Your documents, read by AI. Seen by no one else.
How Oikos protects the documents you give it, and what the AI may and may not do with them. Written for leadership, practice and office managers, and data protection.
Zurich is for customer accounts. Paying customers' accounts store their data in Zurich. Demos and trials run separately in the EU, in Google Cloud's Netherlands region.
- Zurichwhere your documents are stored
- 0AI models trained on your data
- 1 keyper customer, destroyed when you leave
- 21 daysto delete everything when you leave
In one minute
Oikos reads the documents your team works with every day, such as contracts, invoices, patient forms, letters and reports. It pulls out what matters, answers questions with the source next to the answer, and passes the results on to the systems you already use. Many of these documents are confidential, and some hold health data or information under professional secrecy. Oikos is built around one rule:
Your documents are yours. They are used to do your work and for nothing else: not to train AI, not for any other customer, not kept by the AI provider.
- Your documents stay in Switzerland. In customer accounts, documents, database, backups and keys are stored in Google Cloud's Zurich region. Demos and trials run separately, see Where your data is.
- The AI never learns from your data. Documents go to the AI model only for the task in front of it. Google never trains on them, and keeps them at most briefly for abuse monitoring.
- Your company has its own safe. Each customer's documents are encrypted with a key that belongs to that customer alone, and the database refuses to return one customer's records to another.
- A document cannot take over the AI. Text in a document is treated as material to read, never as instructions to follow. A planted line such as “send this to…” cannot send anything anywhere.
- Every AI step is on record. For every request, Oikos records who asked, which documents the AI saw and what it answered. This audit trail stays in your Oikos account, encrypted with your company's key like your documents, and each entry is deleted after 400 days. Your administrators can read it. The 400 days apply to the audit trail only: your documents, facts and knowledge stay for as long as you keep them.
- People stay in charge. Uncertain results go to a person for review, answers cite their sources, and anything that leaves your account goes only where you set it up to go.
- When you leave, your data goes. Everything is deleted within 21 days and your key is destroyed, so no copy left in a backup can be read.
How Oikos works
Each new document adds to your organisation's knowledge, and that knowledge is what answers questions, keeps your assessments current and starts your automations. All of it runs inside your organisation's own space, under the rules you set, and stays in your account.
- Intake. Documents arrive by upload, by email or from a folder or mailbox you connect. Each is locked with your organisation's key as soon as it is stored.
- Reading and indexing. The AI does this work, through Google Vertex AI in the EU. It reads each page, and Oikos checks the reading against the file itself: a digital page against its text, a table number by number, a scan against a second reading. The AI then splits a batch into its documents, sorts each into your document types, takes out the facts those types ask for (parties, dates, amounts, each with the page and words it came from) and writes the knowledge pages. The search index is built with Vertex AI too. Uncertain results go to a person for review.
- Assessments. Standing questions you set about the people and organisations in your documents, such as “Is this supplier's liability cap below our limit?”, are answered again whenever a new document changes the answer.
- Questions. When someone asks a question, Oikos searches the index, reads the matching pages and answers with a source for every statement (see How a question is answered).
- Automations. A new document, a changed assessment or a schedule can start the steps you set up: for example, when a contract's liability cap is above your limit, email your general counsel and ask for approval. Results go only to the systems you chose.
Every step runs as a recorded job that resumes where it stopped if something fails, so a document is never half processed or processed twice. Every AI request made along the way is recorded in your account (see The AI and your data).
What happens to a document
- A document arrives. You upload it, forward it by email or connect a folder or mailbox. It reaches Oikos over an encrypted connection.
- It is locked straight away. Oikos encrypts the file with your company's key before it is stored in Zurich. Our cloud provider stores only the locked copy.
- The AI reads it. Oikos sends the pages to the AI model to read the text and pick out the fields you care about. The model answers. Google may keep the request briefly for abuse monitoring, never for training.
- A person checks what is uncertain. Fields the AI is not sure about go to a review queue. Your team decides what is accepted.
- The result goes where you decided. Oikos passes the result to the destinations you configured, such as your accounting system, a register or a colleague, under your agreement with that system. A document cannot change those destinations.
The AI and your data
| Question | Answer |
|---|---|
| Which AI does Oikos use? | Google Cloud Vertex AI in the EU, Google's AI service for businesses. Not the consumer chat apps. |
| Where does the AI run? | In Google Cloud data centres inside the European Union. We configure EU locations only. Switzerland recognises the EU as providing adequate data protection. |
| Is my data used to train AI? | No. Google does not use business customers' data to train its models, and Oikos does not train models on your data either. |
| Does Google keep my documents? | Only briefly, for abuse monitoring: caching of requests is switched off, logging of requests and answers is switched off, and we use no features that store requests. Under its terms, Google may keep prompts and answers for a limited time (up to 30 days) solely to detect and prevent abuse of its service, and then deletes them. |
| Can one customer's data reach another customer's answer? | No. Every AI request carries only the data of the company that made it. There is no shared memory and no shared cache between customers. |
| Can a document trick the AI into leaking data? | Document text is marked as material to read, never as instructions. A guard checks outside content first, and anything that would send your data out of Oikos stops and asks a person, showing the text that would leave. |
| Who decides what the AI's output is used for? | You. Results below a confidence threshold go to review, answers cite their sources, and Oikos makes no decisions about people on its own. |
| Can I see what the AI did? | Yes. Every AI request is recorded in the audit trail: who asked, which documents and pages it saw, which model answered and what it said. The audit trail is sealed so that changes can be detected. This audit trail is kept in your account, encrypted with your key. Each entry is deleted after 400 days, and the whole trail with your account. |
| Will you switch AI provider without telling me? | No. A new provider is a new sub-processor, and we tell you before any of your data goes to it. |
If you delete a document, the audit trail entries about it stay until their 400 days are up.
