The institute

Research at the Swissi Institute for AI

The publishing institute researches where technical design meets commercial and legal evidence: autonomous agents, supervision at scale, sustainability claims and structured publication.

Swissi Institute for AI

Research programme

Current lines of work

Durable Model-Configuration Attribution and Recovery-Conditioned Mandates for Autonomous Economic Agents

A Distributed-Ledger Framework Binding Acting Configuration, Assurance, and Revocation into One Replayable Record

Walter Kurz, Besnik Alidemi

Swissi Institute for AI

Autonomous agents can initiate payments, purchases and commitments while legal consequence still needs an accountable natural or juridical point. A ledger-backed mandate model binds the authorising party's live assurance state, the agent's profile anchor, the model configuration that decided, the revocation state and a recovery surface into one replayable record. The construct separates three things delegated-agent systems usually compress: who executed, whose authority and funds were used, and what configuration decided.

Contest Without Consensus: Distributed Ledger Design for Machine Actors and Human Signatories

Decidable Delegation, Adversarial Settlement, and the Interface-Corpus Constraint on Machine-Authored Systems

Walter Kurz

Swissi Institute for AI

Every ledger an autonomous agent acts on was designed when the acting party was a person, or a program imitating one, and wallets, recovery phrases and confirmation steps carry that assumption forward. Removing it makes agreement the wrong primitive: among actors drawn from similar distributions agreement is cheap and weakly informative, while disagreement between parties with opposed interests is costly and therefore carries information. The paper proposes contest with a deadline as the trust operation for whatever is not decidable.

Multi-Agent AI as a Nested Principal-Agent Problem in Private Wealth Management

Bargaining-Based Suitability and Context Control under the Legal Framework of Switzerland, Germany and Austria

Walter Kurz¹, Reinhard Magg¹, Stefan Marx², Frank Reinhardt², Florian Kollberg²

¹Swissi Institute for AI²Hochschule für Wirtschaft und Umwelt Nürtingen-Geislingen

Private wealth management in Switzerland, Germany and Austria runs on a directed asymmetry: the advice side is compelled to disclose, while the client side is protected in non-disclosure. A multi-agent system deployed there must recommend for a client whose type it cannot observe and cannot lawfully compel to reveal. The paper models this as a nested principal-agent structure in which the advisor's hidden action re-emerges at the AI boundary, and aggregates admissible facet-agent outputs through the Nash bargaining solution with a context-confidence score reported alongside each recommendation.

Context Substitution in Large Language Model Risk Assessment

A Methodological and Legal Framework for Pre-Judgment and Reputational Externalities in Switzerland, Germany and Austria

Walter Kurz

Swissi Institute for AI

Language-model assistants increasingly issue evaluative judgments about identifiable people, firms and offers without first eliciting situational context. The unobserved gap between the requested judgment and the information supplied is filled with the most salient cluster from the training distribution and communicated as a finding rather than as a contingent prior. The paper identifies context substitution as the root methodological pathology, traces its reputational and legal consequences under DACH personality rights, and proposes context elicitation as an architectural design principle for evaluative AI.

AI-Supported Supervision of Licensed Institutions' Websites by Financial Market Authorities

A Conceptual Framework from Supervisory Practice in Switzerland, Germany and Austria

Walter Kurz, Wojtek Stricker

Swissi Institute for AI

Financial supervisors in Switzerland, Germany and Austria have to review the public web presence of hundreds of thousands of licensed entities for regulatory compliance, a perimeter of roughly 250,000 entities at BaFin, 30,000 at FINMA and 4,400 at the FMA. Manual review at that scale is not feasible, and existing tools target static HTML and fail on applications that render their content only after JavaScript runs. Drawing on interviews with FINMA, BaFin and FMA review teams, the paper derives a reference architecture for an AI-assisted multi-agent supervision system.

Legally Sound Finfluencer Activity through AI-Supported Compliance Review

A Specialised Multi-Agent Framework for Investor Protection in Switzerland, Germany and Austria

Walter Kurz, Wojtek Stricker

Swissi Institute for AI

Financial influencers publish investment content at a frequency and breadth that manual screening cannot follow, which produces a three-sided tension across Switzerland, Germany and Austria. Investors want daily access to comprehensible financial information, creators want to publish regularly without supervisory exposure, and FINMA, BaFin and the FMA carry a statutory investor-protection mandate they cannot discharge by hand. The paper combines doctrinal legal analysis with expert interviews on all three sides and derives a multi-agent reference architecture that serves all three at once.

Greenwashing Risk Perception along the ESG Value Chain

A Qualitative Study at Investment Firms and Supervisors in Switzerland, Germany and Austria

Walter Kurz, Reinhard Magg

Swissi Institute for AI

The study examines how investment firms and supervisory authorities in Switzerland, Germany and Austria perceive greenwashing risk across the ESG value chain. The qualitative design covers both sides of the supervisory relationship, with attention to where risk perception diverges between the regulated entities and the bodies that supervise them.

Publication Without Documents

A Design Science Framework for Addressable and Verifiable Research Records under Machine Readership

Walter Kurz

Swissi Institute for AI

Scholarly communication was digitised rather than redesigned, and the artifact it settled on records how a document should look rather than what its parts are. That was defensible while readers were human, and machines now perform much of the reading, each one reconstructing independently and imperfectly the structure that existed when the author wrote it. The paper proposes a research record addressable at the granularity at which questions are answered and verifiable without refetching the document that contains it, and argues that the binding constraint on citation accuracy is not precision but affordability.