Turning Singapore's Agentic AI Guidance Into a Checklist

Who this is for: Executives, site managers and IT leads at Japanese companies with operations in the Philippines

Reading time: 15 minutes

What this case study covers: How to use voluntary guidance published by a neighbouring country as your own checklist. This is about turning a document with no legal force into a practical tool.

Part 1: Read → Consider the Implications for Your Company

Here is what the addendum consists of.

Item	Detail
Published	17 June 2026, by Singapore's Cyber Security Agency (CSA)
Scope	Agentic AI, meaning AI that decides its own steps and executes multiple actions
How it was made	Following a consultation from 22 October to 31 December 2025, with industry, government and international participants
Status	The measures are voluntary. Explicitly not legally binding, and not applicable to every organisation or environment

What Happened

On 17 June 2026, the Cyber Security Agency of Singapore (CSA) published an addendum to its existing guidelines and companion guide on securing AI systems. It covers agentic AI — the kind that decides its own steps and carries out several operations in sequence.

The addendum was finalised after a public consultation that ran from 22 October 2025 to 31 December 2025, and was developed with industry, government and international partners.

One point matters above the rest: the measures in the addendum are voluntary. They carry no legal obligation, and the document states plainly that they may not apply to every organisation or environment.

Earlier in the same country, in January 2026, the Infocomm Media Development Authority (IMDA) published a governance framework for agentic AI, with real-world cases and further practices added in May 2026.

Why This Matters

"What does voluntary guidance from Singapore have to do with a Japanese company in the Philippines?" is the reasonable first question.

It matters for these reasons:

It is the first time a concrete practical standard has been written down in Southeast Asia
Because it is voluntary, you can select what fits your situation
It comes from a public agency, which makes it usable as a basis for internal discussion
Customers and partners in the region may eventually expect the same standard

In short, it is worth reading not as something to comply with, but as a draft of your own checklist.

The Implication for Your Company

Writing a policy for AI agents from scratch takes time and leaves gaps. Starting from a list compiled by a public agency gives you, at minimum, an inventory of what needs to be considered.

The Philippines does not yet have comprehensive rules in this area. That is precisely why reading the neighbouring country's document early is worth the hour.

Part 2: Key Terms for Executives

Agentic AI — AI that does not simply return an answer to a prompt, but decides its own steps, uses several tools, and chooses its next action based on results. The wider the scope you can delegate, the more room there is for unexpected behaviour.

Least privilege — Granting only the permissions actually needed. Obvious when managing people, easily forgotten with AI agents, which tend to be given "permission to do everything, for now."

Environment segmentation — Keeping test and production separate, and keeping departments separate. It stops one problem from spreading across everything.

Input and output validation — Checking both what goes into the AI and what comes out of it. One side alone is not enough.

Human-in-the-loop — A design where certain actions require a person's approval before proceeding. Deciding which actions belong here is the centre of the design.

Continuous logging and monitoring — Keeping the system in a state where you can reconstruct what happened. You cannot create records after the incident.

Part 3: Applying This to Your Company

Turning the addendum into something checkable at a Philippine site.

Step	What to do	Easy to miss
1	List the permissions	Permissions where nobody can say who granted them and when tend to stay in place
2	Decide which actions need approval	Limit it to irreversible actions and actions that leave the company
3	Decide what goes in the log	Which model executed it is the field most often omitted
4	Separate the environments	Do not give a test agent write access to production data
5	Review periodically	Behaviour shifts when the service provider changes something. Quarterly

The addendum covers supply chain security, access controls, environment segmentation, input and output validation, model and system hardening, human oversight, and continuous logging and monitoring.

Here is how to turn that into something you can actually check at a Philippine site.

Step 1: Build a permissions inventory

For each AI agent in use at the site, write down three things:

Which systems it connects to
What it can do there (read only, write, or send externally)
Who granted that permission, and when

The third item is the one people cannot answer, and unanswered permissions are usually the ones still sitting there.

Step 2: Decide which actions require approval

Requiring approval for everything stops the work. Narrow it with these criteria.

Irreversible actions — sending, paying, deleting, publishing. These go on the approval list.

Actions that leave the building — anything reaching a customer, reaching a partner, or becoming public. Also on the list.

Internal actions you can redo — drafting, internal summaries. Leave these free.

Put the line in writing and you stop spending time on the same judgement call repeatedly.

Step 3: Decide what goes into the logs

Having logs is not enough. What is in them is the question.

At minimum, record when, which operation, against which input, using which model, and what was produced. Which model handled it is the field most often missing, and the one you need most when you later investigate why results changed.

Step 4: Separate the environments

Keep the environment for experimenting apart from the one running the business. Experiment in the same place and a test action reaches production data.

Small sites often say they lack the capacity for this. Even so, not granting a test agent write access to production data is something you can do today.

Step 5: Review on a schedule

The addendum itself asks organisations to periodically re-evaluate whether current controls remain adequate.

AI agents change behaviour when the provider changes something. Put a quarterly slot in the calendar to review the permissions inventory and the approval line.

Across my development career I have made detailed documentation of designs and specifications a requirement, along with review by more than one person. A permission model that lives in one engineer's head cannot be audited, and an addendum you cannot audit is not worth signing. Permission management for AI agents deserves the same treatment.

Part 4: Common Failure Patterns (What Not to Do)

NG1: Dismissing it because it is voluntary

Not reading it because it carries no legal obligation is a wasted opportunity. Being voluntary is exactly what lets you select what fits.

NG2: Adopting the guidance verbatim as internal policy

The opposite failure. Pasting another country's public document into your rulebook produces a list of items that do not match your size, and a policy nobody follows. Select the items and rewrite them in your own words.

NG3: Granting broad permissions at the start

The pattern where a team says "let's just get it running," grants full permissions, and then goes into production that way. Narrowing them afterwards is difficult once the work depends on them.

NG4: Collecting logs that nobody reads

Logs are not the objective. Assign a person and a time — monthly is enough — to look for anything unusual.

NG5: Not telling the local staff

The people using the AI agents daily are your local staff. If nobody explains which actions now require approval and why, workarounds appear.

In the Philippines a spoken agreement is often treated as settled, which is a problem when the thing agreed is who may grant an agent a new permission. Put the permission and the approver in writing. Permission changes deserve the same handling — an explicit item on the agenda.

Three Tips for Getting This Done

Tip 1: Start with one agent — Trying to review everything at the site at once means finishing nothing. Start with the one that carries the most impact.

Tip 2: Write "not yet" rather than "cannot" — For items you are not addressing, record the reason and when you will revisit. A blank leaves the next reader with nothing to judge by.

Tip 3: Cite the source in internal discussions — Basing the argument on a public agency's document makes internal agreement easier to reach. The same holds for explaining it to Japanese headquarters.

Bonus: How to Make Use of PH AI Works

We can work through a permissions inventory for the AI agents at your Philippine site, or the line between actions that need approval and actions that do not, based on your actual operations. Reviewing tools you have already deployed is a fine place to start.