Mental Health Helplines Shouldn’t Record Calls by Default: The Case for AI Voice Agents That Forget

AI voice agents answering mental-health and crisis lines should store nothing that links a caller to what they said. No recording, no transcript, no call record, no saved number. Someone ringing a helpline at their lowest point should not have to wonder whether a tenant admin, a platform admin or anyone reading a database or log could later connect their name to their words. Forgetting should be the default for these lines, and convenience features should give way to it.

Saloma AI’s High Privacy mode is one working example: a per-agent switch that keeps a caller’s identity and the content of the call apart so the two cannot be joined. It is worth examining closely, including where it stops.

Why a helpline should forget by default

On a sensitive line, the safest stored record is no record at all. What is never kept cannot be leaked, misused or joined to something else later. That is the design goal behind High Privacy: nothing stored should let anyone link a person to what they said, asked or needed.

The risk is not only an outside attacker. It sits with the people and systems that routinely touch call data:

• a tenant admin at the helpline itself;
• a platform admin at the software provider;
• anyone able to read the database;
• anyone able to read the logs.

Each is a point where a caller’s identity and words could meet. A helpline that keeps recordings, transcripts and numbers must trust every one of them, indefinitely. A helpline that keeps nothing linkable does not have to.

In Malaysia, recording also carries legal weight. Section 234 of the Communications and Multimedia Act 1998 prohibits the interception and disclosure of communications sent over a network without lawful authority. The Personal Data Protection Act 2010 requires consent before personal data, including voice recordings of private individuals, is collected, used or stored. This is context, not legal advice, and it says nothing about whether any particular helpline is compliant. It does show why every stored call becomes an obligation to manage.

General AI voice platforms, Saloma AI included, are built for businesses. Saloma AI offers live listening, workflows that update customer records and follow-up messages. Those features earn their place on a clinic enquiry line or a support desk. On a crisis line, they are the wrong starting point.

How Saloma AI’s High Privacy mode keeps identity and content apart

High Privacy is a switch set on each agent, intended for crisis, mental-health and care lines. It is opt-in and applies to inbound phone calls only. Once it is on, the platform treats every call as two halves that must never meet.

• Identity: the caller’s number, name, SIP ids, IP address and the exact time of the call.
• Content: the audio, transcript, summary, topic and any tool use.

The identity half is never read. Caller ID is cleared in the phone routing before the voice server, the AI, the logs or the cache receive it. A bug cannot leak data the system never loaded.

The content half is never saved. The conversation sits in memory only for the length of the call and is dropped at hangup. No call ID or session ID is stored, so there is no join key that could later tie a voice to a number.

To hold that line, the mode switches off features that would otherwise store or expose a call:

• recording, transcripts, AI summaries, and sentiment and quality scores;
• caller memory and number-based contact lookup;
• automations such as callbacks, emails, webhooks, CRM updates and SMS;
• supervision, including listening, whispering and barging in;
• outbound use, as the agent cannot join a campaign or dialer.

Only three tools remain: ending the call, searching the knowledge base, and transferring to a human with an anonymous caller ID.

Smart Protection is also off, so there is no number blocking and no decision record. A person in crisis is never cut off by an automated rule.

What is kept, and the safeguards behind the switch

High Privacy keeps only what a helpline needs to run and pay for the service, and nothing that describes a single call.

• Anonymous hourly counters per agent: calls, answers, transfers, short hang-ups and errors, plus total seconds and a duration band.
• Cost and token totals: used for billing and capacity planning.
• No per-call row: there is no entry for any individual conversation.

Small numbers are hidden too. Any count under 5 shows as “fewer than 5”, reads are merged into whole days and the current day is excluded, so a quiet hour cannot point to one caller.

The safeguards matter as much as the switch. The mode fails closed: if the agent settings cannot load, or the phone routing and the database disagree, the call runs in High Privacy and never falls back to normal. The policy is frozen when the call begins, so a mid-call change cannot alter it. One central policy is checked by every point that could store, send or log anything, rather than checks scattered through the code.

Testing backs this up. A canary-style test runs a simulated private call with sentinel values and confirms they appear in no logs, rows, cache keys or outgoing payloads. A build guard fails if a new call-data table has no privacy decision. An operator privacy scan reports “incomplete” rather than “clean” when it cannot see everything.

For administrators, enabling the mode asks for confirmation. Disabling needs a typed phrase and a reason. Locked controls cannot be edited, the audit records admin actions only, never callers, and erasing an agent’s existing call history is offered when the switch goes on.

The strongest objection: what if the caller is in danger?

This objection deserves the most weight. A helpline that keeps no numbers cannot locate a caller who is at risk. It cannot call them back if the line drops. Supervisors cannot listen in, whisper guidance or barge in when a conversation turns serious. Saloma AI says so in its own documentation: if a caller is in danger, the platform cannot locate or call them back.

For a crisis service, that is not a minor gap. Some counsellors and managers will reasonably feel that the ability to trace or follow up a caller outweighs the privacy gained.

The answer is not to pretend the gap away. It is to design the call so that help is reachable within it. Saloma AI’s guidance is that scripts on a High Privacy agent should:

• give callers emergency numbers clearly during the conversation;
• offer a live transfer to a human, one of the three tools still available.

The transfer goes out with an anonymous caller ID. A caller who needs a human can reach one. What the service gives up is the ability to act on the caller’s behalf without their involvement.

That is a genuine trade-off, and no software settles it. Some lines will judge that anonymity encourages people to call who otherwise would not. Others will judge that traceability matters more for their callers. Each service has to weigh this against its own duty of care, clinical guidance and the people it serves, and choose deliberately rather than by default.

Honest limits every helpline should know

A privacy feature is only as trustworthy as the limits it admits. High Privacy is Saloma AI’s own practice about what the platform keeps and shows. It is not a promise about what AI or telephony providers do, because Saloma AI cannot control them. Where a provider offers a no-retention or no-logging option, the platform requests it. No named provider is claimed to be zero-retention.

A helpline considering the mode should weigh these points before advertising it to callers:

• The phone network still knows. The carrier already knows the number called and the time. Stronger anonymity requires callers to withhold their caller ID at the carrier.
• The call passes through AI providers. During the conversation, audio and text must travel through AI providers, so the call exists in memory while it runs.
• Tiny volumes can still hint. Even with small counts hidden, a very quiet line may still suggest that a single call took place.
• Old history stays until erased. Records from before the switch remain unless the administrator erases them.
• Phone only, for now. Browser and website voice entry is not covered. A service should not promote an anonymous web entry point.

None of these weakens the case; they define it. Real AI mental health helpline privacy means storing nothing linkable by default, justifying anything kept, and telling callers exactly what is and is not protected. That honesty earns more trust than any implied promise of total anonymity.

Sanjay W