Legal & compliance
How Teman is described, what it may never claim, how your data is handled, and where the service stands in each market we plan to enter — written from our regulatory research, with the open questions stated as open.
How we describe Teman
One official description — used in the app, the website, contracts and listings — so the service never drifts into claims it cannot stand behind.
The official description
Teman is an AI-guided emotional-support service for adults in Singapore. It offers conversation, guided self-help and structured check-ins. Its frameworks are a practising counsellor’s, held in two curated knowledge bases, and set by a panel of qualified counsellors who take referrals and answer for the service.
Teman is not a healthcare service, not therapy, and not a crisis line. It does not diagnose, prescribe or treat. Its design keeps it deliberately outside clinical and medical-device territory — it produces no clinical outputs. Where a person needs professional or emergency care, Teman’s only job is to say so and hand them to humans who provide it.
This paragraph is the single source of truth for how the service describes itself.
The words we never use
- “Counselling” for what the AI does — counselling describes our human panel, never the software
- “Therapist”, “psychotherapist”, “psychologist” — for the AI, or for any person not actually registered under those titles
- “Diagnosis”, “assessment”, “screening”, “treatment”, “clinical” about the AI — these words pull a product into clinical characterisation, and disclaimers do not pull it back out
- “Clinically proven”, “reduces anxiety”, “treats depression” — no outcome claims without a defensible evidence base
- “Anonymous”, “100% private” — almost never true once data is processed; we promise precise things instead
- “Crisis service”, “suicide prevention”, “emergency care” — false, and dangerous if believed
Singapore: where Teman stands
The legal ground under the design — checked against statutes, ministry guidance and the professional code we build to.
Licensing — no category applies today
Counselling, psychology and psychotherapy are not licensable healthcare services in Singapore today: the Healthcare Services Act schedules sixteen licensable services, and no mental-health counselling service is among them. But the Act defines a “healthcare service” broadly — care of the body or mind, reward or not. We therefore design strictly on the non-clinical side of that line, and watch for any future scheduling.
AI in health — the standard regulators would reach for
MOH and HSA’s AI guidance (AIHGle 2.0, 2026) is written for licensed providers, so it does not formally bind a non-licensed service — but it is the yardstick a regulator would use. Its sharpest rule: consumer AI must not produce outputs that require clinical expertise to interpret safely, even when accompanied by disclaimers. Disclaimers never rescue a clinical design. That is why the AI here is never the therapeutic actor, and every screen says what it is.
Titles and the professions
“Counsellor” is not a protected title in Singapore; SAC registration is voluntary and has been since 2003. Psychologist titles are changing: MOH announced in May 2026 that the five subdisciplines — including counselling psychology — will be reserved for registered practitioners, with a public register to follow. Teman never uses reserved titles, for the AI or for anyone on the panel who is not registered.
The code we build to
We hold the operating design to the Singapore Association for Counselling’s Code of Ethics: truthful public statements (Section I), confidentiality bounded by the duty to protect from harm (B.4), competence and supervision (C.2–C.3) — and no fee sharing with third-party agents (A.4.3). Our panel counsellors’ registration obligations stay theirs.
And A.4.3 is now the live question, not a settled one. Sessions are booked and paid through Teman, which collects the member’s payment and remits the counsellor’s share after deducting an administration fee. That is a fee-sharing structure with a third-party agent, and A.4.3 is the clause that speaks to it. We are not going to argue the point in our own favour on a public page. It is with Singapore counsel, and the answer decides whether this model can open at all. Until it is answered, no panel counsellor is asked to rely on the arrangement — nobody joins the panel on terms that depend on an answer we do not have.
The five rules our design follows
Not values statements — constraints with legal consequences, drawn from the jurisdictions we plan to enter.
The AI is never the therapeutic actor
Engagement, psychoeducation, triage, between-session support, crisis detection — yes. The therapeutic relationship belongs to humans. Several US states now prohibit AI from providing or advertising therapy, defining therapeutic communication to include offering emotional support in response to distress. The AI’s role here is narrower than that line, on purpose.
Scope, not disclaimers
The exemptions regulators provide are for self-help and educational resources that do not purport to offer therapy. Teman is built as support and guided self-help — content that helps you think, not a service that treats you — because a disclaimer never rescues a clinical design.
