# Knowledge & assistant: AI answers with sources > Your company memory with an AI in front of it: it answers questions from wiki, documents, tasks, requests, meetings, customers, invoices and time tracking and names the source for every answer. Behind it is a RAG system that looks up the right passages before every answer. The AI sees only what the asking person may see, runs in the EU and can be switched off per organization. - Page: https://www.ciro-noise.de/funktionen/assistenz - Deutsche Fassung: https://www.ciro-noise.de/funktionen/assistenz.md When someone on your team asks “How do we submit travel expenses?”, the AI in the sidebar of every page answers. It reads the wiki, documents, tasks, requests, meetings, customers, invoices and time tracking, links the source and shows only what the person asking is allowed to see. ## How the AI works in your company The chat in the sidebar of every page is your colleague, and she has her own page for longer conversations. She reads the wiki, the tasks, the requests, the meetings, the customer records, the invoices and the approvals in time tracking. Every answer names its source with a link: the wiki page, the task, the request or the meeting. When Isa finds nothing on a question in your data, she says so instead of guessing. The AI sees exactly what the person asking sees. Anyone without access to finance gets no revenue figures from the chat, and confidential wiki pages stay out unless the person has permission. The same goes for outside AI tools like Claude or ChatGPT and for your own software. They get an access key with exactly the permissions of the person who created it, and they can read and create with it. Only a person deletes wiki pages, tasks, invoices or customer records. Your own software connects through the API with its own key. The AI creates pitch drafts, suggests replies to requests and sums up meetings with their action items. A person makes every decision: a vacation request shows up in the chat as a button, and someone on your team presses it. The AI knows wiki pages and meetings on its own and also learns from curated answers like an FAQ, resolved requests and your public pages. Search finds the page even when it says “sign-off” and you ask about “approval”, in German and in English. Under the hood this is a RAG system (retrieval-augmented generation). Before the AI answers, CIRO NOISE looks up the matching passages in your company knowledge, and the AI writes its answer from exactly those. To do that, CIRO NOISE splits every wiki page, every document and every page of your website into sections at its headings. Search runs two ways, by meaning and by the exact words, so an invoice number or a contract clause is found too. The link below the answer leads to the section it came from, and new or changed content is findable within a few minutes. Files are part of your knowledge too: PDFs, including scanned ones, and Word files from your file storage and from attachments on tasks, requests and wiki pages, plus photos of documents in your file storage. CIRO NOISE reads each file once and creates a profile with its type, a summary, the parties involved, key dates and deadlines. That lets the AI answer “Which contracts expire this year?”. For a deadline from a document, CIRO NOISE reminds the person responsible 30, 7 and 1 day ahead and on the day itself. You can add up to five websites as sources, your own for example, and CIRO NOISE reads up to 100 pages per site and again every week. Whoever maintains the knowledge sees on the “Quality” page which questions went unanswered and which answers your team marked as not helpful. Each of those questions becomes an FAQ entry with one click. CIRO NOISE turns resolved support requests into drafts without personal data, and a draft only becomes searchable once someone approves it. The AI models run in the EU, and your content never trains a model. Each organization has a budget for AI use and a page that shows how much has been used. If you don't want the AI, switch it off for your whole organization, and wiki, tasks and search keep working without it. To check whether an answer is right, click the source below it. ## Answers with a source, within your permissions The AI knows your company and follows the same permissions as your team. - **Check the answer with one click**: Every answer names the wiki page, task, request or meeting it came from, and one click opens the source. When there is nothing on a question in your data, the AI says so. - **Revenue figures only for those allowed to see them**: The AI sees exactly what the person asking sees. Confidential wiki pages stay out, and anyone without access to finance gets no figures from the chat. - **Reads contracts, PDFs