Over the past year, almost every client conversation we've had has included the word "agent". Boards want one, vendors promise one, and demos look impressive. Yet very few Malaysian enterprises have an AI agent doing real work in production.
The gap is rarely the model. Today's large language models are good enough for most business tasks. The gap is in choosing the right first job for the agent, connecting it safely to your systems and proving that it pays back. Here's the approach we use.
What we mean by an "agent"
A chatbot answers questions. An agent also takes actions: it looks up a policy, checks a customer's account, drafts a claim decision, books an appointment or raises a ticket. It does this by calling tools, which are APIs into your systems, within limits you define.
That difference matters, because once an agent can act, you need to think about permissions, audit trails and what happens when it gets something wrong.
Step 1: Pick a job, not a technology
The best first use cases share four traits:
- High volume, low drama. Hundreds or thousands of similar requests a month, where an occasional mistake can be caught and fixed.
- Knowledge spread across documents. Policies, product sheets, SOPs and past tickets that staff currently search by hand.
- A clear hand-off point. The agent can prepare the work and a person approves it.
- A number you already track. Handling time, first-response time, backlog or cost per case.
In practice, good first agents include internal policy and HR assistants, customer service triage, insurance claim intake, sales-support assistants that draft quotations, and IT helpdesk agents that reset access and log tickets.
A useful test: if you can't name the person whose job gets easier on day one, the use case is too vague.
Step 2: Scope it to two weeks
Resist the urge to design the "enterprise AI platform" first. Scope a pilot that can show results in about two weeks:
- One process, one team and one channel (for example, web chat for the claims team).
- A bounded set of documents and two or three read-only tools.
- A test set of 100 to 200 real past cases with known good answers.
- A success metric agreed in advance, such as "drafts accepted without edits" or "minutes saved per case".
Read-only tools come first. Write actions, such as updating records or sending messages, come after the agent has earned trust.
Step 3: Ground it in your data
Agents are only as good as what they can see. Most of the work in a pilot is not prompt-writing. It's preparing knowledge: cleaning documents, removing outdated versions, splitting them sensibly and tagging them so the agent retrieves the right section. This approach is usually called retrieval-augmented generation (RAG).
Malaysian enterprises have an added requirement: language. Customers write in English, Bahasa Malaysia, Chinese and often a mix of all three in one sentence. Test with real messages, not translated samples.
Step 4: Build the guardrails in from day one
- Permissions: the agent inherits the access level of the user it serves, never more.
- Citations: every answer links to the source document so staff can verify it.
- Confidence and escalation: when the agent isn't sure, it hands over to a person instead of guessing.
- Audit logs: every question, retrieval and action is recorded.
- Personal data: mask or minimise personal data in prompts and logs, and know where your model provider processes data. Our PDPA and AI checklist covers this in detail.
Step 5: Measure, then expand
After the pilot, compare the agent against your baseline on the agreed metric. Also track the things that show up later, such as how often staff override it and which questions it escalates. These show you where to improve the knowledge base and which tools to add next.
When the numbers hold, expand one dimension at a time: more users, then more channels, then write actions. Teams that try to expand all three at once are the ones that end up back at the demo stage.
The part that isn't AI
An agent that works in a sandbox still has to run inside your real environment: your mobile app, your CRM, your core system and your single sign-on. In our experience about half the effort of a production agent is ordinary, solid software engineering. That's why we pair AI specialists with the same engineers who build and support our clients' apps.
Want to find your first agent use case?
We run a short discovery workshop that ends with a ranked shortlist of use cases and a scoped two-week pilot. Book a session with our AI team →
