← All case studies
AI concept · PhonePe group split · Capstone, IIT Roorkee PM programme

Dinner was fun. Splitting the bill was maths homework.

I audited PhonePe's group-split flow, found the friction people complain about most, and specified, prototyped and defended an AI feature that splits a photographed bill fairly, taxes included.

DiscoveryOCR + language AIPRD · FigmaGuardrails
My roleSolo: audit, PRD, prototype, metrics
WhenMay – June 2026
ForIIT Roorkee PM capstone
StatusConcept; not affiliated with PhonePe
Real public reviews Thoughts, illustrative Me
A restaurant bill: pizza, paneer tikka, cold coffee, garlic bread, fries, with GST and service charge
Scene 1 · A Goa trip, the last dinner
Meet the group planner

One person pays. Then the maths begins.

Four friends, one bill for ₹1,288: a pizza, paneer tikka, two cold coffees, garlic bread and fries, plus 5% GST and 10% service charge. One person pays. Now they have to work out who owes what.

The group plannerAman had the pizza and a coffee… Riya had the tikka and fries… and how do I split the service charge fairly? Where's my calculator?
🧠
Cognitive load

Itemising, adding up and sharing out taxes by hand is real mental work, done at the end of a night out, under social pressure.

PhonePe's split flow today: 12 steps

Find "Split" inside the Send-money menu1–4 · discovery & entry
Create a group, type the total, divide it5–8 · equally, by share, or custom amounts typed by hand
Settle, verify, check the ledger, chase9–12 · settlement & tracking
Scene 2 · Opening the app
The audit

"Equally" is easy. "Fairly" is a calculator.

I walked through the flow step by step and collected public reviews from LinkedIn, X, Reddit, YouTube and the Play Store. The pattern was clear: the app splits a total, but people want to split what they ate.

A user on Xthe 'split with friends' feature in GPay, Paytm, and PhonePe uses overall shares, percentages, or manual entries… focuses more on the exact items that they want to share
A LinkedIn postsplit-type features are hard to find in PhonePe, contrasting Paytm's clearer 'Split Bill' banner.
🧠
Jobs to be done

The job isn't "divide ₹1,288 by four". It's "make everyone pay for exactly what they had, without an argument."

Five frictions → one AI bet

F1 · Split is hard to findquick win: UI fix
F5 · Chasing friends is awkwardquick win: better reminders
F3 · Receivers see a bare amount, no billfalls out of F2
F4 · GST & service charge split by handfalls out of F2
↓
F2 · Item-by-item maths: the big bet for AI
Scene 3 · The impact-vs-effort board
Prioritising

Five frictions. One worth an AI.

MeDiscoverability and reminders are real, but a UI change fixes them. Itemised maths is different: it's an understanding problem. Read a messy bill, map "I had the pizza" to a line, share out the taxes. And once the app understands the bill, tax splitting and bill proof come for free.
🧠
Use AI only where rules can't reach

A form can't read an arbitrary restaurant bill or understand plain language. That is where AI earns its cost and risk.

Split Expenses screen with a new Scan Bill option marked NEW
Scene 4 · Same screen, one new button
The feature

Smart Split: snap the bill.

"Scan Bill" sits right where the pain is, at the step where you'd otherwise type a total and divide it. Manual entry stays one tap away.

The group plannerOh, I can just photograph it?
🧠
Put the help where the pain is

Features hidden in menus don't get used. The entry point belongs inside the flow, at the exact moment of friction.

Processing screen: bill captured, extracting line items, identifying taxes, calculating shares
Scene 5 · A few seconds
How it works

Read the bill. Understand who had what. Share the taxes.

  • OCR reads every line: item, quantity, price, GST, service charge, total.
  • Language AI maps plain words ("I had the pizza and one coke") to the right lines and people, or you just tap.
  • Allocation logic shares GST and service charge in proportion to what each person ate.
🧠
Show progress, not a spinner

Naming each step ("extracting items", "identifying taxes") makes the wait feel shorter and the result easier to trust.

