Sproutjar is a financial wellness app that helps you clear credit card debt, with an AI voice coach trained on positive psychology and life coaching.

Credit card debt in the Gulf costs about twice what the internet thinks it does

Every payoff calculator, budgeting book and Reddit thread a person in Dubai reads was written for a market where a card charges around 21% a year. Here it is 39% to 44%, quoted at 3.25% a month so it sounds like nothing.

That gap is the whole reason this product is local rather than generic. At 44% the cost of hesitating for a year is not an inconvenience, it is a holiday, a school term, a chunk of a salary. And the mechanics that decide the outcome are the ones nobody reads: pay 95% of your statement and interest still runs from the transaction date, go late twice in six months and the rate itself goes up, keep a dormant card open and it eats your borrowing capacity under the 50% debt burden cap.

Retail credit in the UAE grew 16.1% last year and personal loans 8.6%, so more households are carrying more of this. But bad loans fell to 3.3%, and no public source breaks revolving card distress down by age, income or number of cards. So I wrote the honest version into the brief and kept it: the problem is strongly evidenced, the market size is not, and anyone who tells you the size of the UAE card crisis is guessing.

Never leave a rate in monthly form.

How the research was actually run

I did not ask for a report that agreed with me. I ran my own idea past a deliberately skeptical advisor prompt first: who exactly, what breaks in their week, what they already do instead, and why anything I built would beat that.

Every claim I could not source was tagged as an assumption and carried forward untouched, so the research had something specific to attack. Sources were kept in separate buckets and never blended: first person customer voice from UAE threads, institutional material from the CBUAE rulebook and bank key facts statements, vendor and pricing pages for the alternatives, and academic literature for the coaching claims. Anything that survived only in a vendor's marketing copy stayed labelled as such.

Then the verdicts came back, including on the assumption the first version of the product had been built on. I had believed people avoid their cards out of shame. That got contradicted as a primary explanation, and the product changed shape because of it.

• “Shame appears, but so do job loss, illness, family expenses, accumulating EMIs, cash-flow mismatch, poor habits and misunderstandings.”

  Customer and market research report, verdict on my own assumption

What people actually said

Twenty five first person quotes, mostly UAE threads from the last eighteen months. The pattern underneath them was not embarrassment. It was arithmetic that does not work, plans that die on contact with a bad month, and progress nobody can feel.

People were handing complete strangers every card, rate, EMI, salary and expense they had, in the hope that somebody would build them the plan. The stated need was almost always a number. The revealed need was a sequence: what do I pay, in what order, this month, given my actual life.

The line I keep coming back to is the last one. Someone had already repaid eleven thousand dollars and still could not experience it as success, because every interface they own shows them what is left. That is a design problem rather than a finance problem, and it is the one I could actually solve.

• “just got my salary today and in a blink of an eye its gone (debts and bills lmao)”

  UAE thread, via the research report

• “Please make this plan easy to read and understand without using complicated credit and finance terminology.”

  Someone asking strangers for a payoff plan

• “The accountability is 90% of the value for me.”

  A user explaining what he paid a human coach for

• “I don’t know how to change this or be ‘happy’ with my progress.”

  Someone eleven thousand dollars into paying it off

Nothing on the market is trying to change a habit

The tools people already use are good at exactly one slice each. None of them owns the sentence that matters: this was your plan, this changed, here is the new plan, and here is what we do differently next month.

Budgeting apps categorise the past. YNAB genuinely changes spending behaviour and costs USD 109 a year to do it, but it is a budgeting system with debt inside it rather than a route out. Cleo will talk to you and produce a paydown plan for USD 5.99 a month, and it is the closest thing to a competitor, which is why an AI that gives you a debt plan is not a strategy. Debt Payoff Planner and Undebt.it draw the date beautifully for two dollars a month or ten dollars a year, and then leave you alone with it. Spreadsheets are free, flexible, and stop being true the first month you go off plan.

The alternative that actually works is a human coach. One person in the research paid about USD 3,000 for six sessions and said the accountability was ninety percent of the value. That is the shape of the opportunity and the shape of the problem: it works because someone is in the loop with you, and it does not scale.

