Back to Resources
Compliance6 min read
AI Second Brain and PDPA: A Governance Checklist for SMEs
How Singapore SMEs can connect company files to AI without breaching the PDPA: what data to include, permission rules, vendor checks and a one-page policy.
Can a Singapore SME connect company files to AI under the PDPA?
Yes. The PDPA doesn't ban using AI on company data, but personal data inside your documents keeps all its protections when an AI reads it. You still need a proper purpose, reasonable security, limits on who can see what, and care with overseas processing. A company second brain is compliant when you decide those rules before connecting sources, not after.
This guide turns that into a practical checklist. It is general guidance, not legal advice. For anything high-stakes, check the PDPC's guidelines and speak to a lawyer. For background on what a second brain is, start with our pillar guide.
Which PDPA obligations matter most for an AI second brain
Five PDPA obligations do most of the work when you connect documents to AI: purpose, protection, access, retention and transfer.
• Purpose limitation. Personal data should be used for purposes the individual would reasonably expect and, where needed, consented to. Using client files to answer staff questions about that client is usually within purpose. Mining them for unrelated marketing may not be.
• Protection. You must make reasonable security arrangements. Broad folder sharing that lets any staff member retrieve payroll data through an AI search is a protection problem.
• Access and correction. Individuals can ask what data you hold about them. Know where your second brain's sources are so you can answer.
• Retention limitation. Don't keep personal data longer than needed. Old files you'd normally delete shouldn't be preserved just because the AI can read them.
• Transfer limitation. If your AI vendor processes data outside Singapore, you need to ensure comparable protection, usually through the vendor's contractual terms.
The PDPC's Advisory Guidelines on the Use of Personal Data in AI Recommendation and Decision Systems (March 2024) focus on AI that makes recommendations or decisions, but their emphasis on transparency and accountability is a useful guide here too.
Decide what goes in, what stays out, and what comes later
Sort your sources into three tiers before connecting anything. It's the single most effective governance step.
• In from day one: SOPs, process checklists, templates, product and service information, public marketing material, internal FAQs. Low personal-data content, high value.
• In after a permissions review: client project folders, proposals, meeting notes and shared inboxes. Useful, but they contain personal data, so access must be right first.
• Out until there's a clear reason and controls: HR files, payroll, medical certificates, disciplinary records, NRIC or passport scans, and anything under special confidentiality such as legal privilege or financial client data.
Write the tiers down. That list is the core of your AI policy.
Fix permissions: the AI sees what each user can see
Most business AI search tools are permission-aware, which means your existing sharing settings become your AI access rules. Anthropic, for example, states that Claude's enterprise search only shows results from data a user can access in the original system. Microsoft 365 Copilot and ChatGPT company knowledge work the same way.
That's reassuring only if access is correct. Before launch:
• Remove "anyone with the link" and organisation-wide sharing from sensitive files.
• Put HR, finance and restricted client work in folders limited to named people.
• Review membership of shared drives, Teams sites and chat channels, and remove former staff and contractors.
• On Microsoft 365, consider SharePoint's Restricted Content Discovery for sites that should stay out of AI results.
Check your AI vendor before you connect
Use a business or enterprise plan, never personal accounts, and check four things in the vendor's terms.
• No training on your data. Business tiers from Anthropic, OpenAI, Microsoft and Google generally exclude customer data from model training by default. Confirm this for your specific plan.
• Data location and transfer terms. Where data is processed and stored, and what contractual protections apply to overseas transfer.
• Admin controls. Can an admin choose which connectors are allowed, remove users and see usage?
• Retention. How long conversations and connected content are kept, and whether you can shorten that.
Record what you checked and when. If a client or the PDPC ever asks how you protect data in AI tools, this record is your answer. Our PDPA prompting checklist covers the day-to-day habits for staff.
Write a one-page AI second brain policy
A short written policy is enough for most SMEs, and it's what regulators and clients expect to see. Cover:
• Purpose: what the second brain is for, such as answering staff questions about processes and clients.
• Sources: the three tiers above, and who approves adding a new source.
• Access: who can use it and how permissions are reviewed (for example, quarterly).
• Accuracy: answers must cite sources, and staff must verify anything used for pricing, legal, HR or client commitments.
• Owners: who owns the policy and who owns each core document.
• Incidents: what to do if the AI surfaces something it shouldn't, including how to assess whether it's a notifiable data breach.
Our guide to AI governance for Singapore SMEs maps the bigger frameworks down to a one-page policy you can adapt.
When your second brain starts taking actions
Once your AI can act (sending emails, updating records, filing documents) and not just answer, the governance bar rises. IMDA's Model AI Governance Framework for Agentic AI, launched in January 2026 and updated in May 2026, is voluntary, but its four themes make a good checklist:
• Bound the risks up front: limit which systems and actions the agent can use.
• Keep humans meaningfully accountable: name who approves what, and require sign-off for consequential actions.
• Use technical controls: logs, permissions scoped to the task, and the ability to stop the agent.
• Enable end users: tell staff what the agent does and how to flag problems.
Organisations remain responsible for what their agents do, whether they're built in-house or bought.
Frequently asked questions
Does the PDPA allow using AI on customer data?
Yes, the PDPA doesn't prohibit AI, but its obligations still apply. You need an appropriate purpose (and consent where required), reasonable security, limits on access, sensible retention and proper safeguards for overseas processing. Using client files to help staff serve that client is usually within purpose; using them for unrelated purposes may need fresh consent.
Should HR files go into an AI second brain?
Not at first. HR, payroll, medical and disciplinary records carry the highest risk and are rarely needed to answer everyday questions. Keep them out until you have a clear use case, tightly restricted permissions and a written policy. General HR policies such as leave rules are fine to include, because they contain no personal data.
Is it a data breach if the AI shows staff a file they shouldn't see?
It may be. If personal data is accessed by someone who shouldn't have it, you should assess the incident under the PDPA's data breach notification obligations, which require notifying the PDPC and affected individuals when a breach is likely to cause significant harm or is of significant scale. Fix the underlying permissions immediately and record what happened.
Do I need a written AI policy as a small business?
It isn't specifically required by law, but it's strongly advisable. A one-page policy covering purpose, allowed sources, access, accuracy checks, owners and incident handling shows reasonable care under the PDPA's accountability principle, and it's what clients and auditors increasingly ask to see.
PreviousBest AI Second Brain Tools for SMEs in 2026: ComparedNext How to Build a Company Second Brain With AI: 7 Steps for SMEs
Want to Apply This to Your Business?
We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.