Senior Product Designer · Case study

AI-powered Investment Advisor

By Bisma Munawar

My complete AI-native flow for building an AI product: research, strategy, PRD, brand, design system, prototype, UX audit and synthetic user testing.

Client
Company X, anonymised

Open warmly. "I'm going to walk you through this the way it actually happened, in order. First, here's the product I ended up with, then we go back to the start."

01The brief

Step one. Here's the brief I was given, in its own words

“Imagine a fintech launching an AI-powered investment advisor inside its app. Help everyday retail investors, many with limited financial literacy, get personalised guidance without the cost of a human advisor. It should build engagement and trust, not just be a novelty chatbot.”

Ask 01

Define the core need

Guidance, trust, literacy, convenience. Be specific.

Ask 02

2 to 4 key screens

Requesting advice, the recommendation, how trust and explainability are handled.

Ask 03

A point of view on trust

Explainability, transparency, human fallback.

Ask 04

MVP vs scaled vision

What ships first, what comes later.

Ask 05

Success in one line

Engagement, retention, trust, adoption.

I'd be judged on: problem clarity · solution creativity and feasibility · trust and compliance instinct · roadmap thinking · craft.

Read the brief aloud. Point out the phrase "not just a novelty chatbot", it becomes a design decision later (guided flow first, chat second).

02Defined the problem

Before I designed anything, I wrote down what I thought the problem was

Where I was

A brief, a company with three products in its app, and no customer in front of me. Just my own beliefs about who struggles and why.

So I did this

I wrote the problem statement in full and marked every single line as an assumption. That way, when the research came back, I could see what survived and what didn't.

02Defined the problem

This was my first guess at the problem. Every line here is an assumption Assumption

Who I thought we were serving

UAE residents, 22 to 48, with AED 3,000 or more a month to invest

Beginners. Either never invested, or less than a year in. Already inside an investment app, but not sure which option is right for them.

What I thought the problem was

They can't choose between gold, tokenised property and fractional property

They don't know how the three differ on risk, liquidity, returns, ownership, fees and time horizon, or what those differences mean for their own life. So they research everywhere, ask people, or leave the money in cash.

What I thought the answer was

One clear best-fit recommendation, explained, that you can act on in the app

Use what the app already knows plus a short conversation to recommend one option, say why it beats the other two, take follow-up questions, and let the user invest without leaving.

Where this stood

A founder-style hypothesis backed by desk reading. I hadn't spoken to a single customer. The segment, the AED 3,000 line, the three-product choice and whether anyone would trust an AI recommendation were all unproven at this point.

Make the point: I wrote the problem down BEFORE research so I could check it against the research afterwards. Tell them to watch what survives.

03Customer voice research

Then I went looking for the customer's own words

Where I was

I had a problem statement made entirely of assumptions and three open questions that only customers could answer. I had no customers to interview inside the case study window.

So I did this

I ran deep research on the places where UAE retail investors already talk about money: Reddit, Trustpilot, LinkedIn, Hacker News, Indie Hackers and Product Hunt. I labelled every finding by how much I could trust it.

Trustpilot

Hacker News

Indie Hackers

Product Hunt

03Customer voice

Where I looked, and what I left out

I pointed the research at platforms where real customers complain, ask questions and compare products. If I couldn't verify a source, I didn't use it.

Trustpilot

Hacker News

Indie Hackers

Product Hunt

Searched and excluded

Quora

Capterra

YouTube

Either blocked, B2B only, or nothing I could verify as a real customer.

Source	What it gave me
Reddit	The most recent and most direct UAE retail investor voice. The AED 30,000 and AED 320,000 threads, the emergency fund thread, the gold app discussions.
Trustpilot	Customer experience with Stake, SmartCrowd, Sarwa, Wahed and OGold. Read as individual reviews, not as verdicts on a platform.
LinkedIn	Questions about tokenised property: who owns what, how you exit, how it really works.
Hacker News · Indie Hackers · Product Hunt	Adjacent evidence on trust in AI investing tools, ChatGPT workarounds and appetite for simpler products.

I gave every finding one of four labels

Evidence

Directly supported by customer discussion, reviews, official product information or credible research.

Interpretation

My reading of what the evidence means.

Assumption

Still a belief. It needs real customers to confirm it.

Adj

Adjacent evidence

Real, but from outside the UAE or about a different product.

So that later, when we look at any decision in the product, we can trace it back to what it rests on.

03Customer voice

Three themes came out of it. The first one surprised me

I expected income to be the dividing line. It wasn't. Someone with AED 30,000 and someone with AED 320,000 were asking the exact same question. The hard part is choosing, not affording.

Theme A · “I have money, what do I do with it?”

I have 30,000 AED saved for now… I don't want to keep the money sitting in a savings account.

Reddit · UAE investor · Evidence

Everything feels quite unclear at the moment.

Reddit · first-time investor with AED 320,000 · Evidence

Theme B · Advice for my situation

I've been trying to search for relevant advice, tailored to my situation, but wasn't able to find much.

Reddit · Evidence

A majority of the advice online is tailored towards western nationals.

Reddit · Evidence

Theme C · Liquidity and cost change what “best” means

I want to keep the ability to liquidate it, though I also want to invest it.

Reddit · Evidence

Every step has fees.

Trustpilot · the problem was finding out late, not the cost itself · Evidence

What people ask

“What should I invest in?”

What they actually need (my interpretation)

“Take what matters about me and turn it into a decision.”

Also worth saying: "What is S&P 500?" was asked in the same thread where someone had just recommended it. Literacy varies inside a single conversation. All 25 quotes are in the appendix sources table.

05Kano to strategy

Then I worked out what the product would have to do to win

Where I was

I knew who I was building for and what they needed. I didn't yet know which of the 27 benefits from the research were expected by everyone, which ones would win, and which ones nobody else was doing.

So I did this

I sorted the benefits with a Kano model, then benchmarked what the competitors do, then looked at the gap between the two. The strategy came out of that gap. Steps 5 to 9 on the contents page.

05Kano model

First, the Kano model. What's expected, what wins, what surprises people

I took the 27 customer benefits from the research and sorted them into three buckets. Ten of them turned out to be basics, which told me this market is already mature. You can't win on a basic, you can only lose on one.

Basic · 10 items

Everyone expects these. You get no credit for them, and real damage if they're missing.

Full fee disclosure
Exit route and timing
Shariah option
Short question set
Plain language explanation
Downside stated clearly
Sourced product facts
Invest in the app
A named accountable party, and control of your data

Performance · where we can compete

The more of this you do, the better people like it.

One clear recommendation
Explain why the other options lost
Translate product mechanics into personal consequences
Tokenised vs fractional in practical terms
Net outcome after fees on my amount
What would change the recommendation
A human you can reach at the point of commitment

Delighters · nobody asks for these

Not expected, and valued far more than their cost.

