Which AI model fits a Swiss financial institution depends less on the model's name than on three questions: which data does it touch, does its result have to be traceable and reproducible, and where does it run? Answer those per use case and you land on one of four operating models, and sometimes on the conclusion that no model is needed at all. This holds for investment foundations and pension funds as much as for fund management companies and asset managers, FINMA-regulated or not.
The first operating model is a large language model accessed through an interface in the model provider's cloud. The second is the same kind of model deployed in your own cloud environment, in the region and under the contract you have already chosen. The third is an open model run on your own servers or infrastructure. The fourth is an AI feature embedded in software you already use. The four differ less in the quality of their answers than in where your data flows, who is responsible for operating them, and how easily you can replace the model later.
The first criterion is the data class. Public information, such as published factsheets or market commentary, tolerates almost any operating model. As soon as position data, client data or unpublished figures are involved, the question of where the model runs becomes a governance question. If it runs in your own environment, your data stays where you already host it: for Swiss institutions, typically in Switzerland. If it runs at the provider, that is a deliberate, documented outsourcing decision that has to meet your existing rules on outsourcing and data protection, not a side effect of a trial licence.
The second criterion is traceability. A language model does not necessarily give the same answer to the same question twice. For a draft, a summary or the extraction of fields from a document that a person then checks, that is unproblematic. For a figure that goes to investors or to a supervisor, it is not. There, the calculation belongs in deterministic logic that returns the same result on every run and can be checked step by step. The model may prepare, flag and suggest; it does not calculate and it does not sign off.
That is why the right answer is sometimes: no model. Amula AI's reporting pipeline has been live at IST Investmentstiftung since October 2025, and it is built deterministically on Power BI, with no AI model generating commentary or figures. 39 investment groups are in scope and being rolled out group by group, and monthly effort is down from 20+ to 3–4 hours (an internal estimate based on first measurements). The benefit there comes from clean data logic, not from the choice of a model.
The third criterion is replaceability. Models are superseded at short intervals, and the model that fits best today is rarely still the right one two years from now. Build the use case so that the prompts, validation rules, test cases and logs belong to you and the model behind them stays a replaceable component. An embedded AI feature is available quickly but ties you to its vendor's roadmap; a self-hosted open model gives you full control but requires someone to be responsible for running it: your team, your partner, or both.
In practice, this rarely produces a single model for the whole institution. It produces a short mapping: which use case, with which data class, runs in which operating model, and where deterministic logic deliberately takes the place of AI. That mapping is what the AI Assessment works out: typically in one to two weeks, ending in a report and a live walkthrough with three tiered implementation options. Know your bottleneck? Start with reporting. Not sure where to start? Start with the AI Assessment.