"We should use AI" has reached every board agenda in Malaysia. What it rarely comes with is a concrete answer to the only question that matters: use it on what?
We build AI systems for a living, including one of Malaysia's first AI-based loan approval platforms and a conversational avatar that talks to customers on WhatsApp. So this is not an article about whether AI is the future. It's about which AI projects pay for themselves in an SME, and which ones quietly burn a budget.
The pattern behind every AI project that works
Every AI project we've seen succeed removes a queue. Somewhere in your business, requests pile up and wait for a person: loan applications waiting for review, customer messages waiting for a reply, documents waiting to be read and filed. The queue exists because each item needs a small amount of judgment, and judgment used to require a human.
That's the test. If you can point at a queue and describe the judgment each item needs, AI can probably shorten it. If your AI plan doesn't point at a queue, it's a science project.
Four uses that earn their cost
1. Decisions that follow rules you can mostly write down
Loan eligibility is judgment, but most of it is describable: income patterns, documents, history, risk signals. We built an AI-powered financing portal where applications that once waited days for manual review are assessed in minutes, with documents verified automatically. Processing time fell 70% and approval rates rose 35% across more than 500 SMEs, because the humans now spend their time on the genuinely borderline cases.
The same shape appears everywhere: quotation approvals, claim triage, supplier onboarding checks, invoice matching. If your team applies the same checklist forty times a day, that checklist is software waiting to happen.
2. Customer conversations, on the channels Malaysians actually use
Malaysian customers live on WhatsApp. An AI system that answers there at 11pm, in the customer's language, from your business's actual knowledge, is no longer speculative. We built a digital avatar platform that listens, thinks, and speaks back in real time, powered by retrieval-augmented generation, which means it answers from a specific knowledge base instead of improvising like a generic chatbot.
The honest caveat: an AI assistant is only as good as the knowledge you give it and the escape hatch you design. It must know what it doesn't know and hand off to a human cleanly. The projects that embarrass companies skipped that part.
3. Documents your staff read so nothing else can happen
Invoices, IDs, agreements, application forms. If a person opens each one, extracts the same six fields, and types them into a system, that entire loop can be automated with current AI, with a human checking only the exceptions. This is the least glamorous AI use and often the fastest payback in an SME, because the queue is visible and the hours are countable.
4. Speech, for businesses that run on calls
Speech-to-text has crossed the threshold where transcribing and summarizing calls is reliable infrastructure. We've built the platform engineering behind an enterprise speech AI system: streaming speech recognition and LLM summarization running as production services. For an SME, the practical version is smaller: calls transcribed, summarized, and logged against the customer record, so nothing said on the phone gets lost.
What this costs to run, honestly
Two decisions drive AI running costs.
First, commercial APIs versus self-hosted models. Using OpenAI or Anthropic is fast to build on, and you pay per use. Costs scale with volume and your data goes to the provider under their terms. Self-hosting open-source models (Llama, Mistral) keeps sensitive data entirely in your infrastructure and caps costs at scale, in exchange for a heavier setup. For regulated industries or anything touching customer financial data, we lean self-hosted. For most SMEs starting out, an API is the right first step.
Second, scope. A focused pilot on one queue is weeks of work, not months, and its running cost is typically hundreds of ringgit monthly, not thousands. The expensive AI projects are the vague ones.
The projects to avoid
Skip anything described as "AI transformation" without a queue attached. Skip a chatbot bolted onto a website nobody visits. Skip training a custom model when an existing one plus your documents does the job, which is most of the time. And treat any vendor who won't explain where your data goes as a hard no. Under PDPA, that answer is your responsibility, not theirs.
Where to start
Pick the queue that hurts most. Measure it for a week: how many items, how long each waits, what that delay costs in staff hours or lost customers. That one measurement turns "we should use AI" into a project with a payback period, and it's the exact brief a good development partner needs.
Our AI development work and the SME financing case study show what production systems look like. If you have a queue in mind and want a view on whether AI fits it, tell us about it. Scoping that answer is a short conversation, not a contract.