How We Use AI

Last updated: July 17, 2026

The short version: AI reads and explains, code does the math. AI helps recognize patterns in your transactions and answers your questions in chat. Every number in your forecast comes from deterministic code we can replay and explain. The AI cannot move money, cannot change your forecast without showing you, and its work is never used to train anyone's models.

Most products bury a line about "machine learning" in their privacy policy. We think you deserve the whole picture: which jobs use AI, which models run them, what data they see, and where AI is deliberately not allowed. This page is that picture.

The rule everything follows

AI recognizes, code computes. AI is good at reading messy transaction descriptions and conversations. It is not allowed to do your math. Every projected number, countdown, and red or green signal is produced by a deterministic forecasting engine: plain code that gives the same answer every time and can be audited line by line. When the forecast shows a number, we can show you exactly which assumptions produced it. No number in MMLP is a model's guess.

The four jobs AI does

Classifying transactions. Deciding whether a transaction is income, an expense, a refund, a reimbursement, or an internal transfer. Confident classifications are recorded; uncertain ones become a question for you instead of a guess.
Recognizing recurring patterns. Spotting that a deposit looks like a paycheck or that a charge repeats monthly. Every recognized pattern becomes a visible assumption you can see and edit. Nothing it finds enters your forecast invisibly.
The chat assistant. Answering questions about your money, applying the edits you ask for, and handling support. Its numbers come from tools that query the deterministic engine, never from the model's memory. When it changes something you see a card describing the change, with an undo button.
Summaries and suggestions. Plain-language summaries of your year and suggestions for saving more. Suggestions only ever draw from flexible spending, never from committed costs like rent, childcare, or insurance, and each one is a proposal you accept, edit, or decline.

Which models, for which jobs

We use Anthropic's Claude models, pinned per job and changed deliberately, not silently:

Claude Haiku: transaction classification. A fast, small model for a high-volume, narrow job.
Claude Sonnet: recurring-pattern recognition, the chat assistant, and summaries. A stronger model for the jobs that require judgment and clear writing.

When we upgrade a model we rerun our quality checks first, because a model change can change what gets recognized. Model versions are part of the product, not a moving target.

What the AI sees, and where it runs

Minimum payload. Each job receives only what it needs: the classifier sees transaction descriptions and amounts, the chat assistant sees your conversation and the tool results it requests. No job receives your bank credentials, and none of them can: we never have those in the first place.
Server-side only. Every AI call runs on our servers. Your browser never talks to an AI provider and never holds an AI key.
Never used for training. We use Anthropic's API, and API data is not used to train their models. Your financial life does not become anyone's training set, ours included.
Encrypted at rest. Identifying financial data, including your chat history with the assistant, is field-level encrypted in our database with per-household keys. Details are in the Privacy Policy.

What the AI can never do

Move money. Every bank connection is read-only by design. There is no tool, endpoint, or code path for initiating a transfer, payment, or trade, so the AI cannot be tricked or prompted into one. The capability does not exist.
Change your forecast silently. Every change is recorded as a visible assumption edit with its origin labeled: AI inferred or set by you. Your corrections always override the AI, and AI-applied changes can be undone with one tap.
Guess when it matters. When the AI is not confident about a transaction, the product asks you one concise question instead of quietly deciding. A confident wrong answer is the worst thing a finance product can produce, so uncertainty is surfaced, not hidden.
Compute your numbers. See the rule above. If a number is on your screen, deterministic code produced it.

People are still in the loop

The assistant handles first-line questions, but when you ask for a human, or it cannot answer confidently, it files a ticket and a person follows up by email. AI is how we scale attention, not how we avoid it.

Questions

Ask the assistant in the app, or email hello@mmlp.app. If you think the AI did something wrong, tell us: every report makes the checks above stricter.