> ## Documentation Index > Fetch the complete documentation index at: https://docs.consile.ai/llms.txt > Use this file to discover all available pages before exploring further. # Security & data > Least privilege, EU hosting and encryption with a key per customer: how we handle your data. *[Læs siden på dansk](https://docs.consile.ai/essentials/sikkerhed-og-data.md)* This is how Consile protects and handles your data, in plain words. ## The five principles > **Least privilege** > > The data connectors (Google Search Console, CVR, Dinero, Billy, Ordrestyring > and Business Central, which is in early access) are read-only: your AI can read > your numbers, but never change, create or delete anything. > The write connectors (e-conomic, Uniconta, Shopify, HubSpot CMS, HubSpot > Marketing Email, LinkedIn Ads and Meta Ads) can > also create and change content for you, but only behind safety barriers. By > default, two rules apply: new content is created as a draft or paused, > and actions that spend money, publish, send, post to the accounts or delete > require your confirmation. Without confirmation, you only get a preview. > See [Actions & guardrails](https://docs.consile.ai/en/essentials/actions-and-guardrails.md). > **Hosted in the EU** > > Consile is hosted and operated in the EU (Frankfurt, Germany), and your encrypted > access credentials are stored only there. The answers from your systems are passed > on to the AI service you have connected yourself (for example Claude or ChatGPT). > **Your AI never sees your token** > > Consile holds its own, separate access to your data source, and your AI is > never given it. The login your AI uses with Consile is also never passed on > to your systems (no "token passthrough"). > **Encrypted with a key per customer** > > Tokens, API keys and logins are encrypted with keys that can only be used for > your own account, and that are protected by a hardware-secured key vault. Another > customer's key cannot open your data. > **Access is checked on every call** > > On every tool call, Consile checks that your organization subscribes to the > connector, that your user has been given access to it, that the tool is part of > your access level, and whether the tool requires an administrator's approval. > Only then is the stored access to your system used. If the connection is locked > to one company, Consile also checks that the call stays with that company. > Consile never trusts your AI client to stick to the tools it has access to on > its own. > [!NOTE] > > The safety barriers in the write connectors are built-in defaults. How the > confirmation works, and exactly what requires it, is described in > [Actions & guardrails](https://docs.consile.ai/en/essentials/actions-and-guardrails.md). Consile always > enforces three things on the server: the data connectors have no > write tools, a user on the **Read only** access level cannot use > write tools, and a tool that requires approval runs only once an > administrator has approved the call. ## Access per user An administrator decides under [Team](https://docs.consile.ai/en/portal/team.md) which users have access to which connectors. A user without access to a connector cannot use its tools, even if the organization subscribes to it. With [the Consile Studio membership](https://docs.consile.ai/en/portal/membership.md), the administrator instead gives one licence per person, which covers all of the membership's systems. For e-conomic and Uniconta, the administrator also chooses an access level for each user: **Read only** (28 tools in e-conomic, 21 in Uniconta) or **Full access** (40 tools in e-conomic, 25 in Uniconta). With Read only, Consile blocks all write operations on the server for that user. Both levels cost the same. If the administrator changes the level, a conversation that is already open may show the old tools until the user starts a new chat. ## Calls that need an administrator's yes An administrator can decide that a tool requires approval. The rule applies to the whole organization, including administrators and AI bots. A call to the tool waits, and nothing is carried out until an administrator has approved it under **Team** → **Approvals**. Consile enforces this on the server, whichever AI client is used. See [Approvals](https://docs.consile.ai/en/portal/approvals.md). ## Double authorization (token vending) When you connect, two things actually happen: **1. Your AI logs in to Consile** Your AI client (for example Claude) authenticates with Consile using your Consile login, not with, for example, Google. That is how Consile knows who you are and which connectors you have access to. **2. Consile gets its own access to your data source** You give Consile access to, for example, Google, or you enter an API key in the portal. Consile stores the access encrypted, and where the provider allows it, Consile renews expiring login tokens itself. Meta