Why we don't have an AI feature yet

If you look at fintech product releases from the past two years, one pattern stands out: every company has added an AI feature. AI budgeting insights. AI portfolio recommendations. AI advisor chatbots. The language is similar across products and the execution is often indistinguishable.

Vault & Compass products don't have AI features yet. We want to explain why, because the reasoning matters for how we think about building these tools.

The problem with AI for financial data

Most of the "AI" in current fintech products is doing one of two things: categorizing transactions (which rule-based systems have done for years, and do reliably), or generating natural-language summaries of financial data (which are impressive-sounding but rarely useful).

"Your spending on dining out increased 23% last month" is a sentence a large language model can generate. It's also a sentence you could generate yourself by looking at a number. The LLM wrapper doesn't add insight. It adds a conversational interface around a calculation.

For the AI to add genuine value, it needs to do something that improves your decision-making in a way that the underlying data presented cleanly would not. That bar is harder to clear than most fintech AI features acknowledge.

The trust problem

Financial advice is one of the few domains where incorrect AI output causes direct, material harm. If an AI categorizes your restaurant spending as grocery spending, you get a misleading budget. If an AI recommends a portfolio change that's wrong for your tax situation, you could face unexpected taxes or poor risk exposure.

The liability and trust dynamics around AI for financial decisions are genuinely difficult. Existing regulatory frameworks (investment advisers, broker-dealers) were not built for AI recommendation systems. The industry is still figuring out where AI-generated financial guidance falls on the fiduciary spectrum.

Building AI features on financial data before the trust and regulatory framework is clear enough to do it responsibly is something we've decided not to do for now.

Where we think AI would actually help

This isn't a principled opposition to AI in finance. There are places we think it would genuinely add value:

Document processing in wealth management. Advisors often receive client documents (custodian statements, tax returns, old financial plans) in PDF format. Extracting structured data from these documents is tedious and error-prone when done manually. AI that accurately extracts and structures this data could meaningfully reduce administrative workload.

Natural-language queries against your own data. "How much did I spend on home improvement in 2025?" is a question your Sheetful data could answer, but currently requires writing a spreadsheet formula or sorting a column. A reliable, accurate query interface would be genuinely useful.

Anomaly detection. Identifying transactions that are statistically unusual compared to your patterns (potential fraud, duplicate charges, subscription price increases) is a natural fit for ML approaches and doesn't require the model to generate opinions about your finances.

Report generation. Generating a first draft of a client-facing narrative from structured portfolio data is a task where AI quality is high enough to be genuinely useful when a human professional reviews and approves the output before it goes to a client.

The actual timeline

The reason we're writing "yet" rather than "never" is that these use cases are worth building when the quality bar is high enough and the trust framework is clear enough.

We're not interested in shipping an AI feature because every competitor has one. We're interested in shipping AI features that actually improve outcomes for the people using these products.

When we build them, we'll explain why they're ready.