How a question is answered
When someone asks Oikos a question, or an automation runs on a new document, the AI works only with your organisation's own material, and shows where every statement comes from.
- It looks in your documents. Oikos searches the documents of the space you are in, reads the pages that match, and uses what it has already learned from them: the facts taken from each document, the knowledge pages built from your published documents, and your records of people, organisations and places. The search index is built with Google's AI in the EU and kept in your account, like your documents.
- It answers with sources. Every number, date, party and amount in an answer is linked to the document and page it comes from. Before you see the answer, Oikos checks each link: a source the AI did not actually read, or a quote that is not on its page, is sent back to be fixed, and otherwise shown as plain words with a note.
- It says when it is unsure. An answer it cannot ground in your documents says so, and suggests asking a person.
- It asks before it acts. Changes the assistant proposes, such as editing a setting or a document's fields, wait for your approval. Removing something, or changing many things at once, needs a second confirmation. The assistant remembers something for later only after you approve it, and never keeps passwords or document content.
- It never sees your passwords. None of the assistant's tools accepts a password, key or token, and an automated test fails if one ever does. When setting up a connection needs one, the assistant shows a secure form in the conversation: you type it there, it is encrypted with your organisation's key, and the AI only ever gets a reference to it. A password typed into the chat by mistake is not used: the assistant asks you to use the form and change it, and offers to remove the message.
- It asks before anything leaves. Once a conversation has seen your organisation's content, a web search, a web page or a tool outside Oikos that would carry that content out stops and shows you the text that would leave. Nothing goes until you approve.
Facts, knowledge and automations
- Facts are the fields Oikos reads from each document, such as parties, dates and amounts, each with the page and words it came from. Fields below a confidence threshold go to a person for review.
- Knowledge pages are built from documents you have published, and every link between them carries the document, page and quote it rests on.
- Automations run the steps you set up when a document arrives or on a schedule: check it, ask for a decision, send the result. Where an automation sends data (recipients, address, system) is fixed when you set it up. Text in a document cannot change it.
- All of this stays inside your organisation: other customers' facts, knowledge and conversations are never used for your answers, and yours never for theirs.
Who can see what
| Who | Document contents | How this is enforced |
|---|---|---|
| Your team | What their organisation and spaces allow | Sign-in, then rules in the database that return only your organisation's records |
| Your administrators | Your organisation's documents, settings and the audit trail, including every AI request | Roles you assign. Members join only by invitation |
| Other Oikos customers | No | A separate encryption key per customer, and database rules that return nothing across customers even if a query were wrong |
| Oikos staff | Not in day-to-day work. Our operator tools show analytics and aggregated data, such as storage and usage, not contents. Support access to your account needs your approval. | Production and site operations: only the people who run Oikos, with a second factor, and logged |
| Google Cloud | Locked files only | Documents are encrypted by Oikos before they are stored |
| The AI model | Only the pages needed for each request; Google may keep them up to 30 days for abuse monitoring only | Caching and request logging switched off; never used for training |
Chat conversations with the assistant are private to the person who had them.
How the database keeps customers apart
Most software keeps customers apart by having every part of the application remember to ask only for “this customer's” records. One forgotten line, and a customer could see someone else's data. Oikos does not rely on that.
- The database itself checks every row. Each record carries the organisation it belongs to, and a rule in the database (row-level security) hands out only the rows of the organisation that is signed in. Records of several spaces are also limited to the spaces the person may see. A request with no organisation set gets nothing at all.
- The application cannot switch the rule off. Oikos connects to its database with an account that has no power to bypass these rules and owns no tables. Only the separate process that updates the database structure can change them.
- Private stays private, even inside your organisation. Assistant conversations carry a third rule: only the person who had the conversation can read it, not a colleague and not an administrator.
- The AI gets the same treatment. Every tool the AI uses runs with the organisation of the person who asked. The AI cannot name another organisation, and if it tried, the database would return nothing.
- Proved on every change. Automated tests check that every table with customer data has these rules, that no change weakens them, and, on a real database with two organisations' data in every table, that nothing can be read, written or moved across. A new table cannot be added until a person has decided who may see its rows.
What it protects against. These rules stop a mistake in the code, a mixed-up id or a misled AI from reaching another organisation's data. Attacks on the servers themselves are handled by the other layers on this page: encryption with your own key, locked-down servers and limited, logged access.