Ads is an exception: there you usually have to connect again about every 60 days. Your AI never sees this access. The result: your AI can ask questions, but neither it nor any other customer can reach your access credentials. ## How customers are kept apart > [!NOTE] > > Each customer (*tenant*) is completely isolated. Which data a request may > touch is decided only from the session Consile has verified itself, never from > anything your AI sends along. A tool uses only your own organization's stored > access, and your access is checked on every call. ### One company per connection Some logins reach several companies, for example a Google account with many websites in Search Console or an accountant login in Dinero. In Google Search Console, Meta Ads, LinkedIn Ads, Dinero and Business Central, an administrator can lock a connection to one company. Consile then checks on every call that the call stays with that company, and refuses calls to others. Tools that cannot be limited to one company are not available on a locked connection. If Consile cannot limit an answer to the company, the call is refused instead of answering with data that may belong to another company. See [Company lock](https://docs.consile.ai/en/portal/company-lock.md). In Uniconta, a connection works only in the company (Company ID) it is set up with, even if the server user has access to several companies. ## What we store and log The content from your systems, for example accounting data, orders and search figures, is not stored by Consile. The tools read live from the provider, and the answer is passed straight on to your AI. We log technical events, for example which tool was called, how long it took, and whether it succeeded. If a call fails, we log the error type and any status code, but not the provider's error text. For the usage statistics, we count the number of calls per tool per day. * **The activity log:** For every tool call, Consile stores who made it (a person or an AI bot), which tool, in which system and which agreement, when, the outcome and how long the call took. An agreement is one connection to a system. Administrators see the log under **Team** → **Activity**. Each call is kept for 90 days. See [Activity log](https://docs.consile.ai/en/portal/activity.md). * **Approvals:** For a call that is waiting for approval, Consile stores who asked for it, the system, the tool, the status, the times and who answered. Neither the values in the call nor the system's answer are stored. Consile stores only a fingerprint that can recognize exactly the same call again. The call cannot be recreated from it. See [Approvals](https://docs.consile.ai/en/portal/approvals.md). * **AI bots:** The bot's own log stores the time, tool, result and the address the call came from, for 30 days. See [AI bots](https://docs.consile.ai/en/studio/ai-bots.md). The content of the calls (what your AI sends to the tools, and the answers from your systems) is not logged, not even in the activity log, with approvals or in the bot's log. Tokens, API keys and passwords are never logged. Messages you send us through [Concierge](https://docs.consile.ai/en/concierge.md) are stored, though, so we can answer them. The optional message an administrator can write when a call is approved or refused is also stored. See [GDPR & your rights](https://docs.consile.ai/en/essentials/gdpr.md). ## Your rights (GDPR) * An administrator can disconnect a connection at any time with **Disconnect** in the menu next to the agreement on the **Connections** card on the connector's page. The stored access is deleted immediately, and the subscription continues. Consile does not revoke the access at the provider, so you do that yourself in the provider's settings. * When a subscription ends, Consile automatically deletes the stored access to that connector. * An administrator can delete the whole account under [Account & data](https://docs.consile.ai/en/portal/account-and-data.md). Access is suspended immediately, and all encrypted access credentials are deleted at once. The rest of the account is deleted permanently after a 14-day grace period. The deletion does not cancel your subscriptions, so cancel them under [Subscription](https://docs.consile.ai/en/portal/subscription.md) before you delete the account. * Invoices and payment history are kept by the payment provider, because the Danish Bookkeeping Act requires it. > [!NOTE] > > The Consile service and your encrypted access credentials are in the EU (Frankfurt). > Our login service processes identity data (for example your email) in the US under > EU-approved transfer mechanisms. The full overview, including the list of > suppliers, is in [GDPR & your rights](https://docs.consile.ai/en/essentials/gdpr.md). If you have questions about data processing, write to [support@consile.dk](mailto:support@consile.dk). This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.