Case study · AI product design
Twelve
An AI finance app that plans on a twelve month horizon and answers one question every day: am I okay?
Role
Solo. Research, concept, flow, IA, UI, identity and copy
Type
Self initiated concept
Timeline
3 days, August 2026
Tools
Figma
Independent concept work. Not affiliated with any bank or finance company. The numbers are a worked example, not real account data.
The hardest call
Every money app plans in months. Money runs in years.
The first version of this idea led with a single “safe to spend” number. It was a good number and a weak product, because every budgeting app already shows you a number for the month.
The things that actually break a budget don’t run on a monthly clock. Car insurance is yearly. A car service comes round every six months or so. School fees are termly, festivals are seasonal. In a normal year there are around eight of these, and a monthly view sees none of them until they land.
So I flipped the model. Twelve plans across the next twelve months, and today’s number is what falls out of that plan. The year is why the number is true. The number is why you open the app.
Who this is for
Your balance is a lie
Picture someone on a salary in an Indian city, paying rent and an EMI, with a car and a kid in school. The bank says ₹84,000. Rent leaves on the 5th, the EMI on the 7th, and there’s a car insurance renewal on 22 March they’ve forgotten about.
They fail twice. First they overspend early, because the balance looked healthy. Then in March they put ₹18,000 on a credit card. They could afford it all along. Nothing told them it was coming.
₹84,000
what the bank says is there
8
irregular bills a year that a monthly view never sees
₹18,000
the March renewal that ends up on a credit card
Same number, different claim
Every other app says
Twelve says
₹1,850 safe today, because March insurance still needs ₹4,200 reserved.
Same number. The second sentence is the product, and no monthly app can write it, because it needs to know what March looks like before March arrives.
Why a model, not rules
A rule can’t find a bill it has seen once
Detection
Finding a charge that happened once, eleven months ago, under a merchant name you don’t recognise, and knowing it will come back.
Prediction
The date drifts and the amount creeps up at renewal. Both need an estimate with honest confidence attached, not a fixed rule.
Re planning
Every new transaction re runs the whole year, so today’s number moves on its own, without anyone asking.
The flow
Six ways into the loop, and only two start with you
Setup happens once. Twelve reads 14 months of accounts, finds the irregular bills, and hands them back for you to correct. After that the loop runs every day. Twelve computes, notices, asks and audits itself on its own. You come in through only two doors: asking it something, or asking why.
Every path ends in the same place. The twelve month plan gets re solved and tomorrow’s number changes.

Dark boxes are Twelve acting. Light boxes are you or the screen. The app uses the same colour rule, so the flow and the UI teach each other.
Information architecture
No navigation, because there is one destination
Am I okay, and if not, what do I do about it? Home answers that in two states, on track and off pace. Things you ask for open from inside home. Things Twelve decides you need arrive on top. Nothing sits more than one level deep.

Deliberately not built
No transaction list.
No category chart.
No budgets you set yourself.
No chat.
No day, week or month toggle, because a switcher means the app has stopped having an opinion.
The screens
Nine screens, each one a different way the AI shows up
I ordered these by what the model is doing, not by where they sit in the app. Read the labels alone and you get the whole AI story.
01 · AI proposes
Here is what I found
Setup ends with Twelve handing back what it found in 14 months of accounts: car insurance every March, school fees every April, Diwali every October, a car service it has only seen once. Each one carries a confidence label and an Edit button. The card at the top says it plainly: these are my guesses, not your data.

02 · AI computes
One number, with its reason attached
₹1,850 safe today, and right under it the reason: March insurance still needs ₹4,200. The year ahead sits below as a strip of twelve months, with the next bill and its confidence. Home changes every day without you touching it, because every spend re plans the year.

03 · AI shows its working
Tap the number and it opens its working
Every input it used, and how sure it is about each one. It points out its own weakest guess (the car service, seen once) and lets you switch it off to see the number move from ₹1,850 to ₹1,940. This is the screen that proves it’s a model and not a formula.

04 · You ask
It only offers questions it can answer
There’s no blank chat box. Ask Twelve opens a short list of questions the model has enough data to answer well, grouped by today, the year and the plan. At the bottom it says what it won’t do: guess.

05 · AI answers live
Type a number, watch tomorrow move
A number pad, not a text field. Type ₹14,000 and before you spend a rupee it tells you what happens: past ₹4,200 this starts eating March, and tomorrow’s number drops from ₹1,850 to ₹1,240. Adjust the plan, or leave it. Nothing changes until you say so.

06 · AI admits it doesn’t know
₹8,400 went to Sharma Motors and I’ve never seen it before
When something new lands, Twelve refuses to guess. It asks one question (every year, every few months, or one off), shows exactly what your answer will change, and promises to ask only once.

07 · AI notices
You just spent ₹14,000
This arrives after the spend, on purpose. A UPI payment can’t be stopped mid flight in India, so a screen that pretends to catch you before you pay would be lying. It states the damage in terms of your year (March insurance is now ₹6,400 short) and gives you two real paths. It never blocks anything.

