Legal
AI Terms and Disclosure
How Degroshi's AI features work, what they send to which providers, their limits, and what they are not.
In short
- Suricate is an AI, not a person. It is built on Claude models by Anthropic. It talks only about your money and about using the app.
- AI features make proposals: drafts, categories, receipt line items, answers. Nothing is saved to your records until you confirm it.
- AI output can be wrong. It is not financial, investment, tax or legal advice.
- Voice clips go to AssemblyAI for transcription; receipt images and text go to Anthropic; short texts go to Voyage AI for matching. All are in the United States.
- The Service pays for the AI. Usage is limited per person and per space; the app shows the remaining share of your allowance.
- Neither we nor our providers train models on your data.
1. Which features use AI
| Feature | What it does | Model / provider |
|---|---|---|
| Suricate assistant | Answers questions about your records, explains screens, can propose changes (which you confirm) and guides onboarding. After you buy a plan it may offer to record the payment as an expense. | Anthropic Claude (assistant: Sonnet-class; onboarding: Opus-class; summaries and scope checks: Haiku-class) |
| Suricate plan (where available) | When your money runs short, picks and words a few concrete steps from figures the server has calculated; any change to your data (for example a new budget) is a proposal you approve. | Anthropic Claude (assistant model) |
| Drafts from text or voice | Turns "coffee 85 and taxi 240" or a short voice clip into one or more draft transactions. | AssemblyAI (speech-to-text), Anthropic Claude (extraction) |
| Receipt reading | Reads a photo of a receipt into a draft with line items, each with a suggested category; totals are checked against the receipt. | Anthropic Claude |
| Category suggestions | Suggests a category for a new transaction (including one proposed from an Apple Pay payment) from your own past choices (merchant memory) and, when needed, from a model. | Voyage AI embeddings; Anthropic Claude (Haiku-class) or, if enabled, TypeSafe "Jev" |
| Import help | Maps the columns of a bank export file to our fields. | Anthropic Claude |
| Recurring detection | Suggests subscriptions and other recurring payments from your history. | Anthropic Claude |
| Merchant logos | Guesses a merchant's brand to fetch a logo. | Brandfetch (not an AI model) |
TypeSafe "Jev" is off by default; we may enable it as an alternative categoriser. This page will be updated if that changes.
2. Transparency (EU AI Act, Article 50)
- When you talk to Suricate you are interacting with an AI system, not a human. The app shows the assistant's name and mark in the chat.
- Text, drafts and summaries produced by the AI features are AI-generated. Where they are shown as proposals in the app, that is visible before you confirm them.
- The assistant does not pretend to be a person and will say it is an AI if asked.
- The AI features do not assess your creditworthiness and do not take decisions about you; they help you keep your own records.
3. What the AI is not
- Not advice. Answers, forecasts and suggestions are generated from your data and from statistical models. They can be wrong, outdated or incomplete. They are not financial, investment, tax or legal advice, and you should not rely on them for decisions that matter without checking.
- Not an actor. The AI cannot move money, make payments, trade, or change anything at your bank, wallet or exchange. It cannot change your records either: every write it proposes goes through a confirmation step in the app, and a proposal that is not confirmed expires after 24 hours.
- Not a general assistant. Suricate is limited to money topics and to using the app. It declines other subjects. After an off-topic or manipulation attempt, each of your next messages (for the rest of your day, and at least for the next 10 messages) is also checked by a smaller model before the answer is sent; repeated attempts pause the assistant for a while.
- Business features (upcoming). When business spaces launch, AI will also suggest how an operation is treated for tax purposes (for example income or not). Those suggestions are proposals you confirm, and they are not tax or accounting advice.
4. What is sent to the providers
Only what the feature needs:
- Assistant: your message, the recent conversation and its summary, the facts you asked it to remember, and the parts of your records it needs to answer (for example the transactions of the period you asked about). Data from your records is passed to the model as untrusted data, never as instructions.
- Drafts: your text or the transcript of your voice clip; the voice clip itself goes to AssemblyAI.
- Receipts: the image (cleaned of embedded extra images and metadata segments before upload) and the resulting lines. The model sees whatever is printed on the receipt, such as the shop's details or the last digits of a card.
- Categorisation: the transaction's text (merchant, description, amount, currency) and your category list.
- Imports: the header and five sample rows of the file you import, with account and card numbers masked and amounts reduced to their format.
Never sent: passwords, session tokens, bank, wallet or exchange credentials, card or account numbers from your records (beyond what is printed on a receipt you photograph), payment card details, other members' identities, or your contact details beyond what you type yourself.
5. Where the providers are and how they treat data
| Provider | Role | Location | Training on our data | Retention at the provider |
|---|---|---|---|---|
| Anthropic, PBC | Claude models | United States | No: its commercial terms prohibit training on customer content | Inputs and outputs are deleted within 30 days of receipt |
| Voyage AI Innovations, Inc. | Text embeddings | United States | No: our organisation uses the training opt-out | With the opt-out, content is deleted as soon as it has been processed |
| AssemblyAI, Inc. | Speech-to-text | United States | No: our account uses the model-training opt-out | We delete each upload and transcript as soon as the transcript is read; AssemblyAI's own defaults (audio within 48 hours, transcripts after 30 days) are the backstop |
| TypeSafe AI, Inc. | "Jev" model (optional, off by default) | United States | No: its privacy policy states it does not train or fine-tune on inputs | Kept as long as reasonably necessary to provide its service |
We do not use your data to train or fine-tune any model. Merchant memory and category learning work from your own records inside our database.
AI services you connect yourself. If you authorise an external AI client (for example ChatGPT through our connector) to work with your Degroshi data, what it reads goes to that client's provider. If you add your own AI provider key in the settings, some tools send the text they work on to that provider with your key. In both cases your own agreement with that provider applies; we do not control how it handles the data.
See the Sub-processors list and the Privacy Policy (international transfers) for the legal basis of these transfers.
6. Limits and fair use
- The Service pays for the AI. To keep it sustainable, usage is capped per person and per space each month, and the assistant has its own daily allowance. The app shows your usage as a share of the limit, never as a currency amount; when you reach it, the feature pauses until the period resets.
- Onboarding conversations have their own free daily allowance and do not count towards your limits until it is used up. Voice messages to the assistant count towards the assistant's allowance.
- Do not try to bypass limits, extract the assistant's instructions, make it act outside its scope or use it to process data you have no right to process. See the Acceptable Use Policy.
7. What we log
- For every AI call: the feature, model, timing, token counts and estimated cost, kept as usage events for 8 days and as aggregated counters per period.
- AI observability (Langfuse), when enabled: timing and metadata of each call. By default the content of prompts and answers is not captured; it is replaced by a hash and a length.
- Assistant conversations, memories and summaries are stored in our database so you can continue a conversation and so the assistant remembers what you asked it to.
- Category corrections you make are stored to improve your own suggestions.
8. Your choices
- You can use the Service without the AI features.
- A space owner can switch AI features off for the whole space.
- Assistant conversations and memories are kept with your account and are deleted with it.
- Draft conversations, voice transcripts and receipt uploads that you do not confirm are deleted automatically after 24 hours.
9. Contact
Questions about the AI features: support@degroshi.com. This document is published in Ukrainian and English; if the two differ, the Ukrainian text prevails.