Why Swiss-hosted AI matters for medical data

A triage nurse pastes a patient message into a classification tool and has no way of knowing whether that text gets logged, used to train a model, or moved to another country.

For medical data, Swiss-hosted AI with zero data retention is not optional infrastructure—it is a compliance boundary. Laya Studio runs inference on GPUs in Switzerland and stores nothing you send, which keeps patient messages under Swiss jurisdiction and out of training pipelines. That distinction determines whether you can legally automate patient message triage without building your own data center.

Key takeaways

• Swiss hosting keeps healthcare data under Swiss jurisdiction, which matters for nFADP and GDPR-adjacent workflows.
• Zero data retention means patient messages are not stored, logged for retraining, or shared with third parties.
• A decision API returns calibrated probabilities, not generated text, so it cannot hallucinate a triage label.
• You can route patient messages by urgency or intent without sending PHI to an LLM that trains on your input.

Why does Swiss hosting matter for medical data?

If you are processing patient messages, the physical location of the inference server determines which data-protection regime applies. Swiss hosting means your inference runs on GPUs located in Switzerland, so the data does not leave Swiss jurisdiction during processing. Laya Studio answers on GPUs in Switzerland, and nothing you send is stored (Swiss data residency · Laya Studio). For a health-tech team, that is the difference between a vendor that processes data in an unnamed cloud region and one you can point to on a map.

I have worked with teams that routed patient intake through a US-based API only to discover, months later, that their data-processing agreement did not cover the actual inference location. Under Switzerland's revised Federal Act on Data Protection (nFADP), controllers must know where personal data is processed and be able to demonstrate it (Federal Act on Data Protection · Fedlex). If your vendor cannot name the country, you cannot document the flow. Swiss hosting removes that ambiguity.

That said, hosting location alone does not make an AI system safe. You still need a lawful basis for processing health data, and "Swiss-hosted" does not automatically mean HIPAA-equivalent. It is one layer of the stack, not the whole stack.

What does zero data retention actually mean?

Most AI APIs retain inputs for some combination of abuse prevention, model improvement, or logging. Zero data retention means the provider does not store your input after the inference response is returned. At Laya Studio, nothing you send is stored (Swiss data residency · Laya Studio). The request is processed, a probability is returned, and the text is not kept on disk for retraining or human review.

For medical data AI, this is the difference between a triage tool and a data-breach waiting to happen. If a patient message contains a symptom, a medication name, or an insurance number, you do not want that text sitting in a vendor's log bucket. Zero retention also simplifies your data-processing agreement: there is no vendor-side dataset to audit, delete, or port on request.

The trade-off is operational. Without retained logs, you cannot ask the provider to replay last month's traffic to debug a misroute. You must keep your own audit trail if you need one. I treat zero retention as the default and store my own structured decisions—question, answer, timestamp—separately from the raw text.

How does patient message triage work with a decision API?

You send the patient message plus typed questions, and get back a calibrated probability for each one in a single forward pass. Calibrated probability means the number reflects how often the model is right: a score of 0.80 should correspond to an 80% likelihood that the label applies. Because Laya does not generate text, it cannot invent a triage category that does not exist.

A real week looks like this: Monday morning, the support inbox has 200 patient messages. Instead of reading each one, you submit each message with three questions:

• Is this urgent? (yes/no)
• Does it describe new symptoms? (choice: none / mild / severe)
• Should this route to a clinician or a care coordinator? (choice)

The API returns probabilities for each question in about 120 ms from Europe. You route anything above your urgency threshold to a clinician queue and the rest to standard triage. Because the output is structured, you never have to parse an LLM's free-text explanation to figure out what it meant.

I prefer this over prompting an LLM because an LLM can confidently say "route to emergency" when the message says nothing of the sort. A calibrated score lets me set the threshold myself based on what my clinicians can actually handle.

What do nFADP and GDPR require of AI in healthcare?

Switzerland's nFADP and the EU's GDPR both treat health data as special category data requiring heightened protection. Both regimes require you to know where data is processed, to limit processing to stated purposes, and to be able to demonstrate compliance to a regulator. The nFADP is the current Swiss statute governing this (Federal Act on Data Protection · Fedlex); GDPR applies when you process data of people in the EU.

In practice, that means your vendor must be able to tell you the processing country, the retention period, and whether data is used for training. Swiss hosting answers the first question. Zero retention answers the second. You still own the third: you need patient consent or another legal basis before sending a message to any API, Swiss or otherwise.

I do not claim that using a Swiss API makes you compliant by itself. It removes the vendor as a compliance variable, but your own documentation, consent flow, and breach plan still do the heavy lifting.

How does a decision API compare to an LLM prompt?

This is the table I wish I had before I migrated a triage pipeline. The LLM prompt looked cheaper at first glance, but token costs and parsing failures made it more expensive in practice. The cloud NLP tools returned labels without probabilities, so I could not tune the urgency threshold to my team's capacity.

Laya is compatible with TypeSafe Jev's /v1/systemone protocol, so if you are already using Jev, switching is a configuration change rather than a rewrite. It is an independent service, not affiliated with Convai Innovations or TypeSafe.

What should I look for in a Swiss AI API?

Before signing anything, run this checklist: residency, retention, output type, hallucination risk, audit trail, and pricing. Each one determines whether you can safely hand patient messages to a vendor.

• Residency: Can the vendor name the exact country where inference runs?
• Retention: Is there a written zero-retention commitment, and does it cover logs as well as inputs?
• Output type: Do you get calibrated probabilities you can threshold, or free text you must parse?
• Hallucination: Does the model generate text, or only score questions?
• Audit trail: If the vendor stores nothing, do you have a place to store your own decisions?
• Pricing: Is it priced per question so you can predict monthly cost?

If a vendor cannot answer the first two questions in writing, do not send them patient messages. I learned that the hard way when a "EU-hosted" provider turned out to route inference through a third country during peak load.

If you are triaging patient messages and need a decision API that keeps data in Switzerland with zero retention, get an API key at Laya Studio.

FAQ

Does Laya store my medical data?

No. Laya Studio runs inference on GPUs in Switzerland and stores nothing you send (Swiss data residency · Laya Studio). You are responsible for keeping your own audit trail if you need one.

Is this GDPR compliant?

GDPR applies to processing of EU residents' data, and Swiss hosting helps with data residency but is not a complete compliance solution. You still need a lawful basis for processing health data. Switzerland's own framework is the nFADP (Federal Act on Data Protection · Fedlex).

Can it hallucinate a triage label?

No. Laya does not generate text; it returns a calibrated probability for each question you ask, so there is no generated label to hallucinate.

How fast is it?

About 120 ms from Europe for a single forward pass covering all your questions.

What languages does it support?

English plus 100+ languages.