All articles

Local AI or Cloud AI? A Decision Guide for Swiss SMEs

Local AI, cloud AI or hybrid: a checklist for Swiss SMEs based on data type, professional secrecy, usage, IT capacity and budget.

Local AI7 min read

Most partners and managers at fiduciary offices and law firms now know that their staff want to use AI, and often already do. What is still open is where that AI should run. This guide sets out the three common options and walks you through a checklist so you can make a reasoned choice. It is not legal advice, but it will help you ask the right questions.

Three ways to run it

First, a quick definition of terms. We mean an AI assistant that answers questions about your own documents: engagement letters, file notes, tax records or internal guidelines.

Cloud AIHybridLocal AI
Where are the documents?With the providerOn your serverOn your server
Where does the language model run?With the providerWith the cloud providerOn your server
What goes over the internet?Documents, questions, answersThe question and a few short text excerptsNothing
Is a graphics card needed?NoNoYes
Typical examplesChatGPT Enterprise, Microsoft CopilotYour own server connected to Gemini, GPT or ClaudeYour own server with an open model

Cloud AI is the quickest to set up. Local AI gives you the most control. Hybrid sits between the two: your files stay in-house, and a cloud model is used only to phrase the answer.

The checklist

Go through the table row by row and note which column fits your office.

CriterionCloud AI fits if …Hybrid fits if …Local AI fits if …
Type of datamostly public or internal, non-confidential textclient data, but not large volumes of sensitive personal datahealth data, criminal proceedings, social assistance or similarly sensitive files
Professional secrecyyou are not bound by professional secrecyyou are, and may transfer short excerpts under a reviewed contractyou or your clients rule out any transfer to third parties
Client requirementsclients set no conditionsclients require Swiss or European data storageclients contractually prohibit sharing data with cloud services
Usage volumea few people, occasional use5 to 50 people, daily usevery high query volumes or automated processing of large archives
IT capacitynobody in-house looks after ITan IT partner looks after a serveran IT partner can also maintain GPU hardware
Budgetyou prefer a monthly fee per personyou want low start-up and running costsyou can carry an up-front hardware investment
Internet connectionalways availablealways availablethe network is partly or fully isolated
Timelineyou want to start within daysyou want to start within weeksyou can allow time to procure hardware

How to read your answers

Do not just count the ticks. Some criteria carry more weight than others.

A single tick in the "local" column for professional secrecy or client requirements is usually enough. If an engagement letter rules out passing data to cloud services, no price advantage changes that. The same applies to a network that has no internet connection for security reasons.

Budget and volume usually point small offices towards hybrid or cloud. For an office of 15 people, the pure token costs of a mid-range cloud model are well under a hundred francs a month. The calculation is in our article on AI token costs. At that volume, a dedicated graphics card does not pay for itself.

IT capacity is an honest deal-breaker. A local AI server needs updates, monitoring and someone who responds when it goes down. Without a partner for that, hybrid is the more realistic place to start.

Two provisions of Swiss law matter for fiduciaries, tax advisers and lawyers.

Art. 321 of the Swiss Criminal Code makes breach of professional secrecy a criminal offence. It covers, among others, lawyers, notaries and auditors, together with their auxiliaries. Whether a cloud provider can count as an auxiliary, and under what contractual terms, is assessed differently in the legal literature. Check this with your professional association or a specialist.

Art. 62 of the Data Protection Act (DSG) applies to anyone who learns secret personal data in a profession that requires such knowledge. That includes many fiduciaries who are not covered by Art. 321. In Art. 5(c) the DSG also defines sensitive personal data, such as information on health or on social assistance measures.

There is also a technical question. If the provider is subject to US law, the CLOUD Act allows US authorities to compel it to hand over data, even if that data sits in a European data centre. With hybrid operation, this concerns only the excerpts sent; with local operation, nothing at all.

Three examples

The following firms are invented. They show how the same checklist leads to different results in practice.

Fiduciary office with 8 people. Bookkeeping, VAT returns and tax returns for SMEs and private clients. No health data, no client contracts forbidding cloud use. An IT partner already looks after the office server. Result: hybrid. Files stay in-house, costs are low, and moving to local later remains possible.

Law firm with 25 people, focused on criminal and family law. Files contain information on criminal proceedings and health, and some clients explicitly require that nothing goes to a cloud. Result: local, even though it costs more.

Communications agency with 12 people. Works mainly with public texts and presentations, no professional secrecy. Result: a cloud service such as ChatGPT Enterprise or Copilot is reasonable here and the simplest choice.

Where KADARAG fits in the table

KADARAG covers the "hybrid" and "local" columns. In both, documents, search and logs stay on your server, access rights follow your existing file permissions, and every answer cites the document and page. In hybrid mode, only the question and three to five short excerpts go to Gemini, GPT or Claude. We agree a fixed price after an introductory call.

Your next step

Print the checklist and go through it with whoever looks after your IT. While you do, check two or three of your engagement letters for clauses on sharing data. If questions remain after that, we are happy to talk them through in an introductory call.

See KADARAG answer your own questions

30 minutes, no commitment. We show you how KADARAG works with documents like yours and whether local or hybrid fits.