Result: Aman 480, Riya 402, You 406, each including GST and service, with how the AI split it
Scene 6 · The result
The aha moment

Taxes included. Every rupee explained.

The group plannerThat would've taken me ten minutes and an argument.

Every amount stays editable, the screen says it is AI-generated, and there's a "Was this split correct?" check. (The prototype screen uses placeholder amounts; the try-it demo below does the exact maths.) The shares must add up to exactly ₹1,288, or nothing can be sent.

🧠
Explainability builds trust

"How AI split it" shows the items behind each share. People pay faster when they can see what they're paying for.

Result screen with each person's share
Interlude · Your turn
Be the group planner

Who had what? You decide.

This is the maths Smart Split does for you. Tap a name to change who had each item. GST (5%) and service charge (10%) are shared in proportion to what each person ate, and the shares must add up to exactly the bill.

The Goa bill · ₹1,288

Tap a name to reassign an item.

🧠
Fair beats equal

An equal split looks simple but quietly overcharges whoever ate least. Fairness is the whole promise of the feature, so the maths has to be exact.

Error screen: couldn't read the bill, low image clarity, confidence 38 percent, retake or enter manually
Scene 7 · A blurry photo
When the AI isn't sure

Say "I couldn't read this" instead of guessing.

A cropped, blurry bill gets a clear reason (confidence 38%, bottom of the bill cut off) and two ways forward: retake the photo or enter the amount by hand.

🧠
Honest uncertainty

In payments, a fast wrong answer is far worse than a slower right one. When the AI is unsure, it must say so.

Help screen: the user corrects which items were theirs, with an option to talk to a support agent
Scene 8 · "That coffee was only mine"
Humans in the loop

Easy to correct. Easy to reach a person.

If the AI gets an item wrong you just say so, and the split updates. A real support agent takes over if you report an item twice, tap "Talk to support", or the AI's confidence stays below 60%.

🧠
Never trap the user

If the AI service is down, the app drops straight to the manual split with no waiting. The feature can fail without the flow failing.

Split requests sent, group ledger showing who has paid and who is pending
Scene 9 · The friends' phones buzz
The other side

A request with the bill attached.

Receivers see the items behind their amount, not a bare "Pay ₹402". The bet: proof makes people pay faster, and the planner has to chase less.

A YouTube comment (Hinglish)Bhaisahab kis duniya mai ho notification se dost pese denge gaali dene par toh dete nhi hai"Brother, which world are you in? Friends won't pay because of a notification; they don't pay even when you swear at them."
🧠
Reduce the social cost

Reminders alone don't fix awkwardness. An itemised, fair request removes the reason to argue.

North Star

Weekly ₹ value of bills split-and-settled through Smart Split within 7 days
~1,600 / armA/B test, +5 pp to detect, 80% power
> 2%hallucination rate → kill switch
P95 < 5 sscan to split
< ₹0.50AI cost per scan
Scene 10 · Measuring it
Metrics & defence

Count money settled, not scans.

A scan only delivers value when the money actually comes back, so the North Star counts settled rupees. The guardrail I defended before the review panel: no split can be sent unless every share adds up exactly to the bill total. That adds a step when the AI is unsure, and I accepted the trade-off.

MeA silently wrong amount doesn't just annoy someone. It makes a friend overpay. In money features, trust is the product.
🧠
Vanity vs. value metrics

"Scans" goes up even when nobody pays. "₹ settled in 7 days" only goes up when the whole chain works.

Go deeper

The full case study is six documents: audit, PRD, responsible-AI plan, prototype, metrics and the panel defence.

A concept project for the Future Ready Product Management with Applied AI programme (iHUB Divya Sampark, IIT Roorkee). Not affiliated with or endorsed by PhonePe. Reviews are quoted from public posts. The bill and names in the prototype are sample data.