AI Disclaimer / AI Usage Policy
How Expenu AI features actually work in the codebase: propose and confirm, not autonomous ledger authority — plus the answer-guard safeguard for Ask Expenu style outputs.
Anchored to proposeTransaction/confirmProposal and apps/api/src/ai/answer-guard.ts.
AI’s role
Expenu AI interprets user input (for example free-text transaction descriptions or receipts) and may propose structured drafts. It also powers Ask Expenu Q&A, insights, and suggestions in confirm-first flows.
AI never becomes the financial authority of record. Ledger posts, recurring creation, and similar writes go through existing product services after human confirmation where those flows require it.
Propose → review → confirm
Transaction proposals are created via propose APIs and only become posted transactions when a user confirms (confirmProposal and related routes). Receipt OCR follows the same review-before-post pattern.
Budget and recurring suggestions likewise require explicit confirm actions; members without permission receive errors rather than silent writes.
Not financial advice
AI answers and proposals are not guaranteed financial, tax, investment, or legal advice. You remain responsible for reviewing figures, categories, and counterparties before confirming.
Answer-guard (fabricated figures)
Ask Expenu and related narration paths run assistant text through finalizeAssistantAnswer in answer-guard.ts. Financial numeric literals in the reply are checked against a whitelist derived from tool/financial envelopes.
Figures that cannot be justified from available Expenu tool results are blocked or replaced (including a cannot-determine message when appropriate). This is a real implemented safeguard — not a vague marketing promise — with known residual limits (for example spelled-out amounts are not fully stripped).
Model providers
When AI is enabled, prompts may be sent to the configured provider (OpenAI, Anthropic, or Gemini). Do not paste passwords, API keys, or webhook secrets into AI prompts — product prompts instruct models not to solicit those secrets.
Limitations
Models can misread receipts, mis-categorise merchants, or omit context. Network or provider outages disable AI features without affecting core ledger entry.
We do not claim human-level accuracy, regulatory approval of the AI stack, or that every edge case is caught by answer-guard.
Related
Product overview: /ai-assistant. Security overview: /security. Privacy: /privacy.