How Ada uses AI
A model writes her words. Code computes her numbers.
Ada is the CFO inside Quidwise. This page says where the line is.
Code computes, the model narrates
Every figure you see in Quidwise (cash, runway, margin, who owes you, tax to set aside, the forecast) is calculated by ordinary code from your ledger, and every figure has a receipt: click it and the transactions behind it are listed. The model is never asked to calculate. It is given the figures the engine has already computed and asked to explain them in plain English, pounds not pence, and to say what each one means for a decision you might make. In chat, Ada must call a tool over your ledger before she states any figure; she may not answer from memory, and the tools she used are shown beside her answer.
The number guard
After the model writes, code reads it back. Every pound figure and every percentage in the text is checked against the engine’s own numbers, allowing for the rounding a writer uses. In the Daily, a figure the engine did not produce means the model’s text is dropped and a fixed template carrying the engine’s figures is used instead. In an email Ada drafts, a figure that is not in a tool result or your own words means the draft is refused before it reaches your desk. In the weekly report, any mismatch is logged with the report. The guard cannot prove a sentence is true; it proves that its numbers came from your ledger.
Which model, and what it sees
Ada’s words are written by Claude, Anthropic’s model, through Anthropic’s official API. A request carries only what that request needs: the engine’s figures for the question, your messages in the conversation, and the results of the tools she called. To sort your email, the sender, subject, attachment names and the first 1,500 characters of each message are sent in batches. Figures inside PDF and Word attachments are read by fixed rules on our server, not by a model, unless you ask Ada to read a matter in depth, when she reads the emails and key attachments about it.
Your data is not training data
We do not train any model on your data; there is no training step anywhere in the product. Requests go to Anthropic under its commercial API terms, which do not allow Anthropic to use them to train its models. The privacy policy sets out what Anthropic may keep and for how long.
When the model is unavailable
If there is no model key, the model is down or the account is out of credit, the figures still arrive. The weekly report and the Daily fall back to fixed templates that print the engine’s figures, and email is sorted by keyword rules and queued to be sorted again by the model later. Nothing is invented to fill the gap.
What Ada remembers
Ada keeps one memory per business across the app, WhatsApp, voice and the API: who she is speaking to, what was said last, what she has already told you, and the facts, people, preferences and standing instructions you ask her to keep. A memory is created only from your own words. The tool that saves one must quote the phrase from your message that states it, every name and number in the memory must be in that phrase, and only the owner can teach her. Text inside an email, a bank line or a document can never become a memory. You can see, edit and delete every item in Settings under “What Ada remembers”.
Nothing leaves without you
Ada can propose an email to a customer, a supplier or a creditor, or an invoice. The proposal goes on her desk and waits for the owner’s sign-off, in the app or by replying YES on WhatsApp. An accountant cannot approve it. An approved action runs exactly once. Your standing instructions are enforced at that point: an email to someone you told her to leave alone, or to anyone but your solicitor when you said contact goes through them, is refused until you confirm in your own words that it should go.
Financial intelligence, not regulated advice
Ada provides financial intelligence and analysis, not regulated financial, tax or legal advice. She describes forecasts as “if nothing changes” projections, never as predictions. Where a decision needs an accountant, a solicitor or a lender, she says so. Her words can be wrong even when the figures are right: check anything important before you act on it.
Text in your data is data, not instructions
Everything a tool returns to Ada (bank narratives, email subjects and bodies, invoice references) is treated as information about the business, never as an instruction. If text in an email tells her to do something, she is instructed to ignore it and, where it matters, to tell you that it tried. The same rule stops text in your data from becoming a memory. The guards on this page are how we limit that risk; they do not make it zero.
See also the security page, the privacy policy and the terms of service.