What does an AI financial coach actually see? Here is ours, byte for byte

Publicado 4 de octubre de 2026 · 6 min de lectura

Every finance app is adding an AI assistant, and the pitch is the same: ask about your money in plain English, get an answer. What almost none of them will tell you is what the assistant is handed to produce that answer. For most, it is your transactions: merchants, amounts, dates, account names, sometimes balances. That data leaves the app and goes to a model provider every time you ask a question.

We built Vekfinance so that cannot happen, and the easiest way to show it is to show you a real payload. Below is what the coach received for the question "Should I pay extra on my card or build my emergency fund first?" on a demo account.

What was sent

The server builds a summary from your encrypted data and runs it through an egress chokepoint. Only fields on an allowlist can leave. For that question, the payload contained:

• A net-worth band and whether it is trending up or down.
• An income band and a spending band for the last 30 days, the savings rate as a percentage, and whether net cash flow was positive.
• A debt-to-income band and an emergency-fund runway bucket in months.
• For debts: how many accounts, a total-balance band, weighted and highest APR bands, and a payoff-horizon bucket. No per-card balances.
• A subscription count with a monthly-total band, and for investments a value band and allocation mix.
• The question itself, with identifiers stripped.

That is the whole thing. No transactions, no merchant names, no account names, no exact balances, and nothing that identifies you.

What was withheld

The deny side of the allowlist is enforced in code and tested. For the same question the coach was explicitly never told: the names of the accounts, any individual card's balance or limit, the exact APRs, the exact amount in savings, any individual purchase, and the user's name or email. The disclosure under the answer lists these as "never told" so you can see the boundary, not just trust it.

How the answer still contains real numbers

If the model never sees your figures, how does the answer say "your Sapphire card is at 22.5% APR with $1,200 outstanding"? It does not, directly. The model writes placeholders, tokens that mean "the balance on the highest-APR card" or "spending in category X this month", and our server resolves each placeholder from your actual data with deterministic code before the answer is shown. The model reasons about which numbers matter; the arithmetic is never its job.

How to check it yourself

Ask the coach anything. Under its answer there is a disclosure,See exactly what the AI was told. Open it and you get three views: what was sent, what was withheld, and the raw payload. It is read from the same audit log the request wrote, so it cannot drift from what actually happened. If a competitor offers the same, we would genuinely like to see it.

What the AI is used for, and what it is not

The model powers three things: the coach, the one-paragraph insight at monthly close, and the optional portfolio analysis. Transaction categorization is not AI; it runs your rules, the bank's category and a merchant dictionary, then a review queue. There is no receipt scanning and no image ever leaves your account. The full list is on our AI disclosure page, and the architecture is on private AI budgeting, explained.