AI Assistant Documentation

How to read this document

Sections move from purpose → concepts → calculation process → worked examples → edge cases. Formula boxes use plain mathematical notation for clarity; they are conceptual summaries, not API contracts.

• Point lists call out operator-visible behaviour.
• Callouts mark warnings, limitations, or important qualifications.
• Worked examples use round numbers so you can reproduce the arithmetic by hand.

Purpose

InteliHouse is the AccountHouse assistant for entitled accounts: it answers in a stream with inline interactive charts, keeps conversation history, and shows a privacy card explaining what it can read. It is available on the web and in AH Lite, and in the deprecated AccountHouse app for existing installs. Accounts without the InteliHouse entitlement get a simpler AI Chat instead — real and usable, but without history, charts, or the privacy card. Generated output can be incomplete or wrong and must be verified.

• Ask questions in natural language
• Read streamed answers with inline interactive charts (InteliHouse)
• Summarize permitted records
• Get navigation or organizational help
• Verify before acting

Related intelligence products

• Insights Engine — scheduled suggestion cards about spending patterns (separate entitlement).
• IntelliVis — charts and evidence that explain Insights Engine cards.
• ReceiptVis — receipt scan into expense or Split drafts (separate entitlement).
• These products do not replace verifying AI answers against source records.

Data, consent, and permissions

The assistant operates within the authenticated user's available context and permissions. It does not gain authority merely because a user asks for restricted data.

• Per-area consent toggles cover income, goals, and cards — InteliHouse asks on first use, and until you grant an area it answers without that data.
• Two areas are hard-excluded with no consent toggle: Asset Registry and DocuVault are never read by InteliHouse or any AI feature.
• Household permissions still apply on top of consent — consenting to an area does not reveal records your role cannot see.

Verify generated output

• Open source records.
• Recalculate important amounts.
• Confirm dates, filters, and currency.
• Do not treat citations or confident language as proof.
• Correct source data rather than relying on a conversational memory.

Financial boundary

Prompt privacy

• Do not enter passwords, API keys, card credentials, or unnecessary identity data.
• Assume prompts can become part of operational records under applicable policy.
• Avoid requesting data about people outside your permitted scope.

Actions and confirmation

InteliHouse can suggest actions, but a suggested action executes only after your explicit confirmation — nothing is written to the ledger from conversation alone. Review the exact records, amounts, dates, and scope before confirming.

Smart category learning

Separate from the assistant, a smart-category feature learns your household's own description-to-category conventions and suggests a category when you add an expense. It is suggestion-only: it never applies a category without you accepting it, and it learns only from your household's data.

• Suggestions improve as the household's own history grows.
• Rejecting a suggestion is normal feedback — just pick the right category.
• Smart categories are a per-household entitlement like the assistant itself.

When output is uncertain

• Narrow the request.
• Provide non-sensitive context.
• Ask for assumptions.
• Use deterministic reports.
• Contact official support for account-specific issues.

Privacy, advice limits, and what we do not publish

AccountHouse is a household tracking and organisation tool. It is not a bank, payment processor, tax advisor, or investment advisor. Forecasts, health scores, balances, and alerts are informational.

• Household data is scoped to your household; other families cannot see your ledger.
• Public documentation explains product behaviour and calculation concepts, not proprietary internals.
• Exact scoring weights, anomaly thresholds, cryptography parameters beyond high-level algorithms, and operational detection rules are reserved.
• Generated content may hallucinate, omit context, or become outdated; never use it as sole support for a consequential decision.