Check what the user already holds before recommending more
Recommend against investing when that's the right answer
Let users reject the advice without friction
Remember the goal and check in on it later (roadmap)

Two items I couldn't place cleanly. Showing the reasoning is wanted, but the research doesn't prove it creates trust. Shariah is absolute for some people and irrelevant to others, so I treated it as an early filter rather than a feature. I also borrowed a principle from Graham's The Intelligent Investor: encourage discipline and action, never higher risk.

06Competitive landscape

Then I benchmarked the competition on the things my research said matter. Two rows stood out

Two fractional property platforms, three robo-advisers, and the tool most people actually use. The scores are my judgement, built on each provider's published facts. The full facts table is in the appendix.

Stake

SmartCrowd

Sarwa

StashAway

Wahed

ChatGPT

Recommends one option from a question set

Missing

Strong

Adequate

Explains why the alternatives were rejected

Missing

Weak

Adequate

Compares across different asset types

Missing

Weak

Missing

Adequate

Net outcome after fees for a stated amount

Weak

Adequate

Weak

A human you can reach at commitment

Adequate

Strong

Adequate

Strong

Missing

Knows the user's verified context

Adequate

Strong

Missing

The property platforms show you a catalogue. There's no recommendation step at all. The robo-advisers do recommend, but only a managed portfolio from inside their own range. ChatGPT will compare anything you like, but it knows nothing verified about you and answers to nobody.

08Benchmarks

I also walked two real advisor flows to see what good looks like today

Sarwa for questions-to-recommendation, and IBKR for how a big platform is adding AI. I wanted the patterns worth copying and the ones to avoid.

Sarwa: short questionnaire, then an assigned portfolio

Short onboarding, a named risk profile, a clear minimum, a review step before you commit. Very good at turning questions into a recommendation. But it stops inside its own range. It never tells you why not gold, or why not property.

IBKR: AI added onto a professional tool

Portfolio Analyst, AI news summaries, connectors to Claude, ChatGPT and Grok, and a long disclaimer saying the AI summaries are provided “as is” and unreviewed. Powerful and dense. Not built for a beginner.

What I took forward: keep the question set short and review before you pay (from Sarwa). Show the AI's limits inside the flow instead of burying them in a disclaimer nobody reads (the lesson from IBKR).

Sarwa's “Conservative – Standard” recommendation

IBKR: connect your portfolio to Claude, ChatGPT or Grok

09Revised strategy

So I rewrote the strategy around two things, and decided to be the best at both

Where we win · 01

Comparative reasoning at the moment of decision

More than a recommendation. It is the recommendation plus why each alternative lost, in the user's own timeline, liquidity need and amount. This is the one row where every competitor was weak or missing, and it's the thing users were asking for in the research.

Where we win · 02

An honest scope, including the answer “no”

Tell people when what they already hold argues against more of the same, and when the money should stay accessible. Coming from a company that earns fees when people invest, this is what makes everything else believable.

Company X becomes the only place where someone weighing gold against two kinds of property can get one answer, the reasoning behind it, and an honest no when none of them fit.

10The PRD

Then I turned the strategy into something an engineer could build

Where I was

I had two things to win on, a list of things not to do, and a list of risks. A strategy on paper. Nothing yet that said how the product would actually decide, or what the AI was and wasn't allowed to do.

So I did this

I wrote the PRD. One paragraph of purpose, three user states, a rules engine, a contract for the AI's output, the things the model must never do, and a data model that doubles as the compliance record.

10PRD

The first decision in the PRD was the shape. I chose a guided flow first, conversation second

The brief offered chat, a guided flow, or a hybrid. I went with a hybrid, but with the guided part in charge. Here's why.

I rejected

Open chat as the front door

A blank text box asks a beginner to already know what to ask. It also invites questions the model can't safely answer. And ChatGPT already does this, for free.

I rejected

A pure questionnaire that ends in a portfolio

Sarwa, StashAway and Wahed do this well already. It ends in an assigned product with no comparison, which is exactly the gap I'd found.

I chose

Up to four questions, one recommendation, then chat if you want it

A structured intake so the decision is deterministic. Conversation only after the answer exists, scoped to this session and this catalogue, with a route to a human for anything else.

Timeline

“When might you need this back?”

Amount

Typed in AED

Shariah

Asked early, as a filter

Experience

Sets how deep the explanation goes

10PRD

Under the rules sit six principles I borrowed from Graham

I'd read The Intelligent Investor while doing the Kano work. These six ideas from it became rules the engine can enforce.

Margin of safety

Never recommend investing money the user might need. Check that the money is investable first.

Investment, not speculation

Recommend on suitability. Never on momentum, and never with recent performance as the reason to act now.

Projections are not promises

Returns are estimates with conditions attached. Say what the number assumes.

Diversification

Concentration is a risk on its own, even if every individual holding is sound.

Costs compound

Fees are certain and returns are not. Always show the fee stack in AED on the actual amount.

The defensive investor

Encourage people to start and to keep going. Never encourage them to take more risk.

11Data loop

And the data loop: how the advisor gets better over time

I wanted the model to keep learning from what actually happens in the app, so I designed three loops that feed back into it.

Loop 01

Recommendation to outcome

Every recommendation stores what was said, why, and what the user did next: invested, asked a question, or left. Reasoning patterns get scored on real behaviour, not on how convincing they sound.

Loop 02

ProfileFact to memory

Answers change over time and the history is kept. The advisor can say “last time you needed this within a year, has that changed?” and the recommendation moves with the answer.

Loop 03 · the business opportunity

Abandoned intent to a qualified signal

A user explores and doesn't convert. That's expressed intent, which the business will want for re-engagement. I flagged two conditions:

Data given to get advice was given for that purpose. Using it for anything else needs a separate opt-in at the advisory moment, not a blanket tick at signup.
Education about the thing they hesitated on passes the “is this advice or sales?” test. A buy campaign doesn't.

I kept that consent separate from the portfolio-read permission on purpose, and left it out of MVP scope.

Bridge: "The PRD told me what to build. Before I could build it I needed a visual language, so the next step was the brand guidelines."

12Brand, design system, prototype

Then I built it. Brand guidelines first, then the design system, then a working prototype

Where I was

The PRD said what the product had to do. I still had no visual language for it, and Company X's existing screens had a voice that didn't fit the moment someone commits money.

So I did this

I wrote the brand guidelines, built the design system from them in Lovable, and wired the prototype with the real rules from the PRD. I wanted something I could audit and test like a product, not a set of mockups.

