In August 2025 Ryt Bank opened in Malaysia, a digital bank backed by YTL Group in partnership with Sea Limited. Its assistant, Ryt AI, does more than answer questions about banking. Coverage of the launch described customers sending money and paying bills through a text chat, with DuitNow and JomPAY supported, and paying a bill by uploading a photo of it.
That is the difference between a chatbot and an agent. A chatbot tells you how to pay the bill. An agent pays it.
Fifteen months on, agents are appearing in Malaysian insurance, energy and government services. Here is what they are, which jobs suit them, how they fail and what has to be in place before one acts for your business.
Answering versus doing
| Chatbot | Agent | |
|---|---|---|
| What it produces | An answer | A finished task |
| How it works | One question, one reply | Plans steps, uses tools, checks the result |
| What it can touch | Nothing. It only writes text | Whatever you connect it to |
| What a mistake costs | A wrong answer | A wrong action that has already happened |
A tool is a defined action the software allows the model to call: look up a customer, create a booking, raise a quotation. The model decides which tool to use. Your code decides which tools exist.
Bank Negara Malaysia put it this way in its November 2026 feedback statement on AI in the financial sector: “Agentic AI marks a shift from AI that primarily supports decisions to systems capable of undertaking increasingly autonomous, multi-step actions.”
What Malaysian organisations are doing
Banking. Ryt AI, above. Launch coverage said every action it carries out is verified and encrypted, behind a biometric login.
Insurance. In February 2026 Zurich Malaysia described Z Buddy, an assistant inside WhatsApp. For a motor renewal it checks eligibility, generates a password-protected quotation, sends a payment link and releases the cover note in the same chat. Queries that need nuance go to a live support team. More in our article on insurance claims.
Government. The Digital Ministry said an AI agent in the MyGOV Malaysia app would begin a beta rollout in early August 2026. Where a service supports it, the agent can carry a task out, “subject to user consent and established permissions and controls”. The ministry’s example was a police summons, guided “from the summons payment menu through to completion”. The app had more than 2.9 million users as of July 2026. More in our piece on the public sector.
Energy. On 4 September 2026 Petronas signed with Iraya Energies to add AI agents to myPROdata, the platform investors use to assess Malaysian upstream opportunities. Its stated aim is to cut the time from discovery to first hydrocarbons from 100 months to 50. See what this means for energy contractors.
Look at what the four have in common. Each agent has one narrow job and a fixed set of actions. Each stops for consent, or passes to a person, at a defined point. None is a general assistant let loose on a company’s systems.
Keen, and wary
Organisations are moving. In Workday’s 2025 research, 83% of Malaysian organisations said they were expanding their use of agents. Only 9% of Malaysian respondents were comfortable with agents operating in the background without human knowledge.
Customers are cautious about money. In the Adyen Index 2026, a YouGov survey of 1,031 Malaysian consumers, 52% were uncomfortable allowing an AI to complete a purchase for them.
The people in between are under pressure. In a 2026 TrendAI study across 23 countries, 75% of Malaysian IT decision makers said they felt pressure to approve risky AI implementations, against 66% globally.
Controls exist for exactly that situation.
Jobs that suit an agent
Apply three tests.
- Several steps. If one answer finishes the job, a chatbot is enough.
- Bounded by rules. You can write down what is allowed.
- Checkable. A person, or ordinary code, can tell whether it was done correctly.
Booking and moving appointments passes all three. So does chasing an unpaid invoice, or matching a payment slip to an invoice.
Poor fits are just as clear: decisions about people, such as hiring or rejecting a claim, work where “correct” is a matter of judgement, and work with no record to check against.
How agents fail
At speed. A person who makes a mistake makes it once. An agent with a wrong instruction repeats it for every customer until somebody looks.
By reading the wrong thing. An agent cannot always tell your instructions from text it happens to read. A message, an email or a PDF can contain a sentence written to redirect it. In March 2026 an attacker simply asked an agent run by the research organisation METR to reveal its API key, and it did. We covered that case in our article on AI keys. Anything an agent can read can try to steer it. Anything it holds can be asked for.
By drifting. Bank Negara’s statement tells institutions to monitor agents for “unintended outcomes, privilege escalation and autonomy drift”. As prompts, tools and data change, an agent can end up doing more than it was set up to do.
By looping. An agent stuck on a step keeps trying. Every attempt is a paid call to a model.
Six controls
Bank Negara asks for “clear authority boundaries, robust execution controls and effective human intervention”. It was written for banks. It is a sound brief for anyone.
1. Its own identity, with the least access that does the job. Not a staff member’s login and not an administrator’s key.
2. A short list of tools. AllSet, the WhatsApp platform we are building with Northridge Technologies, is in development and runs on demonstration businesses for now. Its agents have a defined set of skills, each switched on per business.
3. Limits written in code. A cap on the amount, on messages per customer per day, on steps per task. A sentence in the prompt asking the agent to be careful is not a limit.
4. Approval before anything irreversible. A payment, a refund, a deletion, a sensitive message to a customer. Where AllSet relays a message between two parties, the AI drafts and nothing is sent without a person pressing Send. Making that approval quick enough to be real is a design problem in itself.
5. A log of every step. What the agent read, which tool it called, with what values, and what came back.
6. A named owner and an off switch. One person reads the log every week and can pause the agent. In AllSet, when staff take over a chat the AI pauses until it is told to continue.
What it costs to run
A chatbot reply is usually one call to a model. An agent’s task is many, because each step sends the work so far back to the model. So a task costs more than a reply, and the cost varies with the number of steps.
Meter by task, not by month. In AllSet every model call goes through one gateway that records what it cost. Cap the steps, so a task that has not finished stops and goes to a person. Then compare the cost with the hours it replaces, using our costing guide.
A sensible first agent
Choose something internal, reversible and easy to check. An agent that prepares, and does not send, is a good start.
For example: each morning it reads the invoices that fell overdue yesterday, drafts a reminder for each, and puts the drafts in a queue. The accounts clerk approves, edits or discards them.
Run it that way for a month and count how many drafts were sent unchanged. Then let it send only the kind it always got right. If the first agent will face customers on WhatsApp, read what an agent may do there and what Meta charges first.
What to do
- List the tasks in your business that pass the three tests. Pick one.
- Write down what the agent may touch, what it must never do, and where it needs approval.
- Ask the vendor to show you the full log of one completed task.
- Put the proposal through the seven questions to ask before approving any AI project.
One more thing. Other people’s agents are already arriving at your website. Start counting them.
Sources
- Ryt Bank launches as Malaysia’s first AI-powered bank (RinggitPlus)
- Malaysia launches Ryt Bank, its first AI-powered bank (AI News)
- Artificial Intelligence in the Malaysian Financial Sector: Feedback Statement, November 2026 (Bank Negara Malaysia)
- Zurich Malaysia press release, 11 February 2026
- MyGOV to get agentic AI assistant (Lowyat.NET)
- PETRONAS deploys agentic AI to accelerate Malaysia upstream investment (World Oil)
- AI agents gain momentum across Malaysian organisations, according to new Workday research (Workday)
- Agentic commerce in Malaysia: the Adyen Index 2026 (Fintech News Malaysia)
- The AI governance gap in Malaysia (Tech Wire Asia)
- Security update, 31 August 2026 (METR)