routine WhatsApp enquiries into bookings.
Do we need to build AI in-house?
Rarely. Most SMEs are better served by a vendor, but the business should still own the process, the baseline, the success measure and the escalation path.
How long should we run the first process before expanding?
Two to three months against a measured baseline is usually enough to see whether it works and to choose the next process with evidence rather than opinion.