So the target is not calculation, and it is not education. It is behaviour: a new spending habit, a saving habit that survives a good month, and a plan that gets adapted instead of abandoned. Everything in the product exists to serve that, and the dashboard is instrumentation for it rather than the point of it.

• YNAB

  USD 109 / year

  A whole budgeting method, and it genuinely changes spending

  Debt is a category inside a budget, not a route out of one

• Cleo, Debt Reset

  USD 5.99 / month

  Conversational AI, spending data, a paydown plan

  No memory of your plan across months, no adaptation when the month breaks

• Debt Payoff Planner

  About USD 2 / month

  Snowball, avalanche, a payoff schedule and a date

  Draws the date, then leaves you alone with it

• Undebt.it

  Free, or USD 12 / year

  A capable manual payoff planner with several methods

  You maintain it, and nothing happens when you stop

• A spreadsheet

  Free

  Anything you can model, exactly how you want it

  Stops being true the first month you go off plan

• A human coach

  About USD 3,000 for six sessions, in one researched case

  Context, accountability, and real behaviour change

  Priced out of reach of the person who needs it, and it does not scale

Never use a financial product to solve a behavioural problem, and never use a behavioural lecture to solve an insolvency problem.

I built a synthetic user panel of promoters and detractors, and interviewed all five

Before I could reach real customers I wrote five personas and ran a full interview with each: two who should love this, two who are sceptical, and one who should refuse to buy it. The point was not applause. It was to hear, in one place, why someone would pay for this and why someone would not touch it.

Zayd was the most useful. He knows what avalanche means, has built the spreadsheet twice, and abandons it the month reality diverges from it. His objection is the one the whole business has to answer: why is this not just ChatGPT plus Excel. Omar gave me the sentence I now use to explain the product, that the debt cycle is like dieting, austerity then relief then relapse. Leena showed why the mathematically optimal plan is often the personally unsustainable one. And Maya, who is organised and fine, said the quiet part out loud about my own positioning.

Building the panel with detractors in it is the reason it was useful. Five enthusiasts would have confirmed everything and changed nothing. This produced five buying objections and one clear non customer, which is a cheaper way to find them than shipping. None of it is evidence, and I do not treat it as any: a synthetic panel generates objections and hypotheses, and only real interviews can settle them.

Nobody on the panel wanted a forty minute scheduled call with an AI about their credit cards. That matched the research verdict on voice, and it is why sessions are ten, twenty or forty minutes at the person's choice, nothing is scheduled, and Ren is a room in the product rather than the product.

• “I know exactly what I'm supposed to do. That's kind of the problem.”

  Zayd, 28, synthetic persona

• “Once the spreadsheet is wrong, I kind of stop looking at it. Don't tell me I failed. Re-plan.”

  Zayd, 28, synthetic persona

• “It's like dieting.”

  Omar, 37, on paying cards off and rebuilding them

• “I don't want an app telling me I can't go see my family because that's financially optimal.”

  Leena, 30, who sends money home

• “Positive psychology makes it sound less serious, not more serious.”

  Maya, 34, the persona I should not try to acquire

• Aisha, 25

  Two years into her first serious salary, first cards and BNPL

  Pays what is due, occasionally panics and pays extra

  Low. Wants the debt free date, does not want to be talked to like a child

• Zayd, 28

  Tech professional, two to three cards and a personal loan

  Built the tracker twice, abandoned it after a disrupted month

  Medium. Why is this not just ChatGPT plus Excel

• Leena, 30

  Expat, sends money home, debt accumulated gradually

  Follows finance creators, drowning in conflicting advice

  Medium. Will it tell her not to fly home

• Omar, 37

  Family, good salary, cards plus a car loan

  Has cleared the cards before and rebuilt the balance

  Medium to high. Do not teach him what a credit card is

• Maya, 34

  Organised, automated, has a working system already

  Tracks it, automates it, runs her own scenarios

  High. She is the one who told me not to build it

The coaching research, and how Ren was built out of it

I read the literature before writing a word of Ren's prompt, because financial coaching is a field where the marketing is far ahead of the evidence and I did not want to ship a motivational chatbot with a plant on it.