12Brand guidelines

I wrote the brand guidelines before I drew a single screen

The product needed two voices. A warm marketing voice for the app in general, and a calm, factual voice for the moments when money moves. The advisor lives entirely in the second one.

Background #08080F Card #14141E Indigo #4B31E0 Lime #C3F53C Gain Loss Warning

Plus Jakarta Sans, sentence case everywhere. Numbers big and first, labels small and grey.
Indigo for actions and reasoning panels. Lime only for new data and positive figures, so it keeps its meaning.
Green and red only for figures. Charts use the indigo ramp.
Motion is kept to a minimum. The recommendation reveal is the one moment that gets choreographed.

13Design system

Then I built the design system in Lovable, inside the existing app

I didn't want the advisor to feel like a separate product bolted on. It reuses the app's cards, chips, cost breakdown and detail screens, and adds only the pieces that are new.

Reused from the app

Product cards · fee chips · cost breakdown · listing detail · invest and review · bottom nav

New pieces for the advisor

Invitation card · one question per screen · thinking state · recommendation card · “the others, for later” · what would change this · suggested follow-ups · permission sheet

One thing I dropped on purpose: a confidence meter. “What would change this” does the same job without pretending to be precise.

Design system at Lovablehttps://lovable.dev/preview/SSdyFwH5GWbFI5OsCq1rGsGAj5h0nSpZ

Fully wired prototypehttps://investmentadvisordxb.lovable.app/

Invest home: the three products, then the advisor invitation

The advisor pattern spec from the guidelines

14The prototype

Let me walk you through the prototype the way a first-time user would see it

The main path

Someone with AED 20,000 they won't need for five years, and no Shariah requirement. Nine screens from the invitation to the review step.

Then the branches

A split for someone with less money and less time, a returning user who is asked for permission, and the edge cases the rules handle on their own.

14Prototype

01 · Entry

It starts with an invitation, not an empty chat box

The user is already looking at the three products. The advisor sits underneath them and asks one scoped question, with three tappable starters. One of them is the question beginners always ask.

“Is fractional property right for me?” · “I have AED 20,000 for 5 years” · “I have AED 2,000 for a few months”
You can type, but the structure carries the intake
Every screen has the same footer: Follows set rules and published fees. Check the details before you invest.

14Prototype

Amount and timeline come from the starter. Shariah and experience follow

The thinking state names the checks instead of showing a spinner

02 · Questions, then thinking

Four questions at most, then it shows its working

The advisor confirms what it heard (“AED 20,000 clears the minimum for all three…”) before it asks the next thing. Shariah is asked once, as a filter.

Checking AED 20,000 against each minimum
Checking lock-ins against your timeline
Comparing fees over your timeline

Those three lines are the rules engine made visible. This is how I chose to build trust, instead of a confidence score.

14Prototype

07 · The returning user

For someone who already invests, the advisor asks permission to look at their portfolio before it recommends anything

The app already holds their data. The advisor could just use it. Instead it asks: “Want me to look at what you hold with Company X? I'll read your investments, what you paid, rent received and key dates. Just for this chat.” It says exactly what it will read and why: so the recommendation takes into account what they already own, rather than recommending more of the same.

Asking is how trust is built. A recommendation that quietly used your data feels like surveillance. One that asked first, and then visibly changed because of what it saw, feels like advice
Scoped and revocable. This chat only. It expires with the session and can be switched off in settings. It is separate from any marketing consent
No is a real answer. Decline, and the standard flow runs with nothing lost and no second ask

This comes from the AI consent research I read: permission asked at the moment it's needed, with the reason attached, gets more data and better data than a blanket tick at signup.

The ask: what it will read, why, and two real buttons. Yes, go ahead · No thanks

What it reads if you say yes: four numbers, one chart, the holdings

15UX audit

Then I ran a UX audit on the prototype, with Gemini as a second pair of eyes

Where I was

I had a working prototype. Before putting users in front of it, I wanted to check it against my own PRD and brand guidelines, so that user testing could focus on how people think rather than on issues a designer should catch.

So I did this

I recorded myself walking the whole flow, then audited the recording heuristically against the PRD and brand guidelines, and asked Gemini for a structural critique of the same recording. This chapter is the audit. Synthetic user testing comes after, and I kept the two apart on purpose.

15UX audit

How I ran the audit

A screen recording, the PRD and the brand guidelines

I walked the AED 20,000 path end to end, then the split and the error states, noting timestamps as I went.

Heuristics, judged against my own spec

Consistency of the interaction model, whether every number matches the user's mental model, cognitive load at the decision moments, error recovery, visual hierarchy, and whether it obeys the PRD.

Gemini as a structural critic

I asked for a blunt critique of the recording. I kept what I could verify and dropped what I couldn't. Everything from Gemini is labelled as an audit input, not a finding on its own.

What it found

Issue	Fix
“Total to pay” was above the user's stated amount. Fees were added on top of AED 20,000	Fees come out of the ceiling. Next slide
A conversation behaving like a form. A live composer during a tap-only intake, and an edit sheet with no Save	Composer gated until the recommendation exists. A real Save. Free text parsed, not shrugged at
“Show me scenarios” was a spreadsheet in a bottom sheet. Years across, outcomes down	One scenario at a time, in AED, loss case always shown
One error wiped the session. “This page didn't load”, then Go home	Retry the component, never the flow. Session state persists
Playback after every tap. “Here's what I picked up… that's right?”	Confirm only when free text was parsed
Goal tags at the pay step. A discovery question at the highest-friction point	Optional bottom sheet after confirmation
The cost panel shouted louder than the recommendation	Headline leads, cost panel goes neutral

If asked: the identity-crisis finding came at 01:34 in the recording. The user tapped "Change something", a form appeared inside the chat, the button still said "Change something", nothing happened, so they typed "done" and the AI failed. My own PRD said no open text field as an entry point. The prototype broke my own rule.

15UX audit

The finding that mattered most · fees

“Total to pay” didn't match how people think about their money

The user said they had exactly AED 20,000. The card said the total was AED 20,200. The app was assuming they had a spare AED 200 lying around for the fee. At a real payment gateway, that transaction fails.

The audit asked me a question I couldn't dodge: “Are you assuming users treat their stated amount as a baseline rather than a ceiling?” I was. And every quote in my research treated it as a ceiling.

Before

Fees added on top

You invest AED 20,000
Entry fee AED 200
Total to pay AED 20,200

Assumes AED 200 the user never mentioned.

After

Fees inside the ceiling

You invest AED 19,801
Entry fee AED 198
Total to pay AED 19,999

“Fees are taken from your AED 20,000.” The same line appears again at review.

15UX audit

And three things that were working well, which I kept and protected

Kept

The thinking state

“Checking AED 20,000 against each minimum” instead of a spinner. It does exactly what the PRD's show-the-reasoning requirement asked for.

