AI in Malaysia · Internal AI assistants

17 million Malaysians use AI. None of the chatbots know your price list

Public chatbots know the internet and nothing about your customers, products or procedures. An assistant built into your internal system answers from your own indexed documents and records, and shows its source. What that is, what it is good for, where it goes wrong and what it costs.

Pikabyte Engineering 7 min read

68%

of Malaysia's working population, about 17 million people, use generative AI, according to the e-Conomy SEA 2026 report by Google, Temasek and Bain

About 17 million Malaysians use generative AI, equal to 68% of the working population, according to the e-Conomy SEA 2026 report by Google, Temasek and Bain. Your staff are almost certainly among them.

Businesses are behind their own employees. A 2026 study commissioned by Amazon Web Services found that 38% of Malaysian businesses use at least one AI tool, up from 27% a year earlier, and that 67% of those were still using AI for basic tasks such as public chatbots. Microsoft’s 2024 Work Trend Index found that 83% of Malaysian AI users were bringing their own AI tools to work.

A public chatbot can write a polite email and explain a tax term. It cannot tell your salesperson what you quoted a customer last March, which pump replaces a discontinued model, or what your own procedure says to do when a pressure test fails. It has never seen your price list. Staff who need those answers either paste company documents into a public tool, which is a risk of its own, or go and ask the one person who knows.

There is a well-understood way to close that gap. It has an awkward name.

Retrieval-augmented generation, in plain words

The technique is called retrieval-augmented generation, or RAG. It works in three steps.

1. Index. Your documents are read, cut into passages and filed in a way that lets software search them by meaning, not only by exact words. Catalogues, procedures, past quotations, supplier manuals, meeting minutes.

2. Retrieve. When someone asks a question, the system first finds the handful of passages most likely to contain the answer.

3. Answer. Only then does the AI model write a reply, using those passages, and it shows which document and page each part came from.

Three consequences matter to an owner.

  • The model is not retrained. Your documents sit in your own index. Change a procedure today and the answer changes today.
  • You can check every answer. The source is one click away.
  • It can say “I could not find that”. A well-built assistant is instructed to answer from the passages or not at all.

What staff actually ask

In a trading or manufacturing business the questions are ordinary, which is why they add up.

  • What is the warranty period on this brand, and who handles the claim?
  • Which of our fittings is compatible with this pipe size?
  • What did we agree with this customer about delivery charges?
  • What is the procedure for a goods return after thirty days?
  • Has anyone dealt with this fault before, and what fixed it?

Each of these is a phone call, a walk across the warehouse or a wait until the person is back from lunch.

In Microsoft’s 2023 Work Trend Index, a survey of 31,000 workers in 31 markets including Malaysia, 62% said they struggle with too much time spent searching for information in their workday. That is a software vendor’s survey, but few managers would dispute it.

Why it belongs inside your internal system

A stand-alone chatbot loaded with PDFs is a start. Building the assistant into the system your staff already use gives you four things a stand-alone tool does not.

Permissions. A storekeeper should not be able to ask what the sales manager earns. Inside your system, the assistant can only retrieve what the person asking is allowed to open.

Live records as well as documents. A document says what the return policy is. The system knows whether this customer’s goods were delivered on the 3rd. An assistant inside the system can look both up, and should take any number, such as a price or a stock level, from the records and never from its own memory.

Actions with a person in charge. It can draft the quotation or open the job for someone to confirm.

A list of what it could not answer. Every unanswered question is a gap in your written knowledge. Reviewing that list each month is how the company’s documentation improves.

We build this way. In the question bank platform we built for an education company, 475 textbooks and past papers were read into a per-chapter knowledge base, generation is grounded in that material, and editors approve everything before it is published. AllSet, a platform we are developing with a partner in Singapore, answers from a business’s own indexed sources and keeps a list of knowledge gaps when it cannot find an answer.

What the evidence says

The best measured study so far followed 5,172 customer support agents at a United States company and was published in the Quarterly Journal of Economics in 2025. Access to an AI assistant raised the number of issues resolved per hour by 15% on average. Less experienced and lower-skilled workers improved both the speed and the quality of their work.

That pattern is the case for a traditional business. The assistant does not make your twenty-year veteran much faster. It lets the person hired last month answer like someone who has been there for years. We look at that side in when the old hand retires.

In Malaysia, Hong Leong Investment Bank told The Edge in July 2026 that its internal AI draws mainly on the bank’s own policies, standard operating procedures and past meeting records, with people reviewing and approving its output.

Where it goes wrong

Grounding an AI model in documents reduces mistakes. It does not remove them.

A study from Stanford, published in the Journal of Empirical Legal Studies in 2025, tested AI legal research tools from LexisNexis and Thomson Reuters. Both are built on retrieval. Each still gave false or misleading answers more than 17% of the time, though less often than a general-purpose chatbot.

So design for error.

  • Show the source, always. An answer without a citation is treated as a guess.
  • Keep figures out of the model’s hands. Prices, balances and stock come from the database.
  • Let it decline. “I could not find this in our documents” is a correct answer.
  • Keep a person on anything that leaves the building. The assistant drafts. Staff send.
  • Test it with real questions before staff rely on it. We describe how in preparing your documents for AI.

What it costs

The indexing itself is cheap. OpenAI’s published price for its small embedding model, read in November 2026, is US$0.02 per million tokens. Five thousand documents of a few pages each come to roughly ten million tokens, or about 20 US cents. Storing the index can be done in the same open-source database that holds the rest of your system.

The real cost is the software work around it: getting clean text out of scans and spreadsheets, enforcing permissions, connecting to your records, building the screen staff use, and testing. After that comes a monthly running cost that rises with the number of questions asked, which should be metered so nobody is surprised. Ask any vendor for that figure at your expected volume.

Two more questions belong in the first meeting. Where is our data processed, and is it used to train anyone’s model? We set out how to check in where your AI runs.

A first project

  1. Choose one team that answers the same questions every day. Sales support and after-sales service are common choices.
  2. Collect the two hundred documents that team actually uses.
  3. Write down fifty real questions with the correct answers.
  4. Build the assistant for that team only, with sources shown.
  5. Score it against the fifty questions, fix the documents it got wrong, and repeat.
  6. Widen it once the team prefers asking the assistant to asking a colleague.

An assistant needs something to read. If your orders and procedures are not in a system yet, start with what an internal tool replaces first.

Your staff have already decided to use AI. What is left for you to decide is whether it works from your company’s knowledge or from guesswork.

Sources

Written by

Pikabyte Engineering

Pikabyte builds custom business systems, applied AI and secure cloud infrastructure from Malaysia. Talk to us.

The work behind this article

Platform · Education · Assessment content

Textbooks in, reviewed exam questions and learning cards out, in English and Bahasa Melayu.

A content production platform for Tavis, an online tuition provider for SPM students in Malaysia. Subject teams upload textbooks and past papers, AI extracts and generates questions and learning cards against the syllabus, an AI review stage votes on each one, and editors approve what gets published to the live learning platform.

9,300+
questions generated and tracked
5,600+
learning cards, in English and Bahasa Melayu
475
textbooks and papers read into the knowledge base
1,150+
automated backend tests
Tour Question Bank AI Tavis · Malaysia

Start a conversation

Want this applied to your business?

Tell us what is slowing the business down. You will talk to an engineer, not a salesperson, and get a straight answer on whether software is the right fix.