By Ahmad Firdaus · October 2026
Someone in your business has to own what the agent does — and it should be a person with a name.
The question nobody wants to answer
Ask a Malaysian SME owner who is responsible if the new AI agent gets something wrong and you usually get a shrug. That is fair, because the law is unsettled — but the risk is not evenly spread. If an agent acting for your business makes a statement, the person who relied on it is dealing with you. Recovery from anyone upstream depends on contracts you have probably not read closely.
What the FTC chairman said
Reuters reported on 25 September 2026 that US Federal Trade Commission chairman Andrew Ferguson, speaking at a Reuters event in Austin, pushed back on describing AI agents as autonomous actors that "break loose" with "wills and desires of their own", and suggested the developers who instruct agents would be the ones liable for harm. "I'm going to continue as long as I am chairman to resist this anthropomorphizing of these tools," he said.
Read that as direction of travel, not a shield. It comes from a US regulator with no jurisdiction over a Malaysian SME, and it is enforcement philosophy rather than a decided case. What is useful is the framing: agents are tools, and tools have owners. If a regulator refuses to treat the agent as an independent actor, accountability lands on whoever deployed it.
The contract you actually signed with your vendor determines how much risk you can pass back.
Malaysia has no AI liability law — yet
According to the Legal 500 Country Comparative Guide on artificial intelligence (2026), Malaysia has no AI-specific liability regime; AI-related harm would likely be handled through existing frameworks — contract, tort, consumer protection, product liability and data protection. The same guide notes that deployers and operators are typically the primary regulated entities, and that under the PDPA liability rests mainly with the data controller, the organisation that decides how personal data is used.
Reform is coming. Malaysia's National AI Office released a consultation paper on a proposed AI Governance Bill on 10 July 2026, built on a risk-based tiering of AI systems and five principles. The third is accountability: responsibility should remain attributable to identifiable persons and must not be displaced onto the AI system itself. Whatever the final Bill says, that principle describes how a customer, a tribunal or an inspector will look at your agent.
Three ways it actually happens
| What the agent does | Who the customer looks to | Your exposure |
|---|---|---|
| Quotes a wrong or out-of-date price | Your business, which made the indication | Section 12 of the Consumer Protection Act 1999 makes a misleading indication as to price an offence — and it does not matter whether you gave it yourself, through an agent, or on someone else's behalf. |
| Over-orders stock or commits to a purchase | You, as the buyer on record | The supplier enforces the order. Your remedy is against the vendor only if the contract allows it. |
| Misstates a warranty or performance figure | Your business | Section 10 of the CPA 1999 prohibits false or misleading representations, with fines up to RM250,000 for a company on a first offence. "Misleading" includes anything capable of leading a consumer into error, so intent is not part of the test. |
A working desk with a laptop and a calculator — most liability disputes are settled by order records, quotes and approval trails. Photo: Wikimedia Commons (CC0).
Why the vendor contract rarely saves you
Vendor agreements are drafted by the vendor. In practice they cap liability at the fees you paid, disclaim guarantees about the accuracy of outputs, and place responsibility for how outputs are used on the customer. A RM300-a-month agent seat will not underwrite a RM40,000 stock error.
Read your agreement with three questions. Does it cap total liability, and at what figure? Does it require human review before certain actions? And does it give you the logs to prove what the agent did? That last question is where the September 2026 wave matters: Okta's Blueprint Alliance (22 September), RSA's Agent ID (29 September) and SAP's work with Nvidia (28 September) all centre on registering agents, tracing delegation and keeping auditable records. Audit trails are becoming a product feature — and they are your evidence.
What to put in place this month
| Control | What it means in practice |
|---|---|
| Written internal policy | Two pages: which agents are approved, what data they may touch, who signs off. |
| Named human owner per agent | One person per agent, by name, accountable for reviewing output. Not "the team". |
| Approval thresholds | Anything above a set ringgit value, any outbound customer commitment and any price change waits for a human. Start low; raise the limit only after a clean month. |
| Audit logs | A dated record of what the agent sent, to whom, and who approved it — kept at least as long as your dispute window. |
| Insurance and contract review | Ask your insurer whether a cyber or business-interruption policy responds to an agent error, and have the vendor agreement read once by someone qualified. |
| Customer communication rules | Decide who contacts a customer when the agent gave wrong information, and correct it in writing, quickly. |
A weekly stand-up where one named person reports what the agent did is governance most SMEs can sustain.
Framing this for a team with no legal department
You do not need a compliance function. You need three documents and one habit. The documents: an agent register listing each agent, its owner, its access and its limits; a two-page use policy; and a short escalation note naming the decisions that never go to an agent. The habit: a weekly review where the owner reads what the agent did and initials it.
Keep the language plain and the thresholds low enough that people follow them. A rule nobody enforces protects nobody — the point is being able to answer, within an hour, who authorised a wrong message to a customer.
Conclusion
The core of the liability question is simple: your business deals with the customer, so your business answers first. The FTC chairman's September 2026 remarks point to developers being liable for the agents they instruct, and Malaysia's proposed AI Governance Bill already says responsibility must not be displaced onto the AI system itself — but neither will rescue you in a dispute next week. Vendor terms cap the vendor's exposure, not yours.
So move the risk into places you control: a named owner, a ringgit threshold, a log, a written policy, and an insurer that has been asked the question directly.
Frequently Asked Questions
Can I hold the AI provider responsible if their agent gives wrong information?
You can try, and the FTC chairman's position suggests regulators are leaning that way, but your ability to recover depends on the terms you agreed. Most vendor contracts limit liability and place responsibility for how outputs are used on the customer.
Is an agent's message to a customer legally binding on my business?
Generally the customer will treat it as your statement, because it came from a channel your business controlled. Under the Consumer Protection Act 1999 a misleading representation or price indication can expose your company to fines whether a human or an agent produced it.
What is the cheapest meaningful protection?
An approval threshold and a log. A rule that no quote, order or customer commitment leaves the business without human approval stops most costly errors, and the log lets you prove what was sent, when and by whom.
Does our existing business insurance cover an agent's mistake?
Not automatically. Policies differ on whether automated processes count as covered acts, and many cyber policies address breaches rather than commercial errors. Ask your insurer and get the answer in writing.
Ahmad Firdaus covers policy, compliance and tax matters for SMEBuddies. His background spans regulatory reporting and compliance work for small and mid-sized Malaysian companies, and he now advises owner-managed firms. He writes about the rules that quietly affect a business's cash and exposure, and how to meet them without hiring a department. His articles focus on the practical steps an SME can take today rather than the legal detail that can wait for a lawyer.
Who Pays When an AI Agent Gets It Wrong?