Smart Insights in Transactions
Designing an AI-powered transaction search that keeps users in control.
Client:
Candescent
Category:
Product Design, AI
My Role:
Product Designer
Introducing contextual AI into transactions for Candescent's new mobile banking platform
When people open their transaction history, they're usually there for a specific reason. They're looking for a purchase, checking a refund, finding a recurring charge, or trying to understand what they spent. They're not looking to have a conversation with their bank.
I was brought into a new mobile banking initiative with a broad request: surface AI contextually within transactions. The team had a PRD and an early prototype, but no defined user problem or established interaction pattern.
I led the research, interaction design, and UI design, working with another designer and presenting concepts to product, AI, and design leadership.
The challenge became: How can AI make transaction search more useful without turning a straightforward banking task into an AI experience?
Defining the role of AI
Before designing the interface, I needed to understand where AI could genuinely add value. Transaction history is fundamentally task-oriented. People come in with specific things they want to find: a merchant, a date range, a refund, a payment, or a recurring charge.
Traditional search is already good at retrieving that information. AI becomes more useful when the user wants to interpret or explore what they found.
“Amazon spending”
Could summarize spending rather than simply returning a list of Amazon transactions.“Subscription spend”
Could identify recurring expenses across multiple merchants.“Why was I double charged?”
Could identify a potential duplicate and explain what happened.
I also looked at existing transaction research, our limited conversational AI research, stakeholder conversations, and comparable financial experiences such as Amex and Bud.
This helped establish the role AI should play. Search retrieves the information. AI helps users understand it. That difference became the foundation for the rest of the design.
The Approach
AI should add context when it’s useful, but the core transaction workflow stays fast, familiar, and task-focused.
Keeping AI out of the way
Because transactions are a core banking workflow, disrupting the user's primary task was one of the biggest risks. I explored how users could discover AI capabilities without having to learn a new workflow first. My earliest concept used an AI toggle that let users switch between traditional search and AI search.
It gave users control, but it also created a problem: users had to decide whether they wanted AI before they searched. That meant understanding the AI feature became a prerequisite to using it. I moved away from the toggle and toward a model where AI could be available automatically, while remaining visually and behaviorally secondary to the transaction experience.
The user shouldn't have to decide whether to use AI. They should be able to search normally and discover its value along the way.
Expanded insight
One concept allowed users to expand an insight and see more detail directly within the transaction page.
While useful, this pushed the transaction list further down the page. Since seeing transactions is still the primary purpose of the screen, I wanted to preserve as much of the list above the fold as possible.
Insight + suggested action
I also explored pairing an insight with actions such as “Learn more” or “Cancel this subscription.”
This created a clear path forward, but the language was difficult to make specific enough, and automatically suggesting an action felt too directive. An insight should give the user information without implying what they should do with it.
These explorations led to a simpler model. We should surface enough context to be useful, then let the user decide whether to go deeper.
Search becomes the entry point
The final interaction keeps the existing transaction workflow intact. Users open an account and see their transaction history with familiar options to search, sort, and filter.They can either select a suggested query, enter a keyword, or ask a natural-language question.
Regardless of how they search, the transaction table remains the primary result. Above it, a Smart Insight provides additional context about the search. For example, searching for Netflix might return the relevant transactions along with a summary about a recent subscription price change.
The insight is intentionally lightweight. It appears quickly, stays above the transaction list, and can be ignored or dismissed without affecting the search results. This creates a simple progression of Search → Insight → Conversation
The user gets the benefit of AI without having to enter a separate AI mode.
Making deeper exploration optional
The Smart Insight acts as a bridge between traditional search and conversational AI.
If the user wants more context, selecting the insight opens a conversational overlay. Because it sits on top of the transaction experience rather than replacing it, closing the conversation returns the user directly to their filtered results. Within the conversational layer, users can ask follow-up questions and explore patterns in their spending.
Suggested prompt chips reduce the burden of figuring out what to ask next. They also make the capabilities of the system more discoverable.
For example, a user might move from:
“Amazon spending”
to:
“How much of this is recurring?”
or:
“How does this compare to last month?”
The interface gradually introduces more AI as the user's intent becomes more exploratory, rather than asking them to start with a blank chat interface.
Designing for trust
Introducing AI into a financial workflow also meant being careful about how confidently information was presented.
Insights were designed to stay grounded in factual transaction data rather than making broad or speculative claims.
The experience also incorporates AI disclosures to communicate that generated responses can contain mistakes.
This was especially important because the same interaction can move from a straightforward transaction search to an AI-generated interpretation. The distinction between those two types of information needs to remain clear.
The final experience
The resulting experience creates a gradual path from a familiar banking task into financial exploration:
Find
Search for a transaction or group of transactions.Understand
Get an AI-generated summary or contextual insight.Explore
Ask follow-up questions and identify spending patterns.Act
Move toward actions such as setting up autopay, making a payment, unsubscribing, or exploring additional financial data.
The most important design decision was ultimately moving away from the AI toggle. Rather than asking users to decide when they need AI, AI accompanies the workflow and becomes more prominent only when the user chooses to engage with it. The result is a conversational AI experience that adds a new way to explore finances without making AI the reason users have to change how they search their transactions.
Related Projects
