AI Usage Policy for Singapore SMEs: What Staff May and May Not Paste Into Public AI Tools

An AI usage policy for a small business in Singapore does not need to be a legal document. It needs to answer one question for every employee: what may I paste into a public AI tool, and what must I never paste? For most 10-to-50 person companies the workable answer is that anything already public or already generic is fine — draft marketing copy, job advertisement wording, meeting agendas, a spreadsheet formula, code you wrote yourself with no credentials in it. Anything that identifies a person, prices a named client, or unlocks a system is not. Put that on one page, name one person who owns it, and you have addressed the substance of what the Personal Data Protection Act expects of you long before you have satisfied a lawyer.

Why does a 20-person company need an AI policy at all?

Because your staff are already using these tools, and right now you have no idea what has left the building. Surveys of workplace AI use consistently show adoption running well ahead of employer awareness, and the pattern in Singapore SMEs is familiar: someone in sales discovered that a chatbot writes better quotation cover notes than they do, someone in HR uses it to shortlist CVs, someone in accounts pastes in a supplier statement to reconcile it faster. None of them did anything malicious. All of them made a disclosure decision on your behalf, without a rule to follow.

The exposure is rarely the dramatic one. It is a CV with an NRIC number in it, a client contract, a price list your competitor would pay for, or a customer database extract pasted in to "clean up the formatting". Under PDPA, that is your organisation disclosing personal data to a third party. The employee's good intentions are not a defence, and "we didn't know they were doing it" is worse than knowing, not better.

What can staff safely paste into a public AI tool?

Be specific, because vague policies get ignored. A green list that works for most SMEs:

• Content you would publish anyway — website copy, social posts, brochure text, job advertisements
• Generic business writing — a polite chaser email with names and amounts removed, a meeting agenda, a policy outline
• Public information you are asking the tool to explain — a government guideline, a standard, a piece of legislation
• Anonymised or invented examples — "a customer ordered 200 units at an unspecified price"
• Your own working artefacts with identifiers stripped — a formula, a script, a process map

The test staff can remember: would I be relaxed if this exact text appeared on the front page of a newspaper, attributed to us? If yes, paste it. If you hesitate, it belongs on the red list.

What must never leave the building?

A short red list beats a long one. Four categories cover the bulk of real incidents:

• Personal data. NRIC and FIN numbers, full names with contact details, CVs, payroll figures, medical or disciplinary records, customer lists.
• Commercially sensitive material. Signed contracts, tender submissions, margin and cost breakdowns, client-specific pricing, anything under an NDA.
• Credentials and system access. Passwords, API keys, database connection strings, internal URLs. This is the one staff underestimate most, because it usually arrives buried inside a config file they pasted for debugging.
• Anything belonging to a client. Their data is not yours to disclose, and your contract with them almost certainly says so.

Add the escape valve, or the policy will be broken quietly: if someone genuinely needs AI help with red-list material, they route it through an approved business-tier tool with data-retention controls, or they ask the policy owner. Prohibition without an alternative just drives usage underground.

Does PDPA actually prohibit using AI at work?

No. PDPA does not ban AI tools. It holds you responsible for personal data you collect, use and disclose — including when the disclosure happens through a chat window. The PDPC has published advisory guidance on using personal data in AI systems, and the direction is consistent: consent, purpose limitation and reasonable security still apply, and the technology does not create an exemption.

Two duties matter for a small company. First, you must designate a data protection officer, and Singapore has moved toward requiring that the DPO's contact details be registered and kept current — a five-minute administrative task that is easy to leave undone until it is embarrassing. Second, you have breach notification obligations when a data breach is likely to result in significant harm or affects a large number of individuals. Pasting a customer list into a public tool is capable of meeting that threshold. Decide now who assesses that, because you do not want to be working it out for the first time under a deadline.

Where do the human checkpoints belong?

Most AI failures in SMEs are output failures, not input failures — a confidently wrong number in a quotation, an invented regulation in a client email, a hallucinated part specification. Three checkpoints catch nearly all of it:

• Anything that goes to a customer is read by a human before it is sent. No exceptions for "it was only a draft".
• Anything involving money or quantities — quotations, invoices, stock figures, payment instructions — is checked against the source system, not against the AI's summary of it.
• Anything affecting a person's employment — shortlisting, appraisal wording, disciplinary drafting — is a human decision with a human record. This one intersects directly with workplace fairness obligations arriving in 2027, and an AI-generated shortlist with no documented human reasoning is a weak position to defend.

How do you write and enforce this without an IT department?

Budget an afternoon, not a project. Draft one page: purpose, green list, red list, approved tools, three checkpoints, who to ask, what to do if you think you have pasted something you shouldn't have. Name one owner — usually the same person who is already your DPO. Circulate it, get an acknowledgement in writing from each employee, and file those acknowledgements.

Enforcement in a small team is cultural, not technical. Make the reporting route blameless: someone who realises at 6pm that they pasted a client contract must find it easier to tell you than to hide it, because your notification clock may already be running. Review the page every six months, and treat near-misses as the agenda for that review. A policy nobody has ever invoked is usually a policy nobody has read.

Frequently asked questions

Is a paid AI subscription enough to make us compliant?

It helps but it does not finish the job. Business and enterprise tiers typically offer data-retention and training-exclusion controls that consumer tiers do not, which meaningfully reduces risk. But compliance still rests on your own obligations — consent, purpose, security, DPO designation and breach response. Read the specific vendor terms rather than assuming; they differ, and they change.

What if staff use AI tools on their personal accounts and phones?

That is the most common real-world scenario, and your policy must address it explicitly. Company data is company data regardless of the device or account it passes through. State that the red list applies everywhere, and give people a sanctioned tool so the personal-account workaround stops being the path of least resistance.

Do we need to tell customers we use AI?

There is no blanket disclosure requirement, but be careful where AI touches decisions about a person — credit terms, hiring, service eligibility — or where a customer reasonably believes they are speaking to a human. Transparency is cheap; discovering it after a complaint is not. If your quotations or support replies are AI-assisted but human-checked, that is usually an easy and credible thing to say.

Digital Perpetual helps Singapore SMEs put practical AI governance in place — one-page policies, approved tool selection and the human checkpoints that keep AI out of your compliance file.