Neutron

The personal finance space isn't short on products. It's short on products that actually change behavior. The pattern is familiar: connect your accounts, look at your spending, feel vaguely guilty, close the app. The information is there. The gap between information and action is where everything falls apart. Neutron's bet was that an AI layer could close that gap — not by adding more data, but by adding judgment. Surface the right thing at the right time, explain why it matters, and make it easy to act. Less dashboard, more advisor. The design problem wasn't "how do we display financial data." It was "how do we make AI-generated recommendations feel trustworthy enough to act on."

Research focused on how people actually think about money and make financial decisions — not how finance apps assume they do. Discovery sessions surfaced three main patterns:

• People want guidance, not data: When users talked about what they wished their finance app did, almost nobody said "better graphs." They said things like "just tell me if I'm okay" or "I want to know what I should do differently." The mental model was advisor, not dashboard.
• Trust requires a visible chain of reasoning: Users were skeptical of automated recommendations — not because they thought AI was wrong, but because they had no way to evaluate it. "How does it know that?" came up constantly. Recommendations without rationale felt like guesses.
• Conversation felt less intimidating than navigation: When shown prototypes, users responded more positively to natural language interactions than to traditional menus and filters. Asking a question felt lower-stakes than figuring out where to click.

We established these principles to anchor decisions across the product:

• Explain every recommendation: Not a tooltip. Not a help article. An actual reason, in plain language, attached to every suggestion the system makes.
• Keep the human in control: The copilot advises. The user decides. Every recommendation had to be dismissible, questionable, and optional — the AI doesn't act on your behalf.
• Reduce financial complexity: The job isn't to show users their financial data in a new format. It's to translate it into something they can act on without a finance degree.
• Build trust gradually: Trust isn't established in one interaction. The design had to support a progression — small wins first, higher-stakes recommendations later, as users developed a track record with the system.

Conversational interface. The primary interaction model was natural language. Users could ask questions, request summaries, or respond to proactive suggestions — all through conversation rather than navigation. This wasn't chatbot UX for its own sake. It matched how users already thought about financial questions: as things you ask, not things you click through menus to find. Recommendation design. Each recommendation followed a consistent structure: what the action is, why the system is suggesting it, what the projected outcome looks like, and a clear way to act or dismiss. The format mattered. Recommendations that led with the action and buried the rationale got ignored. Recommendations that led with context got engaged with. Explainability layer. Behind every recommendation was a visible reasoning model — not the technical model, but the logic. "You've spent 40% more on dining this month than your three-month average. Based on your current trajectory, you'll be over your food budget by the 25th." That's actionable. "You may want to reduce dining spending" is not. Confidence signaling. Not every recommendation carries the same certainty. Some are based on clear patterns in the data. Others are more speculative. The design surfaced that distinction — not with percentage scores, which felt arbitrary, but through language and framing. "You're consistently spending more than you earn in this category" reads differently than "this might be worth reviewing." Both are honest. Users responded better to the distinction than to uniform confidence. Decision support, not decision replacement. The copilot was designed to make users better at making decisions — not to make decisions for them. Every flow ended with the user choosing something. The AI's job was to make that choice feel informed and manageable.

Early prototype sessions revealed two things worth noting. The conversational interface worked better than expected for exploratory questions — "what did I spend the most on last month?" — and worse than expected for complex multi-step tasks. Users wanted conversation for understanding and wanted clear buttons for action. We kept both. Confidence signaling through language was harder to calibrate than expected. Phrases that felt appropriately uncertain to us read as evasive to users. We ran several rounds of copy testing before landing on framing that felt honest without undermining trust.

This was a 0→1 project. There were no pre-existing users to measure against, no baseline activation rate, no historical funnel. The measure of success here wasn't a metric. It was whether the product vision held up under real user interaction. It did — and the conversational model in particular became the core of Neutron's product direction going forward.