--- title: "Public sector AI: knowledge systems and assistants | YPAI" url: https://ypai.ai/solutions/public-sector/ description: "Knowledge systems and assistants for citizens and caseworkers, document AI for case intake, and the multilingual data that tests them. Built in Norway." source: "src/copy/routes (route /solutions/public-sector/)" --- # The counter closes at four. The questions don’t. > Knowledge systems and assistants for citizens and caseworkers, document AI for case intake, and the multilingual data that tests them. Built in Norway. YPAI builds the knowledge systems and assistants that answer citizens and caseworkers when the service desk is closed, and the multilingual data that tests them before release. Every answer cites the approved source it came from. A named person takes what the assistant may not decide. - **AI Implementation**: Knowledge systems, assistants, document AI - **AI Data & Evaluation**: Citizen languages, evaluation, native review - **Connected Delivery**: One team owns the system and its data - **Run in the EEA**: Private deployment, Norwegian company - Every answer cites the passage it came from. - A named caseworker takes what the assistant may not decide. ## An assistant answering for the public stands or falls on three questions. ### Can it say only what we have approved? - It retrieves from records you have approved, answers with citations, and when the evidence is not there it says so and stops. The answer inherits your access boundaries instead of bypassing them. ### Does it serve every citizen? - Only if the data does. Dialects, minority languages and unstructured questions are in scope from the start. Native speakers record and review, and the assistant is graded per language before release. ### Who answers when it cannot? - A named caseworker. The assistant performs only the bounded actions you intended; everything outside them is routed to a person who owns the exception, with the full exchange attached. ## Buy the system. Buy the data. Or both, from one team. ### Delivered into your environment. - A working knowledge system, citizen or caseworker assistant, or document workflow for a defined public service, with its integrations, permissions and controls. ### Tested in every language you serve. - The citizen questions, documents, evaluation sets and native-speaker review that train, test and prove the assistant before a citizen meets it. ### One team owns the failure. - The assistant and the data it answers from, built and fixed by the same people. Nothing is handed to another vendor. Operated from Norway, EEA residency by default. ## Citizens keep asking after four. The assistant answers from your approved records, or hands over to a person. A benefit change, a deadline, a document list, and a question it must refuse. The assistant answers only from records the service owner has approved, cites them on the answer itself, and routes what it may not decide to a named caseworker. Built for the service you run, inside the access you already enforce. ### Drawn from approved records - Assembled from records the service owner has approved, each one cited on the answer itself. ### The thread stays attached - Every sentence keeps its line back to the passage it came from. A reviewer can open it tomorrow. ### A person takes what it may not decide - Out of scope is routed, not improvised. A named caseworker receives the full exchange. ### I got a letter about my housing benefit changing next month. What does it mean, and what do I need to do? The letter follows a rates-table change. Nothing is required from you unless your housing costs also changed; the letter’s step 2 covers that case. - Housing benefit guide · §4.2 · v7 · approved - Rates table · 2026 · v3 · approved ### Can you make an exception for my deadline? My situation is complicated. That is a decision, not a lookup. The assistant does not decide, and it does not improvise an answer it cannot cite. Routed · named caseworker · full exchange attached In the morning queue. The citizen is told who has it. ### Which documents do I need to send with my application? Three: proof of income, a copy of your tenancy agreement, and last year’s tax assessment. The checklist lists the accepted formats. - Application checklist · v4 · approved ### What is the status of my neighbour’s application? That is not yours to access. The assistant answers inside your own access boundaries, and this request has been logged. access denied · permission-aware retrieval · attempt logged ## Caseworkers decide on the current rule. Post is read on arrival. Every language is tested before release. Three deliverables a directorate or municipality buys from YPAI: a governed knowledge system for the people who decide, document AI for the applications and letters that arrive, and the evaluation that tests the assistant in every language your citizens speak. Each is delivered into your environment and can be bought on its own or together. ### The caseworker knowledge system - Precedent, procedure and the current version of the rule, inside the access the caseworker already has. - A governed library of circulars, guides and decisions with retrieval that returns the passage, its version and its approval state, permission-aware. - Your documents ingested with version and approval state on every record; superseded versions retired, never cited. ### Document AI for what arrives - Applications, attachments and letters read on arrival, classified, and routed with review where judgement is required. - Extraction and classification integrated with the case system, approval gates at every bounded action, a review queue for exceptions. - Annotated applications and attachments from your own casework: extraction ground truth and exception taxonomies. ### Tested per language before release - The service that works in the