08 · AI holds the line
Off pace, and the way back
The same home, second state. ₹740 safe today, down from ₹1,850, with what drifted (food delivery is ₹2,900 over across 11 orders) and two ways back, each with its cost: hold ₹740 a day for 19 days, or spend ₹950 and push part of the March reserve.

09 · AI audits itself
I got two things wrong
At month end Twelve marks its own homework. Car service came in 40% over its estimate, electricity 62% over. It says so, raises the next car service estimate to ₹8,900, and tells you March is still covered.

What I got wrong first
The screens were rebuilt four times
Mistake
The intervention assumed I could catch a payment before it happened.
Fix
Nothing in India can intercept a UPI or card payment in the moment. I reframed it as “You just spent”, a notice after the fact that openly says nothing has changed yet, and added a way in before spending (Ask Twelve) so the “before” moment still exists, started by the person.
Mistake
The screens said they were AI without showing it.
Fix
First person copy and a signature badge still read as a forecasting app. What actually reads as AI is visible reasoning, live answers, named uncertainty and the model changing in front of you. That’s why “Why ₹1,850”, “It asks you” and the live number pad exist.
Mistake
Home led with the year.
Fix
An annual planner gives nobody a reason to open it daily. Today’s number moved to the top with the year as its stated reason, and the twelve month strip dropped to second.
Mistake
The colour was cool on cool.
Fix
A mint accent on a green ground had no temperature contrast. I moved to forest and cream with a single citrus accent, so the moments when Twelve speaks (cream panels) break the dark instead of blending into it.
Trade offs
Every decision closed a door
01
Annual engine, daily surface
The year is why the number is true. The number is why you open the app.
What it cost
Someone skimming home may take it for a budgeting app before they read the line underneath.
02
One question, not a dashboard
Anything that doesn’t answer “am I okay” was cut.
What it cost
No transaction list, no category breakdown. People who like browsing their spending aren’t served at all.
03
An opinion, never a veto
The intervention states the consequence and offers two paths. It never blocks a payment.
What it cost
Some people will spend anyway and end the month short.
04
Correction by touch, not by chat
Every forecast is tappable and editable.
What it cost
A correction a sentence could express easily now needs direct manipulation.
05
Confidence over precision
“Around ₹18,000, medium confidence” is trusted more than a false ₹18,412.
What it cost
The interface keeps reminding you it might be wrong.
06
Two states, one surface
On track and off pace share home, because they’re the same question with different answers.
What it cost
The off pace state can’t have a layout of its own.
07
One way in, and it only offers what it can answer
A set of answerable questions instead of a blank field.
What it cost
A question outside the set has no route.
08
No navigation, because there is one destination
Sheets open from home, takeovers arrive on top.
What it cost
Settings, history and deeper reports have nowhere to live yet.
Identity
Light marks who is speaking
The app sits in deep forest. When Twelve speaks, the surface turns cream, so the system is the thing that breaks the dark. The mark is a year that doesn’t quite close, with the next thing you need to be ready for sitting in the gap. It signs every panel the system speaks through, so it has to survive at 22px.

Numbers are never abbreviated
₹18,000 stays ₹18,000, never 18K.
No shadows, no gradients.
Confidence is written, not implied.
Cream appears once per screen.
What changed
The same ₹18,000 bill, two different Marches
Before (a monthly app)
The balance looks healthy all year. March arrives, the ₹18,000 renewal lands, and it goes on a credit card.
After (Twelve)
The renewal is found in setup, seen twice, high confidence. Twelve sets money aside for it every month, so by August ₹13,800 is already held and only ₹4,200 is left to find. That ₹4,200 is the reason today’s number is ₹1,850, and in March the money is simply there.
The number I’d measure is the share of irregular bills that arrive already fully reserved. A monthly app scores close to zero here by design, because it never sees those bills coming. In the worked example, that’s the gap between a ₹18,000 renewal landing on a credit card and landing on money that’s already set aside. If Twelve works, this share climbs every month, and the month close screen is where you’d watch it happen.
Bills that arrive already covered
The one metric that matters
Stated plainly
Where this breaks
Year one is weaker than year two. Until it has your history, Twelve guesses from category priors and asks questions, and a bill it learns about in July has fewer months left to reserve for, so the daily number drops.
Forecast accuracy for irregular bills needs checking against real statement data before this ships.
And only one of the three rhythms (daily glance, the nudge after a spend, the month close) depends on you remembering to open the app. That’s honest, and it’s also the part I’m least sure about.
What I learned
You can’t assert AI. You have to show it working.
My early screens said “I” everywhere and carried a little Twelve badge, and they still read as a forecasting app. Adding more AI language never fixed that. What fixed it was letting the model do visible things: open its working, answer as you type, admit it doesn’t know, and mark its own homework at month end.
The other lesson came from the payment rails. I designed a beautiful “about to spend” moment before checking whether anything in India could actually trigger it. Nothing can. Now I check the real constraint before I design the screen that depends on it.
More work
Check out more work
ixigo Trains · The four hours nobody designs for
Google Pay Split and Request
Unified Reward Centre · Mobile Product Design



