Manifesto · AI at Visamundi
Everyone "does AI". We prefer to list precisely where we use it, on which data, and where we've explicitly decided not to use it. A single model: Mistral, open-source, self-hosted on our own infrastructure in Europe. A living document, updated when our practices change.
Our mindset
Every task follows the same path: data trimmed to the useful minimum, a model we control, a proposal, then human validation before any action.
Document or email, trimmed to the useful minimum.
Open-source, self-hosted in Europe.
Summary, draft or selector, never a decision.
An operator approves before any action.
Principles
Before the AI feature list, here's what constrains it.
No model accepts or rejects a file. No model classifies a traveller. Visa decisions remain those of consular authorities; ours remain those of human operators.
We don't use AI to score, profile or target travellers. None of our prompts contains the names, addresses or photos of the people we accompany.
We use a single model: Mistral, open-source, self-hosted on our own infrastructure in Europe. No data passes through a third-party API, and nothing trains a public model.
Every generated draft (summary, reply, selector) goes through a human operator before any action. AI proposes, the team decides.
Real uses
Four features. Not one more. If we add one tomorrow, it will appear here before going to production.
Reads passports, tickets and supporting documents to automatically file them in the correct slot.
Generates 3-4 sentences so an operator can identify the state of a file, next steps, and attention points in 5 seconds.
Proposes a draft reply to an incoming email. Always reviewed and sent by a human.
When a government portal changes its forms without warning, the LLM helps find the right HTML field to avoid blocking a file.
Classifies an email as a travel confirmation or not, then extracts destination/dates/reference for ambiguous emails that regex alone can't parse. Bilingual FR/EN.
Red lines
The uses we explicitly forbid ourselves. This list matters more than the one next to it.
Governance
This isn't a frozen marketing page: here's exactly how this document stays alive and who owns it.
It describes the real state of our AI stack. The moment a practice changes, this document changes with it.
Every new AI feature is documented here (model, data sent, human oversight) before it ships.
A question about our AI practices or your data? Our DPO answers you directly.
