TASCA DO BAIRRO Lisboa Bacalhau à Brás14,50 Polvo à lagareiro18,00 Vinho da casa6,00 Pão e azeitonas2,50 Café1,00 TOTAL42,00 € NIF cliente: 2•• ••• ••• READ €42,00 Eating out Tax number hidden Filed

Consumer AI apps · Case study

One Receipt, Two Lives

An independent AI product case study for Air Apps: one photo of a Lisbon restaurant receipt moves through scanner, translator and money apps, and gives a resident and a tourist two different, correct answers. A first real step from 38 separate apps toward one Personal Resource Planner.

Computer vision Language AI Recommendations Privacy by design AI or not, step by step
Scan the receipt ↓ Read the case ↓

Independent product concept written against a public job description. Not an official Air Apps product. Ana, Tom, the restaurant and the receipt are made up. Company figures and Portuguese tax facts come from public sources; each is labelled SOURCED, EXAMPLE or MY VIEW.

Start with the domain

What matters when 100 million people use your apps for small, daily jobs

Tap each card. Every one of them changes what we build, what users feel, and where AI is allowed to act.

0138 apps, one promiseSOURCED38+ apps and 114M+ downloads, while the mission is a single Personal Resource Planner.+
ProductThe value is in the hand-offs between apps, not inside any one app.
UsersToday a receipt scan ends as a PDF. The user does the rest.
AIAI is the glue: it turns one app's output into the next app's input.
02A subscription is a promise kept every daySOURCEDAir Apps One bundles 38+ apps for $19.99 a month.+
ProductEach app must earn its place in the bundle, or the bundle looks empty.
UsersOne wrong total in a money app is a reason to cancel.
AIAI must say "please check this" rather than guess on money.
03Everyday paper carries personal dataSOURCEDIn Portugal, a receipt can carry the buyer's tax number (NIF), linked to the tax authority.+
ProductA "simple" scan is a privacy feature whether we plan it or not.
UsersPeople do not think of a dinner receipt as sensitive.
AIThe tax number is found and hidden before any AI model sees the text.
04Every user lives in a different language and countrySOURCEDTranslate Now offers 300+ languages.+
ProductThe same receipt means different things to a local and a visitor.
Users"42,00" is forty-two euros. The comma is a decimal point in Portugal.
AIReading must respect local number and date formats, or totals go wrong.
05Utility apps are opened for one job, then closedMY VIEWScan, translate, calculate. Nobody browses a scanner.+
ProductThe next step must be offered inside the job, not in a menu.
UsersA suggestion at the wrong moment feels like an advert.
AIRecommendations must earn trust slowly: one useful nudge beats five.

The problem

Same photo. Two people. Two different right answers.

Ana · lives in Lisbon EXAMPLE

"Did this count?"

  • She gave her tax number at the till.
  • She wants the dinner in her monthly budget.
  • She needs to know it is linked to her tax record.
  • She must never see her tax number leak anywhere.
TASCA DO BAIRRO
Lisboa
Bacalhau à Brás14,50
Polvo à lagareiro18,00
Vinho da casa6,00
Pão e azeitonas2,50
Café1,00
TOTAL42,00 €
NIF cliente: 2•• ••• •••
Tom · visiting from London EXAMPLE

"What did I just pay?"

  • He cannot read "Polvo à lagareiro".
  • He thinks in pounds, not euros.
  • He has a trip budget and no idea where he stands.
  • Tax numbers mean nothing to him. Don't mention them.

Today, Scanner Air would turn this into a clean PDF for both of them. That is where the help stops. The rest of the job, reading, translating, filing, checking the budget, lives in other apps or in their heads.

SOURCEDFor Ana, the stakes are real: in Portugal, an expense without her tax number on it does not count toward her yearly tax record.

The essential question

Should AI do this step at all?

I ask it for every step, before choosing a model. On this receipt, AI earns its place in three steps and is kept out of four.

