top of page

Local or Cloud AI: The Data Decides.

5 days ago
4 min read

Whether your practice's AI runs on its own computer or in a provider's data center is decisive for data protection. The sensitivity of the data being processed determines which route is right.


“The server is in Switzerland, so it complies with data protection law.” Some AI products for medical practices advertise with exactly this line. It answers the wrong question. The location tells you where data is stored. It does not tell you who can read it.


Anyone choosing AI for their practice needs answers to two other questions. Where does the AI do its computing? Which data does it get to see?

Where the AI computes determines the path your data takes


Take a report for Swiss disability insurance (IV). With a cloud solution, the physician enters her notes on the practice PC. The text travels over the internet to the provider's data center. There, a large language model writes the draft and sends it back. The report has left the practice, if only for a few seconds.

With a local solution, also called “on-premises,” the computer running the language model sits in the practice itself. The notes go from the practice PC to that computer and back. The report never leaves the practice.


Think of it like lab work. The cloud is the external laboratory: the sample leaves the practice, the equipment is more extensive, and a contract governs how it is handled. The local solution is the in-house lab: everything stays on site, the range is narrower, and maintenance and security are your responsibility.



A practical note: a program installed on the practice PC can still send text to a data center for processing. So don't ask where the program is installed. Ask where the AI does its computing.


Both routes have strengths, and both have limits


The cloud offers the largest and most up-to-date models. It delivers results in seconds, needs no hardware of your own, and is maintained by the provider. You pay on an ongoing basis, per use or by subscription depending on the offer, rather than once for your own hardware. Its limits: without internet it stops, and your data leaves the practice. A server location in Switzerland does not change that.


The local solution has the opposite balance. Patient data stays on site, and the AI keeps working during an internet outage. In return, the models are smaller and, depending on the hardware, slower. Purchase, setup, and maintenance cost money. And data security under Art. 8 of the Federal Act on Data Protection (FADP) rests entirely with you: updates, access control, backups. So local is not automatically secure. It simply means no cloud provider is involved. If someone maintains the computer remotely or backups are stored off site, that also needs clear rules.


For many tasks, speed does not matter. A draft report does not need to be ready in three seconds. It needs to be ready the next morning. A system that prepares the day's documents overnight is done by eight a.m. Going local pays off because of the data, not because of the price.



The data the AI sees decides the route


Under the FADP, health data is sensitive personal data. On top of that, professional secrecy applies under Art. 321 of the Swiss Criminal Code. The term reaches further than many assume: even a patient's name in connection with your practice counts as health data, as a guide from the Canton of Zurich's AI sandbox points out.


This leads to three cases. First, texts without personal data, such as a plain-language fact sheet on flu vaccination. Here, the cloud with a vetted business account is usually the pragmatic route. Second, personal data with no link to patients, such as staff rosters or correspondence with suppliers. Here, the cloud is possible if there is a data processing agreement, the location and sub-processors are disclosed, and your data is not used for training. That rules out free versions. Third, anything involving patients, from reports to appointment reminders. Here, the local solution is the first choice. FMH, the Swiss Medical Association, also recommends giving preference to locally operated or controlled models. For the cloud, professional secrecy requires additional safeguards: the provider must be technically unable to read the data, or it must be contractually bound like an auxiliary of the practice. That takes effort and needs legal review in each individual case.


That leaves the idea of simply replacing names before sending. A local program can remove most identifying details. In a study with real physicians' letters, it removed 99.24 percent of them. But a code in place of a name is not yet anonymous as long as the practice keeps a list linking code and person. The data remains personal data. Sharing less data is still the right move. It just does not replace choosing the right route.


Three steps for your practice


You do not need an IT department to make a sound decision. First: list which AI tools your team uses today, officially or unofficially, and for what. Second: assign each application to one of the three cases. Third: decide for each application whether it runs in the cloud or locally, and record the decision in writing.


The next time you receive a quote, you will know which question comes first: where does the AI do its computing, and who can read the data?


Ready for a first conversation?

In a free initial meeting, we look together at which AI applications in your practice involve which data, and which route fits. Get in touch: https://www.felber-advisory.ch/kontakt


Sources


This article provides general information and does not replace legal advice in individual cases.


 
 
 

Comments

bottom of page