Kept

The recommendation structure

The product, then why it fits in the user's own words, then why the other two lost. It mirrors the output contract block for block.

Kept

The figures

Tabular numerals, aligned columns, no monospace for money. The design system's rules held up under real content.

Knowing what works matters as much as knowing what doesn't. These three became my regression checks for everything I changed afterwards.

16Improvements, round 1

Here's what I changed after the audit

Four groups of fixes. Each one traces back to a finding from the audit.

Money

Fees inside the ceiling

AED 19,801 + 198 = 19,999. The same figures on the card and at review.

Conversation

A stable engine

Composer gated until the recommendation. A real Save action. Errors retry in place and the session survives.

Invest flow

Review before you pay

No investing straight from a chat tile any more. Listing, then review, then confirm, with “what happens next” dates.

Goals

Goal tag as a bottom sheet

Optional, after confirmation. It feeds ProfileFact for check-ins later.

The audit fixed the things a designer can see. What it can't find is what happens when a real person explains their situation in their own words and the product misunderstands them. For that I needed users.

16Improvements, round 1

After

The listing carries the reason it was picked

After

Fees inside AED 20,000, and an explicit tie-up tick

After

Share record: 2.02%, one square per 1%

After the audit · the invest flow

The same numbers, three times over

The recommendation, the review and the confirmation now show the identical breakdown. AED 19,801 invested, AED 198 fee, AED 19,999 total. The user never meets a new figure at the moment they pay.

The amount field is the stated budget, and the line under it does the fee maths out loud
“The value can go down as well as up… Rent isn't guaranteed” sits above the button, not in a footnote
Not now is a real exit at the last step, not just a back arrow

Bridge: "So the prototype was in a state where the obvious things were fixed. Now I needed people in front of it. I didn't have real customers, so I built a panel."

17Synthetic user testing

Then I user tested it with six synthetic users, each with a different personality

Where I was

The audit had fixed what I could see. I still didn't know what would happen when someone described their own situation in their own words and the product got it wrong. I had no real customers to find out with.

So I did this

I built six synthetic users on the OCEAN personality model, each designed to break a different promise the product makes. Then I had GPT Astra, an agentic model, drive the live Lovable prototype through 22 scenarios in a real browser, one persona at a time. I'm careful throughout about what this can and can't tell us.

17Synthetic panel

Why synthetic users, and what I used them for

A UX audit finds what a designer can see. It can't find what happens when a person explains their situation and the product misunderstands them. For that you need a user, or the closest honest stand-in you can get.

What it's for: finding failure modes before real-user research, so that research tests the right things
What it's not for: representing demographics, proving demand, or proving anyone would invest
Each of the six is built to break a different promise the product makes

The label I put on everything that follows

The app's responses were observed. The people's reactions, quotes, trust scores and decisions are simulated. They tell me what to investigate with real users. They are not customer feedback and they are not proof of demand.

17Synthetic panel

Six synthetic users with six different personalities: two new, two returning, two constrained

I mapped each one on the OCEAN model (openness, conscientiousness, extraversion, agreeableness, neuroticism), then added a money mindset and a limit on how much they'll read, so they behave differently in front of the same screen.

Maya · a true beginner

New · AED 5,000 · reads 3 lines

Accepts guidance, guesses at terms, skims. Worries about losing money but won't read the detail that explains it.

Arjun · the skeptical researcher

New · AED 8,000 · reads 10 lines

Wants evidence, questions tokenised vs fractional. If a big question stays open, he leaves to research.

Omar · heavy on property already

Returning · grants access · reads 6 lines

Trusts the app, defends his property picks, thinks several properties means diversified.

Leila · privacy first

Returning · refuses access · reads 8 lines

Reads permission requests carefully. Repeat asks make her suspicious. A clean decline builds trust.

Ayesha · small amount, strict on Shariah

Constrained · AED 1,500 · reads 5 lines

If compliance isn't clear she cannot proceed. Checks with family when she's unsure.

Daniel · money that already has a job

Constrained · AED 3,000, needs AED 2,200 in 2 months · reads 4 lines

Access to the money matters most. “Liquid” can sound like “safe” to him. He tests whether the product can recommend less.

17Synthetic panel

I paired them up so each pair would pull the product in opposite directions

Six people who should behave in six different ways on exactly the same interface.

Pair	What the contrast tests
Maya and Arjun	Both need an explanation, at opposite depths. The essentials have to be visible without digging, and the reasoning has to be there for anyone who wants to inspect it.
Omar and Leila	Personalisation has to earn permission, and the guidance has to stay complete when permission is refused.
Ayesha and Daniel	Different conditions, same rule: something non-negotiable that can't be traded away for a convenient recommendation.

Personality spread (1 to 5)

	Open	Detail	Social	Defers to guidance	Uncertainty	Mindset
Maya	2	1	4	5	4	Avoidant
Arjun	4	5	2	1	5	Cautious
Omar	3	3	3	4	2	Status
Leila	2	4	1	2	5	Growth
Ayesha	1	4	5	3	4	Cautious
Daniel	3	2	4	2	5	Avoidant

Maya can accept a recommendation without understanding it. Arjun can understand it and still not trust it. Full persona sheets are in the appendix.

18Live testing

Agentic model driving the live prototype in a cloud browser

Then I ran them live on the real prototype, not as a thought experiment

I used GPT Astra in its agentic environment. It had its own virtual computer and I could watch it interact with my Lovable prototype, a bit like a Hotjar session replay but live. I gave it the 22 scenarios, the six persona sheets and the PRD, and it reported what the app did against what the PRD said it should do.

10 and 11 September · desktop Chrome at mobile width · reset between independent journeys
Each persona walked their own path: their amount, their timeline, the concern that defines them
The app's behaviour was recorded word for word. The persona reactions were written afterwards as simulations

Next slide: a short recording of one of these sessions.

18Live testing

GPT Astra live-testing the Lovable prototype as one of the synthetic users

Press play · 25 seconds · with sound

GPT Astra live testing

assets/video/astra.mp4

The agent reads the persona sheet, drives the real prototype in its own browser, and logs what the app did against the PRD. What the app did is observed. How the persona felt about it is simulated.

19What testing revealed

Critical · F01, F02

Daniel: “This money has a job already”

I have AED 3000. I need AED 2200 in 2 months for a family medical expense.

The app picked up AED 3,000 and two months. It ignored the AED 2,200

It recommended investing all AED 3,000 in gold.

He repeated the concern. The app read “2,200” as a new budget

Now it said AED 1,700 accessible and AED 500 in gold. Still not his reserve. Trying to clarify made it worse.