and your website too**: Documents from your file storage and attachments and up to five websites count as sources. Search finds a passage even when it says “sign-off” and you ask about “approval”. - **Suggests, a person decides**: The AI creates pitch drafts, suggests replies to requests and sums up meetings. Approvals show up as a button in the chat, and a person presses it. - **Runs in the EU and can be switched off**: The models run in the EU, and your content is never used for training. Each organization has a budget, a usage report and a switch that turns the AI off completely. ## Three questions that come up every day The new hire, the customer call and the founder's Friday afternoon. ### Nele asks: “How do we issue a reverse-charge invoice?” Nele started three days ago and doesn't want to pull anyone away from their desk. Isa answers from the wiki page “Invoices to other EU countries” and links it. Nele can see the page because she has read access to the “Finance basics” area. The confidential page “Terms for key accounts” stays out of the answer, because Nele doesn't have that permission. She reads the source and issues the invoice herself. - Source linked, one click to read it - The permissions of the person asking apply - Confidential pages stay out ### Lena asks: “What did we promise Brandt in the spring?” Ms. Sommer is on the line, and Lena needs the answer now. She asks the AI in the sidebar, and the answer comes from the March 4 meeting summary and the customer record, with a link to both. Still on the call, Lena asks whether there is already a task for it. The AI names BRA-53 on the customer board with its deadline and owner, and Lena opens it with one click. - Meetings and customer record as sources - Every answer links to the place - Questions in German or English ### Kira asks: “Who is out next week, and what is waiting for my approval?” Isa answers from time tracking: Tom is on vacation Monday to Wednesday, Mara on Friday, everyone else is in. Two requests are waiting for Kira, and neither overlaps with anyone else on the team. The requests appear as buttons in the chat, and Kira clicks “approve” herself. Lena and Pascal get the decision right away, and the team calendar is up to date two minutes after the question. - Approval as a button, a human clicks - Team absences from time tracking - Overlaps checked before approval ## Knowledge & assistant in detail The key AI assistant features in one sentence each. - **Answers with sources**: Every answer links the page, task, request or meeting it came from. - **RAG across all your company knowledge**: Before every answer, CIRO NOISE looks up the matching sections from wiki, documents, requests, meetings, tasks and websites, by meaning and by exact words. - **AI with the person’s permissions**: The AI sees only what the person sees. Confidential wiki pages stay out unless the person has the permission. - **Documents with a profile**: PDFs, including scanned ones, Word and text files, and photos of documents in your file storage get a profile with type, summary, parties, dates and deadlines. - **Reminders for deadlines in documents**: When a contract contains a deadline, CIRO NOISE reminds the person responsible 30, 7 and 1 day ahead and on the day itself. - **AI chat on every page**: The AI sits in the sidebar of every page and has its own page for longer conversations. - **Answers from wiki, tasks and customers**: Wiki, documents, tasks, requests, meetings, customers, invoices, websites and the approvals in time tracking. - **Your website as a source**: Up to five websites with up to 100 pages each, read again every week. CIRO NOISE follows the site’s robots.txt. - **Unanswered questions**: The “Quality” page shows which questions went unanswered and which answers didn’t help. One click turns them into an FAQ entry. - **FAQ and curated answers**: Curated answers like an FAQ, resolved requests and public pages count as sources, and the AI knows wiki pages and meetings on its own. Search finds a page even when it uses different words. - **AI in the EU, no training**: The models run in the EU. Your content trains no AI models. - **Drafts and suggestions**: The AI writes pitch drafts, suggests replies to requests and summarizes meetings. Approvals show up as a button that a person presses. - **Outside AI tools**: Claude, ChatGPT and your own software get an access key with the permissions of the person who created it. They read and create, but only a person deletes wiki pages, tasks, invoices or customer records. Your own software connects through the API with its own key. - **Budget and switch**: Each organization has an AI budget, sees usage on one page and can switch the AI off completely. ## Tell Claude. CIRO