Guardrails for marketing agents: approvals, budgets, PDPA
Written rules for marketing AI agents in Malaysia: which actions need approval, how to cap spend and messages, what PDPA means for customer data, and a checklist.
A marketing agent needs written rules before it touches anything real: what it may do alone, what needs a person's approval, how much it can spend or send, and what it logs. If it uses customer data, Malaysia's PDPA applies to you as the business. This is general information, not legal advice, so have a lawyer check your own setup.
Part of the guide: Agentic marketing

On this page
- Why do agents need written rules?
- Which actions need human approval?
- How do you cap money and message volume?
- What does PDPA mean for an agent using customer data?
- What do you log, and for how long?
- What happens when something goes wrong?
- A guardrail checklist
- What should you do next?
- Frequently asked questions
Give a marketing agent access to your ads, messages or customer list and you have given it power. The guardrails decide how much. This guide covers the rules we write before any agent goes live: approvals, limits, PDPA and logs.
Prepared 8 November 2026. General information only, not legal advice. Part of our guide to agentic marketing.
Why do agents need written rules?
An agent follows instructions literally and acts fast. A person who spots something odd stops and asks. An agent that is not told to stop will continue, whether that means sending the same message twice or raising a budget on bad data. Written rules turn "use common sense" into something the system can check.
Good rules are short, specific and enforced in code rather than only in the prompt. "Never spend more than RM 200 a day" in a prompt is a request. The same limit set in the ad platform's account budget is a wall. We use both, and trust the wall. If you are new to agents, start with what agentic marketing is.
Malaysia also has a voluntary national framework. The National Guidelines on AI Governance and Ethics, launched by MOSTI in 2024, are not binding law, but they are a sensible reference when you write your own rules.

Which actions need human approval?
Sort every action an agent could take by what it costs if wrong and whether it can be undone. Reading data is low risk. Public posts, customer messages and spend are not.
Example approval matrix for a sample retailer (illustrative, to be adapted):
| Action | Rule |
|---|---|
| Read analytics, ads and sales data | Agent alone |
| Draft a social post or reply | Agent drafts, person approves |
| Reply to a customer enquiry from an approved script | Agent alone, with a handover to a person for anything else |
| Pause an ad that exceeds its cap | Agent alone, tells the owner |
| Raise a daily budget | Person approves |
| Send a promotion to the customer list | Person approves each send |
| Change prices, delete anything, issue refunds | Never by the agent |
The pattern: the agent may take actions that reduce risk, such as pausing. A person approves actions that spend money or reach many people. See our guides to ads agents and WhatsApp follow-up agents for specific limits on each.
How do you cap money and message volume?
Set caps in the system the agent works through, not only in its instructions.
- Spend: use the daily and lifetime budgets inside Meta and Google accounts, so the platform stops spend even if the agent fails.
- Messages: limit the number of outbound messages per hour and per day, and per customer, in code.
- Changes: limit how far the agent may change a setting in one step, for example a budget by no more than a fixed percentage, and how often.
- Stop switch: one action, available to the owner at any time, that pauses the agent and its permissions.
- Fail closed: if the agent cannot check a limit, it does nothing and alerts a person.
The caps should be small at the start and widened only after weeks of clean logs.

What does PDPA mean for an agent using customer data?
An agent that reads names, phone numbers or purchase history is processing personal data, and Malaysia's Personal Data Protection Act 2010 applies to the business that collects it. The official source is the Personal Data Protection Department at pdp.gov.my. This is not legal advice. In plain terms, the PDPA sets out seven principles: general, notice and choice, disclosure, security, retention, data integrity and access. For an agent, that translates into practical questions:
- Notice and choice: did customers know their data would be used for this purpose, and in which language?
- Disclosure: the data goes to the agent's AI model provider and other tools. Is that covered in your notice, and have you checked where the data is processed?
- Security: who has access to the agent's data, and is it limited to what the agent needs?
- Retention: how long do logs and conversations stay, and when are they deleted?
- Access and accuracy: can you find and correct a customer's data if they ask?
Direct marketing. The PDPA gives individuals the right to tell a business in writing to stop using their personal data for direct marketing (section 43). An agent that sends promotions needs a working opt-out that is honoured quickly and applied everywhere, including in every tool the agent uses.
2024 amendments. Malaysia passed the Personal Data Protection (Amendment) Act 2024. Among other changes, it introduces breach notification duties, direct obligations for data processors and a data protection officer requirement. Different parts take effect in stages, and we do not quote dates here, because commencement dates and guidelines should be taken from the official source. Check pdp.gov.my or ask a lawyer for what applies to you today.
A practical default: send the agent only the data it needs. A reporting agent needs totals, not customer names. A follow-up agent needs a phone number and the enquiry, not the full history.
What do you log, and for how long?
A log is how you answer "what did the agent do?" after a complaint or an error. Keep:
- The time, the action and the data it was based on.
- Who approved it, if approval was needed.
- The message or change as it was sent.
- Errors, refusals and handovers to people.
Review the log weekly at first. Decide a retention period in line with your PDPA retention duty. Logs with personal data should not be kept longer than needed, and old logs should be deleted on a schedule. A log that is never read does not protect anyone, so give one person the job of reading it.
What happens when something goes wrong?
Plan for the mistake before it happens. Write down who is told first, how the agent is paused and how a bad message or change is undone. For example, if an agent sends a promotion to the wrong segment, the owner needs to know within minutes, not at the next weekly review. Alerts should go to a person on WhatsApp or phone, not only to a dashboard.
After the incident, read the log, find which rule was missing, and add it. Guardrails improve this way. If personal data was exposed, treat it as a data protection matter and check your obligations under the PDPA with your adviser.
A guardrail checklist
Before any agent goes live, you should be able to answer yes to each line:
- Every action is on the approval list: alone, with approval, or never.
- Spend and message caps are set in the platform or code, not only in the prompt.
- There is a stop switch and a named owner.
- Customer data is limited to what the task needs.
- Your privacy notice covers the use, and opt-outs are honoured in every tool.
- Logs are kept, read and deleted on schedule.
- A person reviews the first weeks of output.
- Your team knows the rules. Training can be HRD Corp claimable, with eligibility decided by HRD Corp: see HRD Corp courses.
If any answer is no, fix that before launch.
What should you do next?
Start with an AI audit. Quick Scan from MYR 3,500, credited to the build. We map your processes and data, set the rules above, then scope the build. Custom builds start from MYR 12,000, scoped from the audit, and a care plan is from MYR 2,500 per month. A half-day team course for up to 20 people starts from MYR 4,500.
Frequently asked questions
Is this legal advice?
No. It is general information to help you ask the right questions. Check your own situation with a Malaysian lawyer or your company's data protection adviser.
Do I need a DPO?
The 2024 amendments to the PDPA introduce a data protection officer requirement. Whether it applies to you depends on the rules in force and your data volume, so check the current requirements on pdp.gov.my or with a lawyer.
Can the agent use my customer list for ads?
Only if that use fits what customers were told and agreed to when you collected their data, and you honour opt-outs. If unsure, do not upload the list until a lawyer has checked the purpose.
What if the agent makes a mistake?
You, as the business, remain responsible. That is why risky actions need approval, limits are set in code, and every action is logged so you can see what happened and undo it.