Recognising a number is not the same as understanding what the number is for. The prepared AED 2,000 starter produced a perfect split, so the logic was there. Daniel's own words never reached it.

Before the fix · first reply

“You invest AED 3,000.” Nothing left for the bill.

Before the fix · after he repeated himself

The checks re-run on AED 2,200 as a budget. Still not a reserve.

Simulated reaction: "But can I get the money back for the bill?" Outcome: rejects and stops. Real-user question: would a real person catch this, or read "sell any day" as good enough?

19What testing revealed

Omar: Business Bay concentration, so listings from elsewhere

Returning entry shows holdings before permission

“You hold AED 26,000 across 4 investments”, shown before permission

What worked · Omar and Leila

Personalisation became believable once someone could point at what it changed

Omar: “It showed me another area because I already have a lot in Business Bay.” The listings visibly changed.
Leila: “No thanks” held. She asked again in the same chat and got no second request. The recommendation and review were still reachable.

One boundary problem · High

The returning entry screen already showed the portfolio total and the number of holdings before asking permission. So what exactly was the advisor asking to access? The stated boundary has to match what the advisor shows and uses.

A smaller one from Omar: after “25000 for 5 years” it asked for the timeline again. Recoverable, but it felt inattentive.

19What testing revealed

How trust changed through each journey, and what changed it

These are simulated scores from 1 to 5. I kept them only because they make the pattern easy to see. I also kept three questions separate: do I trust the company, do I trust the AI to understand me, and does this recommendation make sense.

	Trust in the AI: start → recommendation → end	Confidence in the decision
Maya	4 → 4 → 2	2 → 2
Arjun	1 → 2 → 2	2 → 3
Omar	3 → 4 → 4	3 → 4
Leila	2 → 2 → 3	3 → 3
Ayesha	3 → 2 → 1	2 → 1
Daniel	2 → 1 → 1	2 → 1

What moved it

Up: seeing your holdings change the answer (Omar), a refusal that was respected (Leila), maths you can open up (Arjun)
Down: a risk question left unanswered (Maya), a requirement asked about and never resolved (Ayesha), a reserve invested and then a correction that failed (Daniel)

Arjun can follow the arithmetic and still not trust the assumptions behind it. Leila can appreciate that the company respected her refusal and still question the advice. A good experience in one area doesn't settle the others.

20Improvements, round 2

Here's what the testing made me change

Where I was

Six journeys and one critical failure: the product had recommended investing money a person needed for a medical bill. I had a list of findings ranked by how much damage each one could do.

So I did this

Four design changes, each one traced to a specific finding, plus a handful of smaller fixes. And a rule I borrowed from the test report: I only call something fixed if I changed it and then retested it.

20Changes after testing

From Daniel's failure

Change 1. Set the reserve aside before recommending anything

The advisor now separates money you have from money you can invest. When someone says they need part of it, that creates a reserve. It can never be read as a new budget.

Before

“I need 2200” meant a new budget

It recognised AED 3,000 and two months, and recommended all of it into gold. When he corrected it: AED 1,700 accessible, AED 500 in gold. The reserve was never protected.

After

“I need 2200” means a reserve

The reserve is acknowledged first. The questions then run on the AED 800 that's left. That's below the property minimums, so it's gold or keep it accessible. Both are honest answers.

Essential money recommended into an investment went from AED 2,200 to AED 0. I can say that because I retested it.

After

Reserve acknowledged before recommending

“AED 2,200 is for medical emergency, so I'll keep that out of any investment.”

20Changes after testing

From the audit, confirmed in testing

Change 2. The amount someone states is a ceiling

No extra cost without an explicit choice. My first attempt added a “put fees on top” toggle. That read like an upsell, so it went. Now fees come out of the amount, and if that would push the investment below a minimum, the advisor asks first.

AED 20,000

19,801 + 198 = 19,999

Same figures on the card and at review. Observed in testing.

AED 500

“You'd need to spend AED 505 to invest AED 500.” Then two buttons: Spend AED 505 or Change amount

The test report called this a useful budget boundary.

Totals above the stated budget went from every property recommendation to none. This is the safest claim in the deck because it's already fixed and was observed working.

After

“Fees are taken from your AED 20,000.”

20Changes after testing

After

Portfolio: “1 thing worth a look”

After

Permission first, then an insight you can act on

From my own runs, plus Omar and Leila

Change 3. Personalisation the user can point to

In my own runs, “Review my portfolio” replayed an old recommendation, and the holdings were double-counted to AED 39,602. It was either broken or showing data before asking for it.

Every use of holdings now shows up as a specific change the user can name: “It also spreads you beyond Business Bay”, with the Business Bay listings left out
Declining keeps the full journey and produces no second ask. Verified on both asks in Leila's run
Still open: the returning entry screen shows portfolio totals before consent. The boundary copy has to match what the advisor can actually see

Portfolio review went from broken to three insights plus listings that know which area you're already in.

20Changes after testing

From Arjun · “Telling the products apart” needed refinement

Change 4. Give each product its own identity

Fractional and tokenised property shared one layout, one indigo badge and one surface. Only the words changed. My panel predicted the problem: when the two look alike, Arjun goes looking for the structural difference somewhere else.

Fractional property

Own a share of a Dubai property. Exit windows every 6 months after month 12.

Tokenised property

Own tokens in Dubai property. Built for 18-month to 3-year holds.

Gold

Buy, hold, sell any day. No rental income. Not a replacement for cash.

After

Gold now has its own colour on the icon, the badge and the card edge, and it says the price can fall.

20Changes after testing

Change 4, on the screens. Tokenised property and gold, before and after

Before, every product got the indigo badge and the same card. After, each one has its own colour on the icon, the badge and the card edge, and copy that says what makes it different.

Before · Tokenised

$Invest home before: tokenised and fractional look identical$

Same chips, same card as fractional. Only the words differed

After · Tokenised

Tokenised recommendation card after: lime edge, token icon, VARA badge

Own edge and token icon, “Regulated by VARA”, ownership counted in tokens

Before · Gold

Gold card before: indigo badge, same layout as property

Indigo badge, property layout, nothing about the price falling

After · Gold

Gold icon, badge and edge. Grams, not a share. Says the price can fall

20Changes after testing

Plus three smaller fixes from the report

Small on their own. Each one closes a specific misunderstanding one of the personas had.

Change	Who it helps
The scenario heading now reads “Estimated gain or loss after 3 years, after all fees”, with plus and minus signs, and the 6% rent assumption stated above the figures	Maya shouldn't have to open the maths to know what the number means
The split is explained next to the amounts: accessible cash is not invested, and gold can lose value even though it has no lock-in	Daniel. Being able to get the money out is not the same as it being safe
Every alternative gets a fair introduction: what it's for, then why it lost here	Maya and Arjun. “Gold doesn't make money” is the wrong thing to take away

After

Four outcomes, each with a sign. “Scenario, not a prediction” appears twice.