Read the photocrumpled paper, bad light, any layout
AI
Receipts have no standard layout. Reading them is a computer vision job, and a fixed template would break on the next restaurant.
Find and hide the tax number"NIF cliente: 2•• ••• •••"
NO AI
It is a labelled number in a fixed format. A simple pattern check finds it every time, runs on the phone, and hides it before any AI model sees the text.
Translate the dishes"Polvo à lagareiro"
AI
Menus are full of local names. Language AI handles context that a dictionary cannot.
Convert €42,00 to poundsfor Tom
NO AI
It is arithmetic with a published exchange rate. An AI model adds risk and no value.
Pick a category"Eating out"
AI SUGGESTS
AI proposes it; the user can change it with one tap. Their correction teaches the next suggestion.
Say whether it counts for taxfor Ana
NO AI
The tax authority's system decides. The app only reports what is on the receipt: "Your tax number is on this receipt."
Warn about the trip budget"€42 of €300 used"
NO AI
Addition against a limit the user set. When to show the warning is a product rule, not a model guess.

Try it

Scan the receipt as Ana, then as Tom. Then break it.

Simulated. Exchange rate and budget are examples.

Three AI jobs, three ways to fail

Where each kind of AI goes wrong on a receipt, and what the user sees

Computer vision

Reads the wrong number

Printed: 1.250,00 €
Read as: 1,25 €
Fix: read with the receipt's country format; if the items don't add up to the total, stop and ask. User sees: "Is the total 1.250,00 €?"
Language AI

Translates too literally

Menu: Prego
Literal: Nail
Meant: steak sandwich
Fix: translate dish names as food, and show the original next to it. User sees: "Prego (steak sandwich)".
Recommendations

Pushes at the wrong moment

First scan ever
"Try our budget app!"
Fix: offer the next app only when it finishes the job the user started. User sees: "Add to your trip budget?" after the second receipt, not the first.

Privacy by design

The tax number never leaves the phone

The job description asks for ethical AI and data privacy. On a receipt, that is not a policy page, it is the order of the steps.

Stays on the phone

  • The photo itself
  • The tax number, found and hidden first
  • Ana's category corrections

May leave, stripped

  • Dish names for translation
  • Total and currency, no name, no tax number
  • Only if the user has turned the feature on

Never

  • Selling or sharing receipt data
  • Using receipts for ads
  • Training on a user's receipts without a clear yes
MY VIEWThe order matters more than the model: hide first, then read, then send. If the order is wrong, a better model only leaks faster.

From 38 apps to one planner

This flow is the first real piece of the Personal Resource Planner

Each arrow is a hand-off that today the user does by hand.

Scannerreads the photo Translatedishes, for Tom Expensesfiles "Eating out" Budget€42 of €300 Plannerthe year, in one view solid: this case · dashed: where it leads

What I would measure

Measures that tell us the hand-off works, not just that the AI ran

QuestionMeasureMust never happen
Did the AI read it right?Receipts filed with no user correctionA wrong total saved without the user confirming
Did the hand-off help?Scans that end in a filed expense, not just a PDFAn expense filed twice from the same receipt
Is the planner forming?Bundle users who start using a second app from this flowSuggestions shown before the user finished their job
Is privacy holding?Receipts processed with the tax number hidden firstA tax number leaving the phone: zero, always

Version one

Ship small, keep the promise, earn the next step

Ship now

Scan to expense

  • Read, hide tax number, file with one-tap category
  • Translation for items, original shown alongside
  • "Please check" whenever items and total don't match
Earn next

Trip and budget

  • Trip budget warnings after the second receipt
  • Currency view for visitors
  • Monthly summary across all scanned receipts
Never

Lines not crossed

  • Telling Ana what she can deduct: the tax system decides
  • Filing anything without the user seeing it
  • Using receipts for anything the user did not ask for

Questions I would expect

Why start with receipts, not a bigger planner feature?

Because one receipt already touches three apps and all three kinds of AI in the job description. If this small hand-off works, the planner is a set of more hand-offs, not a new product.

Why not let AI decide the tax question too?

Because the answer already exists in the tax authority's system. The app's job is to report what is on the receipt, for example "your tax number is on this receipt", not to give tax advice.

Sources: Air Apps website (apps, downloads, Air Apps One) · Air Apps, AI Product Manager job post · RealLX, Your NIF at the register

More AI product case studies

Each one starts from the domain and asks where AI earns its place.

← Back to AI Product