How AI uses your data
Big Picture Money is local-first. AI features are opt-in and disabled by default. This page explains what data may leave your computer when you enable AI and how the app protects your privacy.
Want maximum privacy? Configure Ollama as your provider in Settings. Ollama runs models on your own computer — chat, categorization, and title generation stay on your machine. Nothing is sent to Big Picture or a third-party cloud. See Configure an AI provider to set it up.
What stays on your device
| Data | Stays local |
|---|---|
| Your encrypted database | Always |
| Chat history | Stored only in your local database |
| Work logs under assistant replies (Worked for …) | Stored with that conversation on your device — not uploaded as training feedback |
| Help look-ups the assistant runs | Search the same built-in documentation bundled with the app; results stay on your computer |
| Ollama AI inference | When Ollama is your provider, prompts and model responses stay on your computer |
| LLM categorization cache | Permanent local cache — repeat lookups never call the provider again |
| Embedding neighbor matching | Runs on-device; no network |
| On-device classifier | Trained from your corrections; weights stay local |
| API keys | Operating system secure credential store — not in the database |
What may be sent to a cloud provider
This section applies when you use a cloud provider (OpenAI, Anthropic, Google, and similar). With Ollama, the same prompts are built locally and sent only to your Ollama endpoint on your computer — not to the internet.
When you enable AI, choose a cloud provider, and send a contextual chat message (or when the LLM categorization layer runs), the app may send:
- Transaction summaries you are viewing (payee, amount, date, category, account name)
- Screen context (current page, filters, date ranges)
- Your category tree names (for categorization prompts)
- Your chat message and prior turns in the active thread (trimmed to your context window setting under Settings → AI assistant → Advanced)
- A short summary of the conversation when generating a conversation title (after your first exchange on a new thread)
The app shows a data disclosure modal before the first contextual message with a cloud provider so you can decide whether to continue. Ollama users still see provider details in Settings, but no financial context leaves your device.
Titles, charts, and tables in Chat
| Topic | Behavior |
|---|---|
| Conversation titles | Generated via your configured provider after the first assistant reply. With Ollama, title generation runs entirely on your computer. With a cloud provider, a short summary of the thread may be sent for the title. Titles aim to be neutral and short — they avoid exact amounts and account numbers in the list. You can rename a title; your custom title is not overwritten. |
| Charts and tables | Built from your transaction and report data inside the app. The model chooses filters and chart type; the app computes the numbers. |
| Proposed changes (categories, tags, transactions, goals, views, reports, and similar) | Shown as preview cards only. Nothing is written to your ledger until you click Apply (or consent to a listed batch). Proposed mutations are not a new secret egress path — Apply runs locally through the same app services as the register. |
| Copy | Copies the reply as markdown on your clipboard. The app does not collect thumbs-up / thumbs-down ratings for model training. |
See Using AI Chat for how to use these features.
Redaction
Before data is sent to a cloud provider, the app strips patterns that look like account numbers and other sensitive tokens from prompts. Redaction is best-effort — always review what you share when using cloud AI. Ollama does not need redaction for privacy against third parties because prompts never leave your machine.
Provider choice
| Provider type | Where processing happens | Privacy |
|---|---|---|
| Ollama (local) | Your computer only — inference runs against your local Ollama endpoint | Highest — no data sent to Big Picture or third-party clouds |
| Cloud (BYO key) | The provider you choose (OpenAI, Anthropic, Google, etc.) | You trust that provider with the redacted prompts described above |
| Managed AI (Connected Services) | A provider operated through Big Picture Money | Planned with the optional Connected Services add-on — not available in this release |
For the strongest privacy while using AI, install Ollama, pull a model (for example llama3.2), and select Ollama in Settings → AI assistant (configure the model and endpoint under Advanced). You control the endpoint, model, and when Ollama runs.
Managed AI — a Big Picture–hosted provider path — is a feature of optional Connected Services and is coming soon. Until that add-on ships, Big Picture Money does not operate a hosted AI service.
Disabling AI
- Open Settings.
- Turn off Enable AI.
With AI disabled, no chat requests are made and the LLM categorization layer does not run. Embedding neighbor search and your rules continue to work — they do not require AI to be enabled.