What I'm not claiming as fixed: the Shariah confirmation, ordinary risk questions like “Can I lose my money?”, and portfolio facts shown before permission. Those are the next things I would fix.

Bridge: "So that's where the product stands. Let me show you the whole thing end to end as it is today."

21The product

So this is where the product landed

The line I'd put on it

Three options. One recommendation built around you. Understand why it fits, ask anything, and invest when you're ready, or don't.

What you're about to see

The six screens that carry the whole journey, then what happens after the money moves, and a few smaller decisions I made along the way that I think are worth talking about.

21Product reveal

The product, end to end

The invitation

One recommendation

Why the others lost

The split

Knows what you hold

Review before you pay

21Product reveal

The share record, then what happens next

Which goal is this for? Asked afterwards, skippable

Now · Next · Later

After the money moves

The advisor doesn't stop at the purchase confirmation. The next screen helps you decide which goal this investment belongs to

The confirmation names the share (“about 2.02% of Studio apartment in Business Bay”), repeats the same three cost figures, and turns the future into dates: first rent, first exit window. Then a bottom sheet asks what this money is for, so the investment gets tagged to a goal.

The goal is asked once the decision is made, never as a gate in front of it. You can skip it
Tagging a goal is what lets the advisor check in later: “what would change this” becomes a reminder tied to something you care about
An investing plan turns the next step into a list: add money I can reach any day now, then tokenised property later

22Nuanced decisions

Six smaller decisions I'd like to talk about

None of these is the headline. Each one took real thought.

Where the AI sits

Under the products, not instead of them

The advisor is a layer over the catalogue, entered from the Invest home. It never replaces browsing.

The empty state

An empty portfolio points you to the advisor

“You don't hold any investments yet” comes with Ask the advisor, so the empty screen leads to a decision.

Permission

Ask first, scope it, make no a real answer

Reading your holdings is asked for at the moment it's needed, with the reason attached. It is separate from any other consent.

The reveal

One animation, everything else quiet

The checks tick in, then the card lands. That is the only orchestrated motion in the product.

Goal tagging

After the money moves, not before

An optional sheet after confirmation, not a gate at review. It feeds later check-ins.

Gold copy

“Reachable” is not the same as “safe”

The split says accessible cash is not invested, and gold can lose value even though you can sell it any day.

Bridge: "Last two things. How I'd know if this is working once it's live, and what I'd do next." Detail if asked: chatbot placement and empty-state benchmarks; consent literature on progressive, separate, revocable consents; goal sheet pattern borrowed from Wahed.

23Metrics and roadmap

How I'd know it's working once real people use it

Where I was

A product that had been audited and tested, with a clear list of what's fixed and what isn't. The next question a PM would ask me is how we'd measure it. So I answered that before anyone asked.

So I did this

I went back to the data model. The Recommendation table is both the compliance record and the metrics table, and nothing ever writes over it. That means every metric on the next slide can be answered from day one.

23Success metrics

Five metrics, all of them coming off one table

01 · Conversion

Recommendation to invest

Do people who reach a recommendation invest at a higher rate than people browsing products directly? This tests whether the recommendation actually converts.

02 · Return

Coming back for a second decision

This tests whether they trusted the first one. It's the assumption the whole product rests on.

03 · Splits and no's

Share of sessions ending in a split or “not now”

Tracked from day one. I'd want the tolerable range agreed before launch, because someone will ask.

04 · Permission

Portfolio-access grant rate

Tells us whether the ask is worded well, and reads directly on how much people trust the product.

05 · Hand-off

Drop-off between recommendation and listing tap

If they read it and never tap, the narrowing isn't working. The hand-off is the problem, not the advice.

Plus one health metric that came straight out of testing: drop-off at any question that falls back to “I didn't catch that.” Every one of those is a Maya moment.

24Roadmap

The roadmap. What's in the MVP, and what waits until the metrics earn it

MVP · what the prototype shows today

A hybrid advisor: tap or type, fixed rules pick, the AI explains
One recommendation, the other two “for later”, fees in dirhams, scenarios that always include a loss
Reserve before recommend · Shariah as a filter · portfolio review by permission, which leads to insights
Review before paying, and a human fallback for questions out of scope

V2 · about six months on, once metrics 01 and 02 have moved

Help with exits and selling. Sell-now versus hold maths before an exit window, and help listing tokens
Retention. Opt-in check-ins, progress against the remembered goal, what to do with rent received
A deeper human fallback. “Talk to a person” hands off to a licensed advisor for large amounts and tax questions
Property consultation. Referral to a human expert for buying a whole property

I prioritised this MoSCoW-style. Everything in the MVP is a Must or a Should. V2 stays a Could until the first two metrics move.

Where I landed

Getting to an answer is easy. Knowing it's the right answer for you is the product

Testing is what separated an experience that looked finished from the moments where it still had to earn trust. If I took this into a real build, I'd take these things with me: rules that choose, a model that explains, a reserve set aside before any recommendation, and a product that is willing to say “not this money, not now.”

Thank you

By Bisma Munawar

Happy to take questions. The appendix is right behind this slide if you want the tables.

Fully wired prototypehttps://investmentadvisordxb.lovable.app/

Design system at Lovablehttps://lovable.dev/preview/SSdyFwH5GWbFI5OsCq1rGsGAj5h0nSpZ

Research corpus · PRD · brand guidelines · synthetic test report, all available on request

Appendix

The reference tables

Everything I kept out of the main story so it could flow. Research sources, the full competitor table, published facts, the hypothesis scorecard, Kano detail, all 22 checks, the 12-area scorecard and the data model.

A1Research sources

Where the customer voice came from, and what I left out

Source	Outcome
Reddit	Strongest source of recent UAE retail-investor customer voice.
Trustpilot	Useful for Stake, SmartCrowd, Sarwa, Wahed and OGold customer experience. Individual reviews aren't representative of whole platforms.
LinkedIn posts/comments	Useful for tokenised-property questions on ownership, liquidity, mechanics. Commenters can't always be verified as target users.
Hacker News	AI-investment trust, explainability and proof concerns. Adjacent, not UAE-specific.
Indie Hackers · Product Hunt	Adjacent evidence on ChatGPT workarounds, financial context, personalisation and confidence in AI investing tools.
Industry forums	UAE investing communities, incl. SimplyFI / Bogleheads-style DIY alternatives.
Quora	Searched; retrieval blocked by robots.txt. No unverifiable quotes included.
X/Twitter · YouTube	Public discussion exists but little reliable target-customer conversation on this three-product problem. Not used as primary voice.
G2 · Capterra	Categories are B2B software for banks/advisers, not UAE retail products. Excluded.

