Legal
AI Disclosure
This page describes, precisely, the single way Vekfinance uses artificial intelligence, what data that use does and does not transmit, and how you can verify the description yourself.
De-identified aggregates only
The AI receives coarse bands, ratios, and generic category labels. It never receives raw transactions, merchant or account names, exact balances, or images.
One path, fully logged
There is exactly one route from the app to the model. Every payload that leaves is recorded in your egress audit trail.
Not financial advice
Everything the assistant says is informational. It is not investment, tax, accounting, or legal advice.
Jump to a section
The AI assistant is not financial advice.
The one AI feature
In plain English
There is exactly one AI feature: the assistant. It appears in three places, and all three receive the same de-identified summary. Nothing else in the product makes an AI call.
Vekfinance has a single artificial-intelligence feature: an assistant powered by Anthropic's Claude. It surfaces in three places, and all three are grounded in the same de-identified summary described in the next section:
- The chat assistant, where you ask questions about your finances in plain language.
- The monthly close insight, a single paragraph summarizing the month when you close your books.
- The portfolio analysis, a diversification and cost read of your investment allocation.
No other feature transmits anything to an AI model. Transaction categorization is deterministic and runs entirely on your own server inside the encryption boundary (your rules, the institution-provided category, and a built-in merchant dictionary). The custom report builder is also deterministic. Neither makes an AI call.
What it receives and what it never receives
In plain English
The assistant receives a de-identified summary: coarse money bands, rounded ratios, generic category labels, and relative time periods. It never receives raw transactions, merchant or payee names, account names or numbers, exact balances, or any image.
Before any AI request, your raw data is reduced, on our server and inside the encryption boundary, to a fixed de-identified structure (we call it the InsightContext). That structure, plus the text of the question you type, is the only thing transmitted to Anthropic.
What the model receives:
- Money expressed as coarse bands (for example, "$10k to $25k"), never exact figures.
- Rounded ratios and rates (savings rate, debt-to-income band, weighted APR band).
- Investment posture as ratios only: asset-class allocation percentages, your largest single holding as a percent of the portfolio, and the weighted expense ratio. Never a ticker or a specific holding.
- Generic category labels (groceries, dining, housing). Any category you renamed collapses to a generic bucket.
- Relative time descriptors (current month, prior three months), never absolute dates.
- Counts and de-identified trend descriptors (for example, "dining up 24% versus your trailing average").
- The text of the question you type into the assistant.
What the model never receives:
- Raw transactions or transaction lists.
- Merchant or payee names.
- Account names or account numbers.
- Exact balances or exact amounts.
- Investment tickers or specific holdings.
- Any document, receipt, or image.
- Your name, email, password, other users' data, or session identifiers.
The question you type is your own input and is sent as written, after an automated scan that blocks an email address or account-number-like sequence from leaving by accident. Anthropic states they do not train on data submitted via their API; see their privacy policy at https://www.anthropic.com/legal/privacy. The assistant is optional: the budgeting, debt payoff, categorization, reporting, and transaction tracking all work without ever invoking it.
How you can verify this
In plain English
Every payload that leaves for the model is recorded in your egress audit trail, in plain language, so you can confirm only de-identified aggregates left.
You do not have to take this description on faith. Every request to the model passes through a single enforcement point that validates the payload against a fixed allowlist, scans it for anything resembling personal or raw financial data, and records the exact content in an audit trail you can read at any time under your privacy settings. Each entry states, in plain language, what the model was told and what it was not told.
Known limitations
In plain English
AI can be confidently wrong. Always verify against the actual numbers in the Service.
- AI can be wrong. Large language models can produce confident-sounding output that is incorrect, incomplete, or outdated. Always verify a recommendation against the underlying numbers shown in the Service before acting on it.
- The model reasons over summaries, not exact figures. Because the assistant receives coarse bands and ratios rather than exact amounts, it speaks in ranges and directions. For exact figures, read the underlying screens in the Service.
- Projections are not forecasts. Debt payoff dates and net-worth projections use simplified assumptions (current minimums, seven-percent annual investment return, no new contributions). Real-world outcomes will differ.
- No fiduciary relationship. Using the AI feature does not create any advisor-client, fiduciary, or legal relationship between you and us, between you and Anthropic, or between you and any model.
Your responsibility
In plain English
The decision is always yours. Use AI as a thinking tool, not as a substitute for licensed advice.
You are responsible for any financial decision you make. Use the AI assistant as a starting point for thinking, not as a substitute for advice from a licensed financial professional.
Contact
In plain English
[email protected] for questions about the AI feature.
Questions about how the AI feature works:
[email protected]
Get in touch
Questions about how the AI feature works