A qualified human owns every risk
Risk detection hands over to 24/7 human lines immediately, and stamps the flag on the record where the turn is stored — a stamp that survives deletion of the conversation. The screen that does it runs before the model is called at all. The AI never carries active risk alone — that is the design, not an escalation path.
“Counsellor” describes people
The AI is never called a counsellor, therapist or psychologist — not in copy, not in the system prompt, not in interface strings, not as a persona name. Every human described as qualified actually is, and their registration is their own.
Truthful by construction
No efficacy claims, no outcome promises, no invented supervision. What we say the service does is what it does — and where something is a design intention rather than a shipped fact, we say that too.
Why this discipline pays
Under the EU AI Act and medical-device guidance, what the maker says the product is for drives its legal class. Marketing copy is the highest-leverage decision in the file. The same discipline protects us in Singapore, the region, and everywhere after.
Money, referrals and independence
- A counsellor session is booked and paid for through Teman. Teman collects the member’s payment on the counsellor’s behalf, deducts an administration fee, and remits the balance to the counsellor
- The fee itself is the counsellor’s own, set by them within a published band (S$240–380). We do not discount it, and no part of it is set by how many people they see
- Counsellors are paid a fixed professional fee for the panel work they do — setting the frameworks, taking referrals, answering for the service — never linked to how many referrals they receive
- The business model is the platform — subscriptions for the AI service, plus the administration fee on sessions. It is not a margin on anyone’s clinical judgement, which is why the fee is the counsellor’s to set, and why panel work is paid work rather than volunteered
SAC Code A.4.3 is the anchor, and it is the clause this model has to answer to: members shall not participate in fee sharing arrangements with third party agents. Collecting a member’s payment and remitting a share is exactly the structure that clause speaks to, so the honest position is that we do not yet know whether SAC-registered counsellors can join the panel on these terms. A flat per-session administration charge rather than a percentage is the structure that stands the best chance, and it is a one-constant change if the opinion comes back against a percentage. That question is with Singapore counsel, and until it is answered no counsellor is asked to rely on the arrangement.
Complaints and accountability
- Complaints reach a person, not a chatbot — with a published route and response times
- Anything about the quality of care goes to the panel, and where relevant, to the counsellor’s own professional body
- Panel counsellors carry their own professional indemnity cover; the platform carries its own
- Every safety incident is written up in the review register — what happened, what was decided, and who decided it — rather than closed by a policy
What the AI will not do, how the crisis screen works and what happens when it fires, is set out in Safety, design & limits.
Your data, under Singapore’s PDPA
Mental-health conversations are among the most sensitive data a service can hold. So this section says what protects them, what does not, how long they are kept, and what happens if they are ever exposed — on the machine serving this site, as it is today.
Where your conversations are kept, today
Your conversations are kept on Teman’s own server — a single machine in a Singapore data centre. What that machine is, and what it is not, is the paragraph below.
One machine we control is easier to reason about than a service we do not: the whole of your history is one file that can be inspected, backed up, and deleted from. It is also the same file that has to be protected, so what protects it — and where that protection stops — is the card beside this one.
Your account and conversations are on that server. Your check-ins, bookings and cases are still kept in the browser you used, and do not follow you to another device.
The safeguard, precisely — and what is missing
- In transit, encrypted — the site is HTTPS only
- Passwords are never stored — only a scrypt hash, which cannot be turned back into your password
- Every read and every write is scoped to your account — the server resolves who you are from a session cookie the page cannot read, never from anything the page sends, so no request can ask for somebody else’s conversation
- The service runs sandboxed — a read-only filesystem, one writable directory, and its own service account
- No sale, no ads, and no consent-free reuse — nothing you write is used to improve the system unless you separately, explicitly opt in. That switch is on your Member page: off unless you turn it on, and yours to turn back off. Nothing reads it yet, so today nothing is used either way
Conversations are encrypted in transit. On the server they are protected by the operating system’s file permissions and a service account that can write nothing else — not by encryption at rest. Anyone with root on that machine can read the database file. Access to the machine is the boundary, and we would rather say that than let “encrypted” stand unqualified.
