Where AI runs inside NestWise

No number that affects your finances is ever produced by an AI. Every dollar in the app comes from published government rates or your own inputs, run through auditable formulas in lib/rates.ts and the calculation modules.

AI is used only where it’s good at what it does: reading documents you hand us, and answering navigation-style questions from our public guides.

• The in-app help assistant (visible in the help menu when signed in). Powered by Anthropic’s Claude Haiku, called from our servers. Only reads the public NestWise guides at /guides. Every dollar figure or percentage it says is server-side validated against the guide before we show it to you. If it can’t answer safely, it escalates to a human at hello@nestwise.net.au.
• Inbox document extraction. When you forward a bill, letter, or statement to your private NestWise inbox address, we use Claude vision to extract the specific fields (amount, due date, provider name) so the entry lands prefilled on your dashboard. The original document is stored encrypted in Australia; the extraction is what saves you retyping.
• Payslip / statement OCR (opt-in per-scan). Same pattern as inbox extraction — you upload, we extract structured fields (YTD gross, YTD PAYG, YTD super, employer name, pay period), and the extracted values populate the tax hub.
• Bank-transaction categorisation (when open banking ships). Deterministic rules first; a small model as fallback only when rules can’t classify. Every categorised transaction is editable by you.

• Not for any entitlement calculation. FTB A/B, CCS, PPL, Newborn Supplement, Rent Assistance, Age Pension, Family Tax Benefit — all deterministic, sourced from published Services Australia + ATO rates.
• Not for tax calculations. Refund estimates, marginal rate, offsets, Medicare Levy, LMITO, HELP repayment — ATO formulas in code, no model in the loop.
• Not for “what should I do” recommendations. NestWise deliberately doesn’t give personalised advice (see /legal/tools-not-advice). No AI ever tells you which super fund to choose, whether to salary-sacrifice, whether to refinance.
• Not for cashflow projections. The 12-month + 13-week forecasts are pure arithmetic over your scheduled bills + income.
• Not for “users like you also…” profiling. We don’t build behaviour profiles or lookalike audiences.

The AI assistant runs under strict data-minimisation rules:

• What it sees: the question you typed + the relevant public guide text.
• What it does NOT see: your income, your ATI, your entitlement amounts, your bank data, your kids’ names, your dollar figures — none of it. The prompt-construction layer strips these defensively server-side before the API call.
• What the extraction models see: only the specific document you uploaded. Nothing else about your account.
• What Anthropic sees: the specific prompt we send them (question + guide text, or document image). Under our commercial agreement Anthropic doesn’t train on our API traffic and doesn’t retain it beyond the operational cache.

• Don’t use the help assistant — nothing goes to any AI if you don’t open it.
• Don’t upload documents to the inbox / payslip snap / statement snap flows — the extraction only runs when you initiate it.
• Manually categorise bank transactions — the categorisation model only fills gaps; every row remains editable.

Report it at hello@nestwise.net.au with “AI accuracy” in the subject. We treat AI-caused mis-extractions or unsafe assistant answers as first-priority bugs: full trace review, root-cause fix in the guardrail layer, notify affected users if a pattern is found. See /legal/data-breach-response for how we handle it if the impact spreads.