NOISE records it. ### Question with a source > How long is our notice period for maintenance contracts, and where is that written down? The notice period is three months to the end of a quarter, and notice has to be in writing. So to cancel as of Dec 31, it has to arrive by Sep 30. It's written down in two places: • in the wiki under Contracts › “Maintenance contract: terms”, section 4 • in Maintenance_contract_template_2026.pdf, Section 9 Term Support gave the same period in May and July. ### Put together onboarding > Lena starts in support on Monday. Pull together from the wiki and knowledge what she should read in her first week, and create a page for her. Lena's page “Welcome, Lena: your first week” is ready. She reads in this order: • Mon–Tue: support basics, escalation and tone in replies • Wed–Thu: the 20 most common customer questions with answers, and the complaint process • Fri: the product training as a PDF, then a Q&A with Jonas Jonas Brandt helps her with support questions and Petra Lang with anything finance. With SUP-14 she checks off the week by Friday. ### Look up a rule > Can we copy customer data into a test system? What does our policy say, and what did we tell the last customer who asked? Test systems get pseudonymized data. Real customer data may only go in with the data protection officer's approval and must be deleted after 30 days. For Winter GmbH, Section 4 of the data processing agreement rules out testing without pseudonymization altogether. Support promised the last customer on May 14: “tests only with made-up data”. That matches the policy. Sources: • Wiki › Data protection › “Test systems” • Data processing agreement Winter GmbH, Section 4 ## Frequently asked questions ### What can I ask the AI? Anything that's in CIRO NOISE, like “How do we submit travel expenses?” or “What did we promise Brandt?”. The AI searches the wiki, tasks, requests, meetings, customers, invoices and time tracking approvals and answers in the sidebar chat or on its own page. ### Is the AI assistant a RAG system? Yes. RAG stands for retrieval-augmented generation: before every answer, CIRO NOISE looks up the matching sections from your wiki, documents, requests, meetings and websites, and the AI writes the answer from those sections. That is why every answer shows its source, and when search finds nothing, the AI says so. ### How do I know where the answer comes from? Every answer names its sources with a link: the page, task or request it came from. One click and you can read it yourself. When the AI finds nothing on a question in your data, it says so instead of guessing. ### Which documents does the AI read? PDFs, including scanned ones, Word files and text files up to 15 MB, plus photos of documents in your file storage. They can sit in your file storage or be attached to a task, a request or a wiki page. Each file gets a profile, and an answer only shows a file to people who may open it anyway. ### Does the AI respect our permissions? Yes. The AI answers with the permissions of the person asking: management gets finance figures, the team gets what it may see anyway. Confidential pages stay confidential, in the answer too. ### Where does the AI run? In the EU. Your data sits on servers in the EU, the AI runs on models in EU regions, and your content trains no model. You switch the AI on or off per organization. ### Can I use Claude with my company's knowledge? Yes. Claude connects to your organization via MCP, the open standard for AI tools. There Claude searches knowledge and files and reads wiki pages, meetings and deals, with the permissions of the person who grants access. You use Claude with your company's knowledge without copying anything. ### What does the AI assistant cost? It's included in the price per seat: the first two seats are free, and each additional seat is €25 a month plus VAT, all features included. Each organization has an AI budget, and you see its usage on one page. ## CIRO NOISE at a glance - CIRO NOISE is a workspace for small businesses and teams of up to about 50 people: wiki, tasks, customer requests, finance with quotes and invoices, customers (CRM), meetings, knowledge with AI, social media, password vault, and time tracking behind one login. - Price: the first 2 seats are free, with every feature. Each additional seat costs €25 per month plus VAT. No credit card to start. - Privacy: servers and AI run in EU data centers, content never trains AI models, and permissions apply per person and feature. - AI connection: Claude and other AI assistants connect through MCP, the open standard for AI tools, with exactly the permissions of the person. - Provider: CIRO GmbH, Eching, Germany. Contact: hello@ciro.gmbh, phone +49 89 24 88 31 22.