Why this exists
Your agent talks to your customers, and what they write is personal data. Under the GDPR and equivalent laws you decide why and how that data is used, which makes you the controller, and Miral handles it on your instructions, which makes us your processor. The law requires that relationship to be written down. This is that document. It forms part of Miral's terms of service for every client and applies automatically from the moment a workspace is created; no signature is needed, and a signed copy is available on request at support@meetmiral.com.
What we process, for whom
Data subjects: your customers and prospective customers who talk to your agent, your staff who use the dashboard or reply in your channels, and anyone mentioned in the knowledge you connect.
Categories of data: the content of conversations, including any name, email address, order number or other detail a person types; the knowledge, documents and pages you connect; account details of your staff; and technical logs of the service operating. Where you connect a store, the order data the agent looks up. You should not route special-category data (health, beliefs, biometrics and the like) through your agent; if your business requires it, tell us first.
Purpose and nature: generating replies to your customers from your knowledge, handing conversations to your staff, showing you what was said, and, where you enable it, taking the actions you have permitted such as looking up or refunding an order. Storage, retrieval and transmission as needed for that, and nothing else.
Duration: for as long as your workspace exists, and up to 30 days after its deletion for the removal of backups.
Our commitments
- We process personal data only on your documented instructions, which are these terms, your settings in the dashboard, and anything further you agree with us in writing. If we believe an instruction breaks the law, we tell you.
- We do not use your data, your conversations or your documents to train AI models, and we do not sell them or use them for advertising.
- Everyone who works on Miral and could see your data is bound by confidentiality.
- We keep the technical and organisational measures described below, and improve them as the service grows.
- We help you meet your own obligations: answering requests from people whose data you hold, security, breach notification and impact assessments, to the extent the request concerns what we process for you.
- We tell you without undue delay, and in any event within 72 hours of knowing, if a personal data breach affects your data, with what we know and what we are doing about it.
- At the end of the service you can take your data with you from Settings and delete the workspace yourself; whatever is left in backups goes within 30 days, unless the law requires us to keep it.
- We make the information you reasonably need to check these commitments available to you, and allow an audit, at your cost, once a year with at least 30 days' notice, or sooner after a breach or when a regulator requires it.
Your commitments
You have a lawful basis for the data you connect and for the conversations your agent has. You give the notices your customers are owed, including that they are talking to an AI agent where that is required. You do not route data through your agent that you have no right to process, and you use the controls in the dashboard, such as pausing the agent, limiting its actions and deleting conversations, to meet your obligations.
Sub-processors
We use the following providers to run Miral. Each receives only what its role requires, and each is bound by data protection terms at least as protective as these.
- Railway (hosting and the Postgres database that holds conversations, knowledge and accounts). Region: EU West.
- Qdrant Cloud (the search index of your knowledge, one collection per client), on AWS in Frankfurt, Germany.
- OpenRouter and the model providers it routes to (generating replies and classifying messages). Conversation text and the relevant parts of your knowledge are sent to produce a reply; providers are contractually bound not to train on it.
- Google (converting your knowledge into search embeddings, and audio or video understanding where a customer sends one).
- Resend (sign-in codes, invitations and handoff notifications to your staff).
- Stripe (subscriptions and invoices; we never see a card number).
- Cloudflare (DNS, and serving this website and the chat widget script).
- Browserless (a headless browser, used only when you start a browser task from your assistant).
Channels and systems you choose to connect, such as Slack, Shopify, Google Workspace or Meta, are not our sub-processors: you connect them, and they process data under your agreement with them.
We will email workspace owners at least 30 days before adding or replacing a sub-processor. If you object on reasonable data protection grounds and we cannot resolve it, you may end the service and we refund any prepaid period you have not used.
Where data goes
Your conversations, knowledge and accounts are stored in the European Union. Generating a reply sends the conversation and the relevant knowledge to model providers who may process it outside the EU, including in the United States. Those transfers rely on the European Commission's standard contractual clauses or on the EU-US Data Privacy Framework where the provider is certified, together with the safeguards described here. Data is not retained by the model providers for training.
Security
- Every connection is encrypted in transit. Credentials for the channels and systems you connect are encrypted at rest with a key that is never stored beside them.
- Each client's data lives behind its own boundary: its own knowledge collection, its own process where one is dedicated, and every database row keyed by the client. One client's data is never read to answer another's customer, and our tests assert it.
- Access to production systems is limited to the people who operate the service, over authenticated connections, and recorded. The dashboard records who signed in, who changed what, and when.
- Public endpoints cap message size, session count and request rate, and a per-client spend limit stops abuse from running up a bill or a queue.
- Backups of the database are kept apart from the key that unlocks stored credentials, so a copy of one is useless without the other.
Requests from people whose data you hold
If someone who spoke to your agent asks us to access, correct or delete their data, we do not act on it ourselves: we tell you within five working days and act on your instruction, since the conversation is yours. You can delete a conversation, an article or the whole workspace from the dashboard without asking us.
Liability and precedence
Liability under this agreement is subject to the limits in the terms of service. Where this agreement and the terms differ on the handling of personal data, this agreement wins. Where the law requires more of a processor than is written here, the law wins.