Three findings did the most work. The working alliance matters: a meta analysis across 27 samples and 3,563 coaching processes found the quality of the relationship moderately associated with outcomes, and with fewer unintended harms. Monitoring closes the loop: across 138 studies and 19,951 participants, interventions that increased progress monitoring improved goal attainment, more so when progress was recorded and reported back. And no branded framework wins: the evidence favours flexible integration of solution focused work, cognitive behavioural coping and strengths, not a proprietary model.

Positive psychology, read properly, is not compulsory optimism. A coach should be positive about human possibility rather than positive about every human experience, and prematurely reframing someone's fear about a 44% balance into gratitude is bad coaching. The evidence on financial coaching specifically is genuinely mixed: one randomised evaluation found improvements in money management, savings, debt and financial confidence, while a 2026 systematic review of eleven controlled studies found small effects, heavy heterogeneity and methodological weaknesses in ten of them. So the product describes itself as structured support for behaviour change, and never as a cure.

Those findings became rules in Ren's prompt rather than vibes. Ren asks permission before challenging, teaching or suggesting. Ren decomposes a global self judgment into a behaviour before agreeing with it. Ren reflects using the person's own words and their actual balances. Every session ends with one commitment the person believes they can keep, and every session opens by returning to the last one, kept or not, without a score.

1. 01

   Understand

   Every balance, rate and minimum in one place

2. 02

   Plan

   A payoff order and a date derived from real numbers

3. 03

   Act

   One commitment for the month, chosen by the person

4. 04

   Adapt

   The month broke, so the plan changes, not the person

5. 05

   Continue

   Next session opens on the last commitment, kept or not

6. 06

   Debt free

   The date arrives, and the habit outlasts it

A positive psychology coach should be positive about human possibility, not compulsively positive about every human experience.

The escape route, localised

The payoff engine is the easy part. The judgment is knowing which layer someone is on, because avalanche versus snowball is irrelevant while the balance is still growing.

At 39% to 44%, choosing snowball for the feeling over avalanche for the maths costs more here than in any American guide. So the product leans avalanche, and clears a trivial balance only where that releases a minimum payment or closes a card. This is the order Ren works down, before any payoff method matters.

1. 01Stop creating new principal
2. 02Separate a survival problem from a spending problem
3. 03Protect a small anti relapse reserve
4. 04Reduce the price of the debt
5. 05Choose a payoff order the person will actually execute
6. 06Free up cash flow
7. 07Defend against relapse
8. 08Rebuild reserves, then rebuild wealth

• The rights layer

  The counselling obligation here sits on the creditor, not a free charity. A genuinely useful right, attached to a genuine conflict of interest. Use it, knowing who the counsellor works for.

• The riba layer

  User triggered by design, never a setting and never a default. The moment someone says avoiding interest matters to them, the plan moves fully into that frame. Ren describes what a contract does, and sends rulings to a scholar.

Qatar: the same habit at a third of the price

Sproutjar was researched in Dubai, where a card costs 39% to 44% a year. Doha is the second market, and the first thing the research had to do there was refuse to carry the UAE numbers across.

Every Qatar issuer I read publishes the same headline: 1% a month on a revolving balance, which the product shows as 12% a year. Doha Bank writes that the figure follows Qatar Central Bank guidelines, and I have left it at that rather than quote a directive I could not read in the original. The detail underneath is where the money is. Commercial Bank charges 2.25% a month, 27% a year, when your salary does not land at the bank. Cash advances run from day one at every issuer with a 3% to 4.5% fee on top. QIB, an Islamic bank, has no late fee at all: a late payment triggers a 2% commitment to charity, which is a different thing to explain to a person than a penalty. And a QIB minimum is 5% of the balance against Commercial Bank's 3%, so two people with the same debt have different floors.

