Solution
Secure AI assistants
AI use may already exist, with or without a decision behind it. The decision covers the permitted tasks, the data, the product and its contract, access, retention, training use, transfers, logs and how answers get validated.
The gains from the assistant
Answers expose verifiable sources
When an answer uses the corpus, it links to the passages it retrieved. Citation makes checking possible; it proves neither the source nor the conclusion.
Processing becomes a documented decision
Local, private-cloud and managed options get compared against real flows and the product terms. No-training use and retention are separate properties: you configure and verify each on its own.
The tool inherits your permissions
Permissions get filtered at indexing and at retrieval, then tested profile by profile. Connectors, caches, logs and configuration errors stay in the threat model.
A cost you can predict
The model and the architecture get selected on quality, latency, volume, operations and the full cost observed during the pilot, not on a leaderboard or an invented price norm.
In practice
- Search across your documents with the exact passage quoted, rather than an unverifiable answer
- Access rights taken from your directory and applied to every question asked
- Self-hosted open models or frontier models in a Swiss region, depending on how sensitive the corpus is
- Conversation history, technical logs and retention configured separately for the product, operating need and obligations
- Explicit guardrails: what the assistant refuses to do, and what it hands back to a person
- Proportionate usage logging: categories, failures, cost and latency without retaining content or identity beyond the declared need
Systems involved
- SharePoint, OneDrive, Nextcloud and internal file shares
- Microsoft 365 and Google Workspace for identity and permissions
- Document management and line-of-business software over APIs
- Knowledge bases in versioned Markdown
- Mail and internal chat tools
Service lineArtificial intelligence →
Which documents
The same work, against each sector’s own constraints. Every card opens the full sector.
Banking and insurance
An assistant grounded in the normative corpus
Directives, credit policies, product terms and circulars queryable in plain language, with the exact passage cited and directory permissions applied to every answer.
Energy and utilities
Assistant on technical documentation
Diagrams, procedures and equipment sheets queryable from the field, with the source cited.
Healthcare and life sciences
Assistant on confidential documents
Open models served from your own machines, with directory permissions applied to every question.
Manufacturing
Queryable technical documentation
Drawings, routings and procedures reachable from the floor in plain language, with the passage quoted.
Retail and e-commerce
Internal assistant for store teams
Product sheets, return conditions and till procedures queryable in plain language, with the passage quoted on every answer.
How it runs
Corpus
We choose what the assistant may read, and above all what it won’t. A narrow clean scope beats a broad doubtful one.
Prototype
A first version on a real subset, judged on questions you already ask. That’s where you find out whether the corpus holds up.
Rollout
Permissions wired to the directory, interface installed, guardrails in place, and a session for the team that will use it.
Upkeep
Corpora age and questions shift. We read the log, fill what is missing, remove what goes unused.
To go deeper
The EU AI Act now applies: four questions for a Swiss company
A Swiss address does not settle whether the EU AI Act applies. Start with the company’s role, the EU connection, the effect on people and the evidence needed before production.
Kimi K3, Soofi S and Apertus: an open-model snapshot
A dated snapshot, not a permanent ranking: release status, weight precision, internal benchmarks, total cost and deployment conditions.
Internal AI without a US cloud: realistic in 2026?
Open models, Swiss hosting and local deployment: a dated method for selecting internal AI without reducing sovereignty to the server's location.
Test the fit: Secure AI assistants
Describe the context, constraints and decision you need to make. The first conversation qualifies scope, boundaries and the next useful step.
Describe the situation