AI Company Wiki for SMEs: Use Your Knowledge Securely

Friday, 4:40 pm, at a property management firm in the canton of Aargau. A tenant calls: the heating at the Bahnhofstrasse building has failed. The property manager in charge is on holiday. Who's the heating contractor? Is there a maintenance contract with an on-call service? Where's the key to the boiler room? The answers exist. They're in an email from 2023, in a folder on the server, and in the head of a colleague who is currently lying on a beach.

Almost every SME knows moments like this. It has little to do with poor organisation. In small companies, knowledge grows informally. That works as long as everyone sits in the same office. As soon as someone is off sick, leaves or starts new, it gets expensive.

Three trends are making the problem worse right now:

• Skills shortages and staff turnover. If you need to get new people productive faster, you need knowledge that's written down.
• Retirements. Many SMEs will soon lose people who have “just known how things work” for twenty years.
• Shadow AI. Employees paste client figures, tenant correspondence or patient questions into private AI chatbots because they want fast answers. Often the company has no idea.

That last one is the most dangerous, because it stays invisible until something goes wrong.

In property management, knowledge is tied to the building

A property manager often looks after well over a hundred rental units. For every building there are things no property management software records properly: who the caretaker is, which plumber also comes at weekends, what the laundry room rules are, where the key box hangs, which owner wants a call before any repair over CHF 2,000.

Nobody notices as long as the person in charge is around. Then she's off sick, and that's the day a water leak comes in. Or a new property manager takes over 150 units and needs six months to learn their quirks. On top of that come the tenant questions that sound the same every week: rent adjustments after a reference rate change, service charge statements, notice dates, defect reports.

An AI company wiki answers two kinds of questions here: property-specific ones (“Who's the caretaker at Bahnhofstrasse 12?”) and process questions (“How do we handle a defect claim?”). Both come with a link to the source and the name of the person responsible for that page.

In accounting and fiduciary firms, it is tied to clients and deadlines

In an accounting firm, a lot of knowledge sits in the details. Which client files VAT using the flat-rate method? How does a client want receipts filed? Which expense policy applies at Mr X's limited company? Which cantonal specifics apply to the tax return?

During year-end season, junior staff ask their experienced colleagues twenty times a day. Each question takes two minutes, but it pulls a specialist out of deep work. When a long-standing accountant leaves, part of the client knowledge leaves with her.

Then there's the data protection risk. Under the revised Swiss Federal Act on Data Protection, Art. 62 FADP makes it a criminal offence to intentionally disclose secret personal data learned while practising a profession. It applies not just to doctors and lawyers but to every profession that requires knowledge of such data. The fine can reach CHF 250,000 and is normally imposed on the responsible individual, not the company.[2] When an employee pastes a client's balance sheet into a US chatbot, that's not a minor slip.

In medical practices, it is tied to procedures

Medical practices see high turnover among medical assistants. Every new assistant has to learn hygiene plans, equipment reprocessing, phone rules, emergency procedures, billing rules and the quirks of the practice software. Usually that happens by shadowing and asking. The doctor ends up as the help desk for everything.

An AI company wiki helps with the internal procedures: “How do we reprocess the ultrasound machine after an exam?” or “Which appointments can I book without checking first?” Patient data explicitly does not belong in the wiki. Professional secrecy under Art. 321 of the Swiss Criminal Code applies to it, which is exactly why an on-site setup makes sense here, where even the language model never leaves the practice.[3]

Where the data lives. Every client gets their own instance. It runs either on a server in a Swiss data centre or on hardware at your premises. There's no shared database with other clients.

Where questions are processed. We use open-source language models. In the standard setup the model runs at a data centre operator in Switzerland or the EU, under a data processing agreement. In the premium setup it runs on a dedicated machine at your site, so neither content nor questions ever leave your building. That's the setup we recommend for medical practices.

Who can see what. Read permissions apply to pages and to the chat. Search filters by permission before the AI gets to see anything at all. Access is logged.

Which data doesn't belong in there. The wiki holds processes, property details and expertise, not client files, tenant records or patient data. We agree this rule with you in writing and train your team on it. Data minimisation is the most effective data protection there is.

Contracts and documentation. RakenAI processes data on your behalf. We sign a data processing agreement under Art. 9 FADP[2] and Art. 28 GDPR and document the technical and organisational measures: encryption, backups, access control, update process.

Training. For companies in the EU, Art. 4 of the EU AI Act requires providers and deployers of AI systems to take measures that support the AI literacy of their staff.[5] The amended article does not prescribe a specific level. Our onboarding training covers the basics: what the assistant can do, where its limits are, and what must never go in.

No software vendor can sell you data protection compliance. A system becomes compliant through its architecture, contracts, rules, and people who follow them. We deliver the architecture and the contracts, help with the rules, and coordinate with your data protection adviser where needed. This article isn't legal advice. How we handle patient data specifically is covered in Privacy-first AI.

1. First conversation. We listen: which questions cost your team the most time today? Who gets interrupted constantly? Where is your data now?
2. Analysis and concept. We decide which area to start with, which hosting option fits and which roles exist. We measure the current state, for example the number of internal questions per week, so we can prove the benefit later.
3. Capturing knowledge. This is where most of the work is. We run interviews, go through existing documents and record voice messages. That produces draft pages your specialists review and approve. For property managers and accounting firms we bring a ready-made base structure.
4. Setup and testing. We install the system on your server, connect sign-in and test it with real questions from your daily work. We only move on once the answers are right.
5. Training and launch. A short session for all staff, a deeper one for page owners.
6. Operation and upkeep. We take care of updates, backups and monitoring. You don't need your own IT department.

We recommend starting with one area, such as property management or year-end closing. Once that works, the wiki grows step by step.

Every page has an owner and a review date. When the date passes, the owner gets a reminder with a direct link. The assistant collects questions it couldn't answer. That list shows every month where knowledge is missing, and it's the best to-do list for upkeep there is. Additions come in by voice message, so nobody has to block an afternoon for one page.

If you like, we'll also handle the editorial side: you send a voice message or a document, and we turn it into the page.