AI Information and Rules for Using the AI Assistant

This is an English translation of the Polish „Informacja o AI i zasady korzystania z Asystenta AI”. The Polish version is binding; in case of any discrepancy, the Polish version prevails.

Annex 3 to the CoinTrack Terms of Service

1. You are talking to artificial intelligence

The AI Assistant in CoinTrack is a computer program, not a human. In the app it is called “Max”. It has a name only to make talking to it easier. Its answers are generated by an artificial intelligence model. We mark this at the start of every conversation and with every answer (Article 50(1) of Regulation (EU) 2024/1689 - the AI Act): in the corner of the chat you always see the “AI Assistant” label, and at the start of every conversation a notice that you are talking to a program, not a human. Proposed changes in the chat, entry suggestions from receipts and category suggestions during import are also generated by AI. We sign every answer, including one with a proposed change, “Max · AI”, and we label a category suggestion “AI suggestion”.

2. Which features use AI

Feature	What it does	Where it runs
AI Assistant (chat)	answers questions about your budget (e.g. "how much did I spend on food this month?"), suggests entries from a description ("I spent PLN 40 on fuel"); at your request it changes a bottom bar shortcut; at your request it also changes a Task (e.g. its title, due date, amount or repeat settings): it shows the change on a card and saves it only when you click the confirmation button	in the cloud: the Google Gemini model on Google Cloud AI, called by our server; it processes data in the EU
Receipt Scanning	reads the total, date, shop and items from a receipt photo (the “Scan a receipt” screen or a photo attached in the chat) and adds a suggested entry to the “To approve” queue	in the cloud: a Google Gemini model on Google Cloud’s AI platform, with model processing in the EU; we do not store the photo
Category suggestions during import	suggests categories for operations from a CSV file whose category we do not know from your earlier choices; you can turn this off on the import screen	in the cloud: a Google Gemini model on Google Cloud’s AI platform, with model processing in the EU

Availability and limits: the AI Assistant and Receipt Scanning are included in the Plus Plan (100 messages and 100 receipts per Family per monthly limit period) and the Premium Plan (300 messages and 500 receipts per Family per monthly limit period). The message limit is shared by everyone in the Family; each person has their own conversation with the AI Assistant. For a yearly Plan the limits reset every month. Each person in the Family may scan at most 50 receipts a day and send at most 150 messages to the AI Assistant a day. A scan that could not be read does not count. File import and category suggestions are included in the Plus and Premium Plans; you get AI suggestions for at most 20 file imports a day per Account - receipt scans do not count towards this (above that limit, and also for a file with more than 500 different merchants, import works without AI suggestions). These features are not available in the Free Plan (except during the Trial Period). Details in Pricing.

3. Your data and AI

The conversation goes to Google. Our server - not your browser - sends your messages, a receipt photo to be read and the Budget Data needed for the answer to Google (Google Cloud AI). From Tasks, the AI Assistant gets the amounts planned in a given month, as part of the total of planned expenses, and, when it needs them to answer (e.g. you ask about Tasks or ask it to change a Task), also their titles, due dates, amounts, repeat settings and status (e.g. overdue, done). An expense saved when you check off a Task is named after the Task's title and is sent to Google like any other expense. We never send Task notes. If you type the content of a note in a message yourself (e.g. an account number), it goes to Google as part of the conversation. We do not send your e-mail or name. During import, AI receives only cleaned merchant names and operation descriptions, their type and direction and the names of your Family’s categories - without amounts, dates, account numbers, transfers to and from private individuals or operations whose description points to specially protected data.
Google processes this data as our processor, under the Google Cloud data-processing addendum, only on our instructions: it does not use it to train AI models or for its own purposes. The AI model processes data in the European Union. Google stores a prompt only when its automatic filters flag a possible breach of the service’s usage policies - solely to check abuse, for up to 90 days, in the EU. To answer faster, Google’s service may also by default keep fragments of prompts in a cache, in the EU, for up to 24 hours. Details are in the Privacy Policy.
We store conversation history in the CoinTrack database together with the proposal cards and the record of changes made in the chat (what the AI Assistant proposed, the data before and after the change, and when you approved or undid it). We delete them when you click “New conversation”, and at the latest when the Account is deleted. We do not store receipt photos - we delete them after reading, so they are not in the chat history or in the “Download my data” copy either. That copy contains the conversation together with the proposal cards, but not the record of changes (the data before and after the change, or the dates of approval and undo). We will send a copy of the record if you ask at contact@cointrack.pl.
Receipt Scanning requires your consent. A receipt may reveal health data or other specially protected data. Before the first scan we ask you for explicit consent (Article 9(2)(a) GDPR). You give it once and can withdraw it in Settings → Privacy (Terms of Service § 13(8)).
Only what you approve affects your budget: an entry suggested by the AI Assistant, the result of reading a receipt (amount, date, shop, items) or an imported operation with a suggested category - just like a manual entry.
We do not train AI models on your data - neither on conversations, nor on receipts, nor on Budget Data.
We count how many messages and scans the Family has used (for the limits). Each time AI is used we record the type of feature, the model, the number of tokens and the cost - without content - and keep this for 13 months.
If a device is shared, other people with access to your signed-in Account can see the chat history.