capital’s dialect, tested in the others before a citizen meets it. - An evaluation set built from the questions citizens actually ask, graded by native speakers per language and question class, re-run on every source update. - Speech and text recorded and reviewed by native speakers; where a language fails, data is collected and the class is re-graded. - When does a benefit change take effect? - What counts as proof of income? From the first day of the month after the rates-table revision. No action is required unless housing costs also changed. A current payslip, a tax assessment for the last year, or a statement from the paying agency; the checklist lists accepted formats. - A change that follows a rates-table revision takes effect from the first day of the following month. - Monthly rates by household size, effective 1 May. - Proof of income: payslip, tax assessment, or a statement from the paying agency. - Income is assessed on documented amounts for the assessment period. Inside the caseworker’s access · permission-aware Application · Housing benefit · received by post - **Benefit**: ok. Housing benefit - **Period**: ok. From 1 May - **Tenancy agreement**: ok. Attached - **Tax assessment**: ok. Attached - **Proof of income**: flag. Missing Approval gate before any letter is sent Arabic · deadlines · data collected · re-graded Released · graded in every language ## Build the system. Knowledge systems, citizen and caseworker assistants, document AI and case workflows, delivered into your environment. - [Enterprise knowledge assistants](https://ypai.ai/enterprise-knowledge-assistants/) - [AI assistants and chatbots](https://ypai.ai/ai-assistants-chatbots/) - [Document AI workflows](https://ypai.ai/document-ai-workflow-automation/) ## Build the data it runs on. Citizen questions, documents, evaluation sets and native-speaker review, traceable through the assistant’s lifecycle. The AI Data & Evaluation service line - [Speech data](https://ypai.ai/speech-data/) - [Data collection](https://ypai.ai/data-collection/) - [Data annotation](https://ypai.ai/ai-data-annotation/) - [Evaluation](https://ypai.ai/data-solutions/evaluation/) ## When the failure sits in the data, the team that built the assistant fixes the data. A circular missing from the library, a dialect the evaluation set never covered, a question class nobody graded. No hand-off between a system integrator, a model vendor, an annotation company and an evaluation supplier. Each service line remains independently purchasable. ## Serving every citizen is a data problem. We are also the data company. ### Nordic depth ### European ### Beyond Europe - the languages your citizens arrive with ### Code-switching Recorded and reviewed, not machine-translated coverage. The exact language list is set with the engagement. ## Built and run from Norway, under the rules you already answer to. YPAI is a Norwegian company operating under GDPR. The assistant runs in your environment, inside the access boundaries you already enforce. - **Jurisdiction**: Norway, under GDPR - **Data residency**: EEA by default, defined per project - **Deployment**: Private and EEA, in your environment - **Processing terms**: Standard DPA, subprocessors and transfers defined per project - **Erasure**: 30-day end-of-contract erasure - **EU AI Act**: Article 10 documentation, produced as part of delivery ## What we tell agencies before they start. ### [Data residency vs sovereignty: why GDPR is not enough](https://ypai.ai/blog/compliance/eu-speech-data-sovereignty-gdpr-not-enough/) ### [CTO’s guide to sovereign AI: architecture and costs](https://ypai.ai/blog/infrastructure/ctos-guide-sovereign-ai-architecture-costs/) ### [EU AI Act Article 10: data governance checklist](https://ypai.ai/blog/compliance/eu-ai-act-article-10-data-governance/) ## Bring us the service that has to answer after four. Start with a service or journey, the documents that should govern its answers, and the languages your citizens speak. YPAI will define the shortest credible path from the current state to an assistant you can hold accountable. Pilots are scoped to your specification. Scope and commercial terms are agreed per project before the pilot begins, every pilot carries a SOW, acceptance criteria and a DPA, and production is a separate decision after pilot review and acceptance. Enquiry details are treated as confidential. One brief. A project lead reads it, not a queue. A named project lead reads this and comes back to you with the first scoping questions. ## Frequently asked questions The questions a service owner asks before an assistant answers in the agency’s name. ### Can the assistant answer from documents we have not approved? No. Retrieval is limited to records with an approval state you control. When the evidence is not in the library, the assistant says so and stops, or routes the question to a named caseworker. Every answer carries its citations, so a reviewer can open the record behind it. ### Who decides when a question is outside the assistant’s scope? A person. The assistant performs only the bounded actions agreed in scoping. Anything outside them is routed to a named caseworker with the full exchange attached, and the citizen is told who has the case. The assistant never improvises a decision. ### How do we know it works in the languages our citizens actually speak? It is graded per language before release, against an evaluation set built from the questions citizens actually ask. Speech and text are recorded and reviewed by native speakers, and the evaluation re-runs on every source update. ### Where is the data processed? In the EEA by default, with private deployment in your environment where the engagement requires it. Residency, subprocessors and transfer controls are defined per project in the DPA. YPAI is a Norwegian company operating under GDPR.