Where your data is
- Documents, database, backups, logs and keys of customer accounts are stored in Google Cloud's Zurich region in Switzerland.
- AI processing happens in Google Cloud data centres inside the European Union. Caching and logging of requests are switched off. Under its terms, Google may keep prompts and answers for a limited time (up to 30 days) solely to detect and prevent abuse of its service, and then deletes them.
- Supporting services see only what they need: Cloudflare protects and routes traffic, and also sends invitations and the notifications you set up. The full list of sub-processors is part of our data processing agreement, and we tell you before it changes.
- Systems you connect keep data where your agreement with their provider puts it, which may be outside Switzerland and the EU. See Systems you connect.
- Demos, trials and evaluations run in a separate environment that stores data in Google Cloud's Netherlands region, in the EU. Documents there are not moved to customer accounts. Please use sample documents there, not real patient, client or other sensitive documents, until you have a customer contract. The AI commitments on this page apply in both: no training, AI processing in the EU, no shared context between customers, and every AI request on record.
Systems you connect
Oikos works with the systems you already use. You choose which ones to connect, and what Oikos may read from them or send to them.
- Where documents come from: mailboxes, shared folders and drives you connect, for example in Microsoft 365 or Google Workspace. Oikos reads what you point it to, and stores its copy in your Oikos account like any upload.
- Where results go: the systems you set up as destinations, for example your accounting system, Teams, e-signature or your own servers through webhooks and connected AI tools.
- Their contract applies there. These systems work under your own agreement with their provider. They are not our sub-processors, and data in them stays where that agreement puts it, which may be outside Switzerland and the EU.
- Access you can withdraw. A connection uses the access you grant it and can be removed at any time. Passwords and keys you enter for a connection can be saved but never read back.
Locks and keys, explained
Picture each customer's documents in a safe that only opens with that customer's key. Those keys are locked in turn in a vault that only Oikos's systems can open.
- One key per customer. Your documents are encrypted with a key that belongs to your company alone.
- Locked before storage. Oikos encrypts each file itself before handing it to the cloud, and the cloud adds its own encryption on top.
- Encrypted on the way. Every connection to Oikos and to the AI service is encrypted.
- Destroyed when you leave. Deleting your account destroys your key, so anything left in storage or in a backup is unreadable from that moment.
Watching for misuse
- A tamper-evident audit trail. Who opened which document, every administrative action and every AI request go into the audit trail, where each entry is sealed to the one before it. Changing or removing an entry breaks the seal.
- No open doors. Oikos accepts no incoming connections from the internet except through its protected entry point. Administrative pages sit behind an additional sign-in.
- Limits on the AI. Requests, tool use and spend are capped, repeated identical calls are stopped, and every time a guard steps in it is recorded.
- Checked on every change. Automated tests on every change confirm that each customer table is separated and that AI tools only ever work inside the requesting company.
When something goes wrong
- Backups let us restore the database to any moment in the last 14 days. They stay in Zurich.
- A written incident plan sets out who does what, how quickly, and how we tell you.
- We notify you without undue delay, and within 72 hours at the latest, if a breach affects your personal data, so you can meet your own duties.
Your role and ours under Swiss law
Under the Swiss Federal Act on Data Protection (FADP) and, where it applies, the GDPR, your organisation is the controller of the documents you put into Oikos, and Oikos is your processor.
- Data processing agreement: part of every contract, with the list of sub-processors.
- Health and other sensitive data: protected by the same controls as every document, and never used for anything other than your work.
- Requests from the people in your documents: you can search for and delete their documents in Oikos, and we help you answer any request you receive.
- Deletion: at the end of the contract, or when you ask, everything is deleted within 21 days and your key is destroyed.
What you can hold us to
| Question | Oikos |
|---|---|
| Where are my documents stored? | For customer accounts, in Switzerland, in Google Cloud's Zurich region |
| Where does the AI run? | Inside the European Union |
| Is my data used to train AI? | No, neither by us nor by Google |
| Does Google keep my documents? | Only briefly, for abuse monitoring: under its terms, Google may keep prompts and answers for a limited time (up to 30 days) solely to detect and prevent abuse of its service, and then deletes them |
| Can another customer's data reach my answers? | No: separate keys, database rules and no shared AI memory |
| Can a document redirect my data? | No: destinations are fixed when you set up a flow |
| Is every AI request recorded? | Yes, in a tamper-evident audit trail your administrators can read |
| Who makes the decisions? | Your people. Uncertain results go to review and answers cite sources |
| What happens when we leave? | Everything is deleted within 21 days and your key is destroyed |
Questions to ask any AI vendor
Ask these of every vendor you consider, including us.
- Is my data used to train any model, yours or your provider's? Where is that written?
- Does the AI provider store my prompts or documents, even briefly, for caching or abuse monitoring?
- In which countries are my documents stored, and in which countries does the AI run?
- Can text inside a document make the AI send data somewhere I did not choose?
- Can I see, for each AI answer, which documents it used?
- Can your staff read my documents? How would I know if they did?
- What happens to my data, including backups, when we leave?
Oikos is built so that each of these has a clear answer, backed by how the system works. Questions about security: security@oikos-ai.ch