4. AI model


Model	Google Gemini - only generally available (GA) model versions, not preview versions (currently: Gemini 3.5 Flash-Lite) - provider: Google, on Google Cloud AI (Gemini Enterprise Agent Platform, formerly Vertex AI)
Where it runs	the “eu” location - the AI model processes data in the European Union
Requirements	a browser that meets the requirements in the Terms of Service (§ 4(1)) and an internet connection; nothing to download

5. Limitations - what you need to know

AI can make mistakes. The model may misunderstand a question, mix up amounts, misread a receipt (e.g. an illegible print, abbreviated names) or present false information convincingly.
Check before you approve. The AI Assistant does not save anything without your confirmation. Before approving an entry, check the amount, date and category. For a change to a Task, the card shows what it was and what it will be, based on the data in CoinTrack, not on what the AI Assistant wrote. All Family Members will see the change. You can undo a one-off expense added in the chat, a change of a bottom bar shortcut and a change of a Task with the "Undo" button on the card within 24 hours of saving, if nobody has changed that data in the meantime and, for a Task, nobody has checked off its due date. Undoing a change of a Task does not restore due dates removed by a change of the due date or repeat settings; the card warns you about this before you approve.
Answers about numbers are based on the data in CoinTrack. If the data is incomplete (e.g. some expenses are missing), the answer will be incomplete too.
A category suggestion may be wrong. Check the category before approving an imported operation. Duplicate checks and matching with your entries do not use AI - they may still link an operation with the wrong entry, which is why we leave the decision to you.
The AI Assistant answers in Polish and in English.
In a long conversation the AI Assistant takes only its most recent part into account. You still see older messages in the history, but the AI model does not use them in its reply. If you come back to an earlier topic, repeat the key details.
What the AI Assistant will not do. The AI Assistant does not conclude or change contracts, does not buy, change or cancel a Plan, does not pay, does not withdraw from the contract, does not give or withdraw consents, does not delete the Account or a Family, does not invite or remove Family Members and does not hand over a Family to another person. You do these things yourself on the right screen in the app. The AI Assistant will point you to that screen.

6. This is not advice

The AI Assistant does not give investment, credit, insurance, tax or legal advice.
The AI Assistant does not recommend specific financial instruments (shares, bonds, funds, ETFs), crypto-assets, loans, credit or insurance and does not assess your creditworthiness. Asked for such a recommendation, it will refuse and suggest contacting a licensed adviser.
It may give general budgeting hints (e.g. which expense categories are growing, how much is left to reach a goal) - these are for information only. You make financial decisions yourself.
The AI Assistant is not designed to push you towards purchases or loans or to exploit a difficult financial situation. We do not earn money from recommending financial products.

7. Rules of use

Do not use the AI Assistant to:

enter other people’s data you have no right to;
attempt to bypass security measures or limits or to gain access to other people’s data;
create unlawful content;
send automated, mass queries (scripts).

8. External AI assistants (Connected AI apps, MCP)

Connected AI apps are not available yet. The rules in this section apply from the day they become available in the app.

You can connect your own external AI assistant (e.g. Claude, ChatGPT) to CoinTrack with an access token (Terms of Service § 14). This is not the CoinTrack AI Assistant.
Connected AI apps are available only to the Family Owner on the Premium Plan. A token is valid for at most 90 days and you can revoke it at any time.
When you connect an external assistant, your Family’s data within the chosen scope goes to that assistant’s provider on your instruction - often to servers in the USA - and is processed under its rules, including as regards model training. Check its privacy settings.
We recommend a “read-only” token with a short validity period. Tokens with write access let the external assistant change and delete Family data.
The external assistant’s provider is responsible for its answers.

9. Reporting errors

If the AI Assistant gave a harmful, offensive or clearly wrong answer or misread a receipt, report it to contact@cointrack.pl (describe what it was about and give the approximate date of the conversation). You can report systematic errors as a complaint (Terms of Service § 16).

10. Model changes

The AI model may be updated (e.g. to a newer version that understands Polish better). We announce a material change (e.g. a change of the model provider) in advance in accordance with Terms of Service § 17.