What does not change is the behaviour. The plan still has to be built from the person's own salary, payday and obligations, the minimum is still a constraint rather than the ambition, a windfall still needs a rule decided before it lands, and a balance transfer still has to pass the same four tests. What changes is the arithmetic Ren speaks and the screen shows: the currency, the rate, the fee schedule, and which bank's official page the live retrieval is allowed to read. That switch already exists in the code, keyed on the person's country, and the table below is what it reads for Qatar.

• QNB

  Retail tariff, credit cards

  Rate

  1% a month, 12% a year

  Minimum

  Not stated in the tariff

  Grace

  None on cash advances: interest runs from the transaction date until repaid in full

  Late

  0.25% a month; over limit QR 100

  Cash

  1% a month plus a 4% fee, minimum QR 10

  QNB tariff of charges, Qatar · read 2026-09-05

• Commercial Bank

  Schedule of charges, all card packages, salary transfer customers

  Rate

  1% a month, 12% a year. Without salary transfer: 2.25% a month, 27% a year

  Minimum

  Classic product page: repay from 3% of the balance

  Grace

  Up to 45 days interest free on the Classic product page; the tariff counts a maximum of 15 grace days after the statement, none on cash advances or fund transfers

  Late

  0.25% or QAR 100, whichever is greater; over limit QAR 150 once 2% over

  Cash

  1% a month plus 4% (Diners) or 4.5% (Visa, Mastercard)

  Commercial Bank personal banking tariff of charges · read 2026-09-05

• Doha Bank

  Charges for personal customers, all consumer credit cards

  Rate

  1% a month on all transactions, 12% a year, stated as per QCB guidelines

  Minimum

  Not stated on the charges page

  Grace

  Not stated on the charges page

  Late

  QR 100 per missed payment or 0.25%, whichever is higher; over limit QR 100

  Cash

  4.5% of the amount or QR 20, whichever is higher. Fast Cash instalments at 0.75% a month, 9% a year, over 6 or 12 months

  Doha Bank charges for personal customers · read 2026-09-05

• QIB

  Visa Gold, Platinum, Infinite and Digital revolving cards, Islamic structure

  Rate

  1% profit a month on the unpaid balance, 12% a year

  Minimum

  5% of the balance

  Grace

  Up to 45 days when the statement is paid in full

  Late

  No late fee: a 2% charity commitment on the unpaid amount, sent to charity

  Cash

  3% withdrawal fee; listed as not applicable on the Digital card

  QIB Visa Platinum credit card, fees · read 2026-09-05

Read off each bank's own page or tariff on the date shown, not from a comparison site. Banks change these without notice; the figure in a live call comes from the same retrieval, with its own source and time.

Refuse to carry the UAE numbers across.

Why this is voice first, and not another chat window

A text agent answers questions. That is not the problem here. The problem is a person who already knows the answer and cannot keep doing it, and the thing that moves that is a relationship with something that talks back.

Three reasons voice earns its place. Saying a number out loud is a different act from typing it: people admit the real balance, the loan they left out of the spreadsheet, the month they hid, because speech is confession and a form is data entry. Tone carries what the text never sends, so when someone flattens out on the sentence about their family, Ren has a signal a chat box does not get and can slow down, ask, or drop the agenda for the one that matters. And voice has a personality: pace, warmth, the pauses. You do not build accountability with a text field, you build it with somebody who is expecting you.

Ren is an ElevenLabs conversational agent joined over WebRTC from the app, with a signed session minted server side so no key reaches the browser. The prompt is written from the coaching research, not from a persona sketch: permission before challenging, teaching or suggesting, reflection in the person's own words, one commitment at the end. Turn taking took as much work as the words, because the default failure of a voice agent is interrupting somebody who is thinking; the silence timeout is long, and Ren never asks whether you are still there. Seven server tools point at this app's own API routes, so mid call Ren can read live balances, run the payoff plan, log a commitment, capture a belief and file an offer.

I designed around the honest limits rather than through them. The research verdict on voice as a modality is unsupported, and no persona wanted a forty minute scheduled call, so the person picks the length and the agenda before the call, nothing is scheduled, quiet mode lets someone whisper on a Metro platform and read the reply on screen, and every number Ren says is also on screen behind it. If a real cohort tells me voice is dead weight, the journey survives without it: Ren can also be typed to, in the same room with the same memory, and before anyone's first call a one-time card says what Ren does and does not do.

