Banking apps show where the money went, but not what to do about it. I designed an assistant that finds what changed in your spending, explains why and suggests a plan. Whether to use it is up to the person.
Concept, August 2026Solo project4 daysFigma, FigJamClaude, GPT
7 key screens
14 states, from a spending overview to an active plan
4 days
from the problem to a UI kit: research, flow, screen design and components
01Role and format
02Context and task
Banking apps show the balance, history and categories, but don’t answer “what exactly can I change”. Out of four everyday problems I picked the most common one. Percentages in the text are a guide; exact amounts are in the app.
Turn spending history into one clear action: show what changed and why, and suggest an option.
4 → 1
Out of four financial problems I chose one
+10%
Spending this month is above usual. The app shows the amount, but not the reason
+40%
That’s how much delivery grew. In the total it’s barely visible
1 question
“What changed and what do I do about it?” The product answers in a few steps
03What AI is for here
Here AI is neither a calculator nor a chat. It looks for patterns in spending and answers the question “what to change”.
Not “delivery went up”, but “it went up by 40%: more orders, mostly on weekdays”
Not “you spend a lot”, but “40% more this month than you usually do”
You can’t cut rent, but you can cut food orders. AI only suggests the latter.
If you show what changed and what to do about it, spending is easier to control than by going through a budget by hand.
04Core principle and path
The architecture shows what the product is made of. The user flow shows how a person moves through it: Notice → Understand → Explore → Decide → Act. I show one main path and the places where you can decline or go back.
Core principle
Every tip comes with an explanation. The decision is the user’s.
People see why the app suggests something and can check it.
TrustExplanation
Every tip has a “Not for me” button. The app doesn’t push an action.
ControlChoice
Before agreeing, you can try an option on and see how much it saves.
What ifSimulator
Each screen has one job: understand, choose or confirm.
FocusOne screen
The essentials first. Details open when a person wants to dig in.
Essentials onlyDetails later
One main path
Each step answers one question. People don’t have to analyze, decide and calculate at the same time.
05Problem → solution
For each decision: the problem, what I did, why, and what I traded off.
Problem
Categories show where the money went, but not what to change. Delivery spending went up: is that a lot or as usual?
Solution
The first screen shows the month’s total and the deviation from the norm (e.g. +10%), briefly explains the reason and gives 2–3 observations instead of a dashboard.
Why
People remember two observations, not twenty numbers.
Trade-off
Some data is hidden. The rest is behind “View more”.
Problem
A category total doesn’t explain the reason: why did delivery cost more than usual?
Solution
We compare a person with themselves: “usually 5 orders, now 7, delivery spending +40%”. We say “more than usual”, not “too much”.
Why
A comparison with your own habit is clearer than an abstract amount and doesn’t make you feel guilty.
Trade-off
It needs spending history. For a new user AI has nothing to compare yet.
Problem
The tip “spend less on delivery” says nothing: by how much and how?
Solution
A specific option: “cut weekday delivery, that’s about 25% of this category”. Next to it, the reason and two buttons: “Explore options” and “Not for me”.
Why
An insight without a next step is just information. “You could try” leaves the choice to the person.
Trade-off
The more precise the tip, the higher the risk of guessing behavior wrong.
Problem
It’s unclear how much to change a habit: by one order or by two?
Solution
The habit slider recalculates savings instantly: from 0% to ≈25% of the category. People find a realistic level themselves.
Why
People try options on and choose themselves instead of accepting a ready answer.
Trade-off
The calculation is an estimate (labelled “AI estimate”). Numbers must not sound like a promise.
Problem
After the analysis people close the app: it’s unclear where to start.
Solution
Three goals: small (≈15% of delivery spending), recommended (≈25%), ambitious (≈35%). Then a plan: one action, a goal for the month and progress.
Why
The flow should end not with “I spend too much”, but with a first step.
Trade-off
Personalization vs simplicity: I only ask what changes the tip. Automation vs control: AI calculates, the person confirms.
06Visual language
Large numbers, a calm background, lots of air. Color carries meaning: green is a benefit, violet is AI.
The typeface was chosen for how well numbers read on mobile screens.
“AI Insight” and “AI Recommendation” with a violet accent. It’s clear at once that this is AI.
Violet accentLabel
Every hint comes with a short explanation of why it appeared.
ExplanationTrust
AI is built into the screens as a layer, not a separate dialog.
LayerNot a dialog
07UI kit
New screens, such as plan progress, are assembled from ready components.
Navigation & system
Status Bar, Header, Back Button, Bottom Container
Buttons
Primary, Secondary, Text, Disabled
Cards & content
Result Highlight, AI Insight, Progress, Transactions
Controls & fields
Segmented Section, Slider Card, Track, Thumb
Data display
Transaction Row, Progress Track, Trend Badge
Utilities
Icons, Divider, Bullet Icon, Avatar
08States
Without spending history AI has nothing to compare, so the first day and possible failures are designed separately.
09Final UI
From the month’s total to an active plan. Each slide is one step of the path.
10Results
Typical tracker → AI Spending Coach
What people see
AmountsWhat changed
Who analyzes
By handAI, with a reason
Outcome
“Too much”One plan
The product wasn’t launched and the effect wasn’t measured. It’s a concept: the hypothesis hasn’t been tested with users yet.
Ready for development
11Reflection
Whether AI explanations are clear and whether people trust the savings estimate. For now it’s a hypothesis, not research.
Usability test
Empty, loading, error and “no history” are designed, but not tested in the flow.
Edge cases
How the system behaves with irregular spending.
Testing with data
What I’ll do
How I’ll know it worked
Test the hypothesis from “Why AI”.
Usability test, 5 people
Judge how the recommendation works, not how the screen looks.
Share of accepted recommendations; share of plans kept to the end of the month
A progress screen and plan recalculation.
People don’t drop the plan after a week
Irregular spending and a new user with no history.
“No history” and “error” states are clear without hints
I’m happy to walk you through the logic, decisions and mockups.