One more, because it is easy to assume otherwise: your messages do reach the AI provider, and that provider is outside Singapore. Identifiers are stripped on the way out — NRIC and FIN numbers, phone numbers, email addresses, links, handles and postal codes are replaced with markers before the call, and the words themselves stay in the store on this server. Names, places and employers in ordinary sentences are not removed. A pattern cannot find “my supervisor at DBS, Priya” without either missing it or shredding the sentence it is in, so that is not attempted rather than half-done. The pass reduces what crosses the border; it does not make the text anonymous, and we do not call it that.
How long we keep it, and what happens if it leaks
If you use Teman without an account, what you say is kept too. That changed on 9 October 2026 and it is worth saying plainly rather than burying: a conversation held before you sign up is stored on this server, on the same terms as one held inside an account, and the team running Teman reads it while the service is new. The footer links this page from every page of the site, and this is where it is stated in full — the same 24 months apply to those conversations and the same age-based delete covers them; what a guest does not have is the button, because there is no account for a delete to be scoped to. If you would rather not be read at all, 1771 is answered by a person who is not Teman.
Conversations are deleted after 24 months with nothing said in them — counted from the last message, not from when the conversation began, so one you still use is never purged and one you have left behind does not sit on the machine indefinitely. It is the same real delete as the ones you can ask for below, and what survives is the same textless record: when it went, how many messages went with it, and that the reason was age rather than a request.
When a reply is read aloud, the audio is kept with the conversation — that is what makes the speaker button still work when you open a past conversation months later. It is stored beside the message it reads, it is deleted with it (a delete takes the voice as well as the words), and it ages out on the same 24 months. Nothing is kept for a conversation you never signed in for: a guest's replies are read aloud at the time and not stored.
If data is ever exposed it is handled as an incident rather than filed under a policy: opened on the machine, assessed as notifiable or not, and — where it is — notified to the Commission and to the people affected, inside the region’s strictest clock of 72 hours. Singapore’s PDPA sets no hour count; 72 is the bar we hold ourselves to, not a local deadline. The assessment and the reason behind it are recorded either way, because an incident that was assessed and closed is the record the law asks for, and an incident quietly dropped is not.
What you can ask for
- Deletion while you are signed in — delete any single conversation from Chat sessions, or every conversation at once from Member. The messages are destroyed, not hidden, and the spoken audio goes with them: a delete is a real delete, and it is not recoverable
- What survives a deletion, and why — a record with no text in it: the dates, how many messages went, and whether the crisis screen had fired. It exists so that a conversation the crisis screen fired in is not indistinguishable from one where it stayed quiet — the fact that a screen matched, and when, is the one thing a deletion does not take with it. It holds no message, and no title — a title is your own first sentence, so it goes with the rest
- Access, correction and export — a copy of your data and anything wrong put right, within statutory timelines. Today this is answered by a person over email rather than self-service
- Closing your account entirely — not built yet. It needs rules for bookings you have paid for and are owed, so it is being written rather than improvised. Until it exists, ask and it is done by hand
- A human — data requests and complaints are answered by people, not a chatbot
The data groundwork is built once, to the strictest standard in the region rather than four times over as each market opens — Indonesia’s GR 33/2026 is the tightest bar in it, and the one this section is written against.
Where we plan to operate
Regulatory posture market by market, from the October 2026 research pass. Colour = how open the door is for this design.
Singapore Home — open
First market. No licence category for counselling today; a deliberately non-clinical design; the PDPA and AI-governance foundation described on this page is built here first.
Malaysia Phase 2 — gated
The Counsellors Act 1998 gates the service of counselling for a fee, not only the title. Any AI entry must be a tool, never the counselling service itself — a structure we take to Malaysian counsel before shipping.
Indonesia Watch — data strict
The region’s strictest data bar: 72-hour breach clock, mandatory impact assessments that name AI explicitly, compliance deadline January 2027. We build to it once, globally.
Thailand Open today — window closing
No law directly controls counselling psychology today, while clinical psychology is licensed — and a 2025 petition is pushing to change that. A market to enter early or watch, not to assume.
Philippines Not an early market
The Psychology Act captures “psychological counselling … coaching” with penalties, and a separate law bars operating a guidance or counselling centre without a Board permit. Avoid until the structure allows.
US & EU Design discipline, not launch
Some US states permit emotionally supportive AI with guardrails; others prohibit AI therapy outright. The EU lets intended-purpose wording drive classification. Both reinforce rule 01: the AI is never the therapeutic actor.