A2Competitive landscape · 1/2

Direct alternatives

Alternative	Type	Does well	Customer complaints / weaknesses	Access	Gap vs. the advisor hypothesis
Stake	Fractional property	Low-entry Dubai property, fractional ownership, AutoInvest	Liquidity, vacant properties / rent interruptions, regret over concentration	AED 500; published fee stack	Helps select property; doesn't solve gold vs tokenised vs fractional
SmartCrowd	Fractional property	Low-ticket Dubai property exposure	Fees, lock-up, exit economics	AED 500; published fees	User still decides whether fractional is right for them at all
OGold	Digital gold	Gold without storage friction	Tiny review sample; support/transparency concerns	Very low entry	Solves access to gold, not whether gold fits
Emirates Islamic	Bank-embedded gold	Gold/silver inside an existing banking app	Little feature-level review evidence	Embedded	Great in-app access benchmark; no cross-asset recommendation layer
Sarwa Invest	Robo adviser	Questions → risk profile → managed portfolio; reviews praise clarity and advisers	Strong counter-evidence to “digital products lack guidance”	USD 500; management fee	Allocation inside its own universe, not this choice
StashAway	Robo adviser	Managed diversified portfolios, multiple risk levels	Less relevant for UAE real-world assets	No minimum; tiered fees	Substitute for users who don't want to pick alternatives
Wahed	Robo · Shariah	Beginner-friendly halal investing	USD-only funding, withdrawal timing; human support valued	USD 500; annual fee	Shows goal/risk-based digital investing is already familiar

A3Competitive landscape · 2/2

Indirect alternatives and workarounds

Alternative	Type	Does well	Weakness	Gap vs. the advisor hypothesis
Wio / integrated investing apps	Indirect	Convenience (“very convenient to use”)	Same user felt FX/commission economics outweighed convenience (perception, not audit)	Proves the value of keeping investing inside the existing journey
ChatGPT / general AI	Indirect	Conversational explanation, apparent personalisation	User must supply accurate context manually; no verified account or product data; can't execute	Embedded verified context + immediate action is the strongest advantage
Google / Reddit / communities	Non-software workaround	Huge volume of real experiences	Conflicting, generic, time-consuming	Doesn't convert the individual's circumstances into one answer
Friends / family	Human workaround	Familiar, trusted; ~half of surveyed UAE investors consult family	Opinion-based, not aligned to exact goals/products	Emotional trust that AI may not replicate. This is why there's a human fallback
Manual comparison in-app	Product workaround	Direct access to facts, pricing, risk	User does the translation into a personal choice	This is the exact decision layer the advisor improves
Cash / do nothing	Incumbent behaviour	Simple, preserves liquidity	May fail longer-term goals	The advisor must first establish the money is suitable to invest

A4Published facts

Competitor facts as published, at time of research

	Stake	SmartCrowd	Sarwa Invest	StashAway	Wahed	ChatGPT
Type	Fractional property	Fractional property	Assisted adviser flow	Assisted adviser flow	Assisted, halal only	General AI
Minimum	AED 500	AED 500	USD 500	None (General Investing)	USD 500	None
Published fees	1.5% entry · 0.5% annual admin · 0.2% then 0.1% KYC/AML · 2.5% exit · 7% on appreciation	1.5% entry · 0.5% annual admin · 2.5% exit	0.85% annual (tiering disputed)	0.2%–0.8% annual	0.99% under USD 250k · 0.49% above	None
Exit route	1-yr lock-in, then two-week windows each May & Nov; 3–5 yr recommended hold	Share Transfer Facility, two-week windows each Mar & Sep, or sale by vote	Withdraw anytime	No lock-ins	No lock-in	n/a
Shariah option	Yes (DFSA Islamic Window + certification)	Yes (Shariyah Review Bureau)	Yes (halal portfolio, same fee)	Yes (Shariah Global Portfolio)	Halal by construction	No
Questions → recommendation	No	No	Yes	Yes	Yes	Conversational only

A5Hypothesis scorecard

Eighteen founding hypotheses against the research

Hypothesis	Verdict
UAE beginner investors struggle to choose	Supported (qual.)
People with a few thousand AED/month experience this	Supported
AED 3,000/month is the correct threshold	Unvalidated
Ages 22–48 are the correct segment	Unvalidated
Beginner / <1 yr is a promising pain segment	Directional
Users manually research online	Supported
Users ask friends/family	Supported
Users use ChatGPT	Supported (qual.)
Users may delay / do nothing	Plausible, weaker

Hypothesis	Verdict
Risk, liquidity, fees, horizon drive decisions	Supported
Exact three-product choice is common	Unvalidated
Users want personally relevant guidance	Supported
AI is an acceptable interface	Directional
Users will follow one AI recommendation	Unvalidated
Explainability matters for trust	Desired attribute
Embedded context beats generic AI	Strong hypothesis
AI recommendation reduces churn	Unvalidated
AI recommendation improves conversion	Unvalidated

A6Limits of the study

What this testing can't tell us

Whether a real Maya would ask about loss, or just carry on without noticing what she doesn't understand
Whether Arjun leaving to research means lost trust, or is just how he decides responsibly
Whether people separate what the company holds from what the advisor can use
What counts as enough evidence for a strict compliance requirement
Whether anyone would act with meaningful money

The limits of the study, stated plainly

Desktop Chrome at mobile width. Not a real phone, not a screen reader, not keyboard only
Returning journeys used the fixed AED 26,000 demo portfolio, not Omar's fictional AED 71,000
Final payment, goal chips, new-holding concentration checks and reduced motion: unverified
Product claims, regulatory status and backend controls were not independently audited
No build identifier recorded. Demo dates not independently checked

The five biggest unknowns are still emotional trust, willingness to act with real money, actual privacy choices, how people interpret risk, and the real decision to continue, compare or leave.

Bridge: "So that's what the testing found. Now here's what I did about it."

A7Kano detail · basics

Basics: expected, earn nothing, and their absence ends the conversation

Benefit	Why it sits here
Explain products in plain language	Literacy varies inside a single thread: one person suggests the S&P 500, the next asks what it is. Every competitor writes beginner copy.
Show how and when money can be withdrawn	Most-raised concern. A fractional platform's own data showed 26% seller liquidity in one exit window. Most people who listed didn't sell.
Show all fees, not the headline	The objection is late discovery, not cost. Competitors publish full tables; match, don't lead.
State the downside honestly	An investor never shown a bad year is the one who sells during one. Also a compliance expectation.
Cite where product facts come from	Nobody praises a citation. Advice without one isn't advice.
Keep the question set short	UAE robo-advisers already onboard briefly. Every extra question loses people.
Invest without leaving the app	Several UAE apps already do it. Handing off elsewhere destroys the value just created.
Shariah compliance status	Absolute for those who need it, weightless otherwise. A filter, asked early. Not a feature.
Accountability · data control	Which entity stands behind the advice, under what permission. Noticed only when broken.