• 

  The person sets the contract before the call: how long, and what would make it worth it. Quiet mode sits next to it.

How I actually worked: research, brief, PRD, flows, then build

The order was deliberate and it is the part I would defend in a review. Research fed a positioning brief, the brief fed a PRD of forty two numbered features, the PRD fed the user flows, and only then did anything get built.

The brief came first because it decides what the product is allowed to be, down to the language that is banned rather than discouraged. Success criteria in the PRD are written behaviourally so they can be failed: read ten of Ren's lines aloud and none should sound like an assistant.

1. 01

   Research

   Skeptical advisor prompt, sources kept in separate buckets

2. 02

   Brief

   Who it is for, the method, the legal posture, the banned language

3. 03

   PRD

   42 numbered features, unhappy paths weighted the same

4. 04

   Flows

   Onboarding, session, post call, relapse

5. 05

   Build

   Live balances, seven server tools, a working call

42

Numbered features, unhappy paths weighted like the happy ones

4

User flows mapped before a screen was drawn

0

Streaks, badges, points, feeds, text chatbot

• Flow 01Onboarding

  Opens on what the money is for. The first exchange may not ask for a number.

• Flow 02The session

  A contract before a conversation: length, intent, and an agenda proposed from what changed.

• Flow 03Post call

  The commitment is written in the person's own words and filed, or the app says nothing landed.

• Flow 04The month that broke

  No red, no restart. A re plan, and the next session opens on the last commitment.

Do not compete on conversation. Compete on continuity.

The build, and the plumbing that makes it a product

Next.js and TypeScript, Convex for live persistence, ElevenLabs over WebRTC for the call, seven server tools wired to this app's own API routes.

That plumbing is the difference between a voice demo and a product. When Ren says the rent on your debt is one thousand three hundred and twenty eight dirhams this month, the number came out of the same database the screen is reading, and the payoff engine behind it runs month by month amortisation over every card at its own rate.

The demo I judge it on is the balance transfer. Mid call I asked Ren what to do about the Emirates NBD card. It read the live balance and the 3.69% monthly rate out of Convex through its debt snapshot tool, worked the transfer against the fee and the promotional window, said the annual figure out loud rather than the monthly one, and proposed it. The offer appeared as a card in the call and filed itself on the Plan screen, unapplied, because Ren cannot move money and will not record transferred principal as progress. The person confirms it after their bank has actually done it. A coach that acts on a mishearing is worse than no coach.

That offer is now checked live. Context.dev is wired server side: Ren's tool searches the web for the bank's own pages, reads the published terms off the official page with fact checking on, and returns the promotional rate, the window, the revert rate and the fees with the source URL and the retrieval time attached. The key sits on our server, never in ElevenLabs and never in the browser. When retrieval fails, Ren says so in the call rather than reading the dated table out as this morning's rate.

A second audit, run against the research rather than the screens, changed the shape of the month. The plan now runs on a monthly amount derived from the person's own salary, payday, essentials, obligations, money sent home and known lumps, with their own figure allowed to override it and the screen saying which is in use. Payday brings a review in four numbers: debt at the start of the month, new borrowing, principal repaid, interest burned, with what each card received and one reflection in the person's words, and the date recomputes from that. Cards carry a due day, a statement day, a minimum and a type, BNPL included, and a statement can be pasted in when memory is not enough. Every rate is shown as the simple twelve times the monthly figure, because that is the only annual number a Gulf bank's own paperwork supports. And the four tabs read Today, Ren, Plan, You, in the person's words rather than mine.

• 

  Today, then called Home. The plant, the date, and the interest named as rent, the number nobody quotes you when you take the card.

• 

  Plan, then called Growth. Every card at its own rate, what it costs this month, and what has already gone. One card marked Start here.

• 

  You. Their reason, their own past success, the beliefs they have said out loud, and now the numbers Ren works from, all in their words and editable.