Roadmap

Related: Changelog · Blog · Integration research framework

created_date: 2026-01-15 last_modified_date: 2026-09-26 last_modified_summary: Rewritten for the person-centred product and the people who fund and run this work. Investor framing removed; pilot targets are stated as hypotheses the pilot tests; language status corrected (6 offered, 3 speaker-reviewed).

What this is for

People who arrive in Switzerland through the asylum system spend months or years in accommodation run by an operator on behalf of a city, a canton or the Confederation. What happens there — whether a house is calm, whether someone gets an answer, whether they take a first step into language and work — shapes how the rest of their life here goes, and what it costs everyone.

This product is the everyday layer of that work: who is waiting for an answer, how each person is doing, where they could go next, who lives with whom, and how a house settles its conflicts. It sits beside an operator's case-management system and gives residents a portal in their own language.

Not the goal: occupancy, throughput or fewer staff. The goal is that the time staff have goes to people rather than to searching and firefighting — and that the difference can be measured.

How we will know whether it helps

Measured in a pilot, against a baseline, not asserted:

Question	Measure	What the product can do about it
Are houses calmer?	Incidents per 100 resident-months; conflict-related moves	Influence (matching, conflict ladder, house rules)
Do people get answers?	Median time from request to first staff answer	Influence (Eingang)
Do people take steps?	Documented language and work/volunteering steps per quarter	Contribute
Are staff relieved?	Minutes spent on documentation and searching, sampled	Influence
Do residents feel heard?	Short anonymous survey in their language	Contribute

The targets written into earlier versions of this page (−30 % incidents, −50 % relocations, −40 % mediation hours) are hypotheses, not results. A pilot that finds no improvement is a valid result and will be reported as one.

Now (to end of 2026)

Done:

• Person-centred staff navigation: Eingang, Klient*innen, Einsatzplätze & Angebote, Wohnen, Zusammenleben.
• Eingang: every request a resident raised reaches everyone who may answer it, with a one-press "Übernehmen".
• Einsatzplätze: AI-assisted entry from a pasted advert (permit route and publishing stay human), one catalogue with activities, one-tap interest in the portal.
• Household fit notes instead of per-person problem flags; no religion or origin proxies in matching.
• Resident portal offered in 6 languages: German, English and French reviewed by speakers; Ukrainian, Russian and Arabic complete and marked as machine-assisted. Farsi/Dari, Tigrinya, Turkish, Albanian and Somali are in progress and not offered until complete.
• A no-account demo on its own instance with invented data.
• AI use disclosed at /ki-datenschutz and wherever staff meet it.

Next:

• Pilot measurement tooling: baseline import, staff time sampling, an anonymous resident survey, aggregate-only reporting (no cell below six people).
• Human review of the portal languages most asylum seekers read — Tigrinya and Dari first.
• Impressum and privacy statement; a subprocessor list.

Later (2027)

• One installation serving several operators, each with its own data, rule book and settings — needed before a second organisation uses it.
• Hosting in Switzerland and an AI model under a data-processing agreement (or AI switched off) before real data is processed in regular operation.
• CSV exchange with the case-management systems operators already use.
• Aggregated indicators aligned with the Integrationsagenda Schweiz, for the city, canton and integration offices — never personal data.

Engineering invariants

1. Every score is explainable; the product recommends, a person decides.
2. Functional needs only — never diagnoses, asylum-procedure details, religion, politics or origin.
3. No verdicts about individuals from counts; aggregates only for reporting.
4. Money is integers, balances are computed over full history, votes snapshot their policy.
5. Safety is never put to a vote, and safety conflicts never start at the bottom of the ladder.