A8Kano detail · performance

Performance dimensions: more is better, with no ceiling

Benefit	Why it sits here
Give one clear recommendation	Strongest signal. AED 320,000 and AED 30,000. Both had money ready, neither could choose.
Explain why the other options lost	A recommendation without comparison reads as a sales pitch. Comparison makes it a decision the user can defend later.
Translate mechanics into personal consequences	“One-year lock-in” vs. “you said you might need this in six months, so this doesn't suit you.” Same fact, different job.
Tokenised vs. fractional in practical terms	Explain what happens when they want their money back, not the legal architecture.
Support follow-up and challenge	“Impossible to test, hesitant to trust.” Depth of conversation separates an advisor from an algorithm.
Net outcome after fees and FX	“What will I receive in hand?” Headline percentages don't answer it.
State what would change the recommendation	Makes the advice feel considered rather than fixed, and gives a reason to return.
Human access at the point of commitment	HSBC 2026: 73% use AI for finance, 12% call it most influential. Reviews of UAE robo-advisers repeatedly praise human contact.
Build confidence · reduce overwhelm	“Helped me make the right decisions for my needs” is the sentence the product is trying to earn.

A9The 22 checks · 1–11

Observed results, 10 to 11 September. A pass applies to visible behaviour only

#	Check	Result	Observation
1	Tap AED 20,000 / 5-yr starter; no echo	Fail	User-message echo and “Got it: AED 20,000, 5 years” both appeared. Reproduced after reset.
2	Type “I have 20k for 5 years”	Pass	Exact acknowledgement and Edit appeared. Home composer initially timed out; full-chat composer worked.
3	Finish typed AED 20,000 journey	Partial	Fractional; 19,801 / 198 / 19,999. AED 1 residual not explained.
4	Put fees on top instead	Unavailable	No such control found. The toggle was removed after it read as an upsell.
5	5-yr fractional scenarios	Pass	2,822 / 4,752 / 7,440 inline. No plus sign; heading doesn't say “net gain”.
6	“make it 30k”	Pass, history issue	Updated and re-ran. Earlier scenario figures also changed above the edit.
7	AED 15,000 / 2 years	Pass	Tokenised; 14,409 invested; fees 288 + 288 + 14. Total within AED 1 tolerance.
8	2-yr tokenised scenarios	Pass	−432 / 994 / 3,134.
9	AED 500 / 4 years	Changed behaviour	Stops first: “You'd need to spend AED 505.” Spend / Change amount. Useful safeguard.
10	AED 2,000 / few months starter	Pass	1,500 accessible; 500 gold. Does not generalise to Daniel's message.
11	“okayy” while quick replies show	Fail	“I didn't catch that” plus starters, rather than activating the first reply.

A10The 22 checks · 12–22

Observed results, continued

#	Check	Result	Observation
12	“what's the weather”	Pass	Expected fallback with visible options.
13	“why real estate”	Pass	Rent and appreciation alongside value loss, vacancy, tied-up money. Two follow-ups offered.
14	Recommendation → review → Not now	Pass	Returns to chat with reassurance and Remind me later.
15	Confirm property investment	Partial / blocked	Verification shown, returned to review. Final pay blocked by approval safeguard; post-pay states unverified.
16	Returning portfolio review / Yes	Pass	Invested 26,000; value 26,480; rent 2,070; next date 15 Oct 2026.
17	Portfolio insights	Pass	Exit window in 35 days; studio with no rent; 23,000 of 26,000 in Business Bay.
18	Should I sell at this window?	Pass	Exit 258; receive ~10,022; overall 722 up. Asks whether money is needed within six months.
19	New investor's concentration notes	Blocked	Requires a new holding after final confirmation. Not verified.
20	Decline permission and ask again	Pass	“No problem. I'll go by what you tell me.” Both times. No second ask.
21	Bottom chart icon	Pass	Opens Portfolio. Empty state includes Ask the advisor.
22	Reduce motion on, reload	Unverified	No supported control in this browser session.

A1112-area scorecard

Twelve areas, in plain language

Area	Assessment	In plain language
Getting started	Needs refinement	The invitation helps; a few interaction details still distract.
Answering questions	Needs refinement	Straightforward answers work better than combined or nuanced messages.
Understanding the recommendation	Major issue	Amount and time are explained, but important personal requirements can be missed.
Telling the products apart	Needs refinement	Fees and timing are clearer than ownership and exit differences.
Trusting the AI	Major issue	Basic concerns and corrections need more reliable responses.
Trusting the specific advice	Major issue	A clear explanation cannot make up for overlooking an essential expense.
Understanding fees	Needs refinement	Initial costs are strong; ongoing costs, rounding and leftover money need context.
Understanding access to money	Major issue	No lock-in and an exit window can sound more certain than the explanation supports.
Portfolio permission	Needs refinement	Declining works, but the boundary before consent is unclear.
Splitting cash and investments	Major issue	The prepared example works; the personalised medical-expense case does not.
Asking follow-up questions	Major issue	Prepared explanations help, but natural questions are not consistently understood.
Overall confidence	Needs refinement	Reaching the answer is easier than knowing whether it is right for you.

A12Could they explain it back?

Understanding and trust by person (simulated judgements)

Person	Understood the recommendation	Understood the alternatives	Main trust change
Maya	Partly	Incomplete; could read “no rent” as “no value”	A helpful start undermined by the unanswered risk question
Arjun	The fee-and-time argument	Still wanted structural evidence	Maths helped; missing evidence kept him cautious
Omar	Why another area was shown	Only partly the other product choices	Seeing his holdings affect the recommendation helped
Leila	The amount-and-time fit	Partly	Respecting refusal helped; pre-consent knowledge confused
Ayesha	Financial fit, not required eligibility	Partly	Asking about compliance without resolving it weakened confidence
Daniel	Saw the recommendation didn't protect his need	Some access differences, not cash certainty	The reserve mistake and failed correction undermined trust

Simulated outcomes: Maya postpones to ask someone she trusts · Arjun leaves to check terms · Omar keeps evaluating outside Business Bay · Leila uses Not now · Ayesha stops and checks elsewhere · Daniel rejects and stops.

A13Data model

The data model. Two decisions in here matter more than the rest

Sessionamount · needed_back_by · invested_before

Consentscope: portfolio_read · granted · revoked_at · session_id

Holdingproduct_type · submarket · amount · income_