PlannedBuilt-in AI: what would be worth having?
We haven't built an assistant you can ask questions about your finances, and the reason is security rather than effort. Two things concern us. The first is prompt injection. Transaction data isn't clean input: merchant names, memos and descriptions are text we receive from other systems, and anyone who can get text into your transaction history could try to plant instructions in it. Feeding that into a model that also has access to your accounts is a bad combination. The second is simpler, which is that piping everyone's financial data through a model by default widens the blast radius if anything goes wrong. So the first thing we shipped is the AI connector (MCP) instead. You connect your own AI client, you decide when it runs, and the connector controls what it can see. That doesn't make prompt injection disappear, but it does mean we're not the ones quietly feeding your data into a model, and you can see exactly what's being shared. What we think is worth building in is narrower: - Fraud and anomaly detection. Flagging a charge that doesn't fit your pattern before you spot it on a statement. - Spending habits. The slow drifts that are hard to notice month to month: subscription creep, a category trending up, a recurring charge that changed amount. Both of those run against your own history and don't need you to ask anything. What would you want here? And is there anything you'd rather we never do?
AI
Aaron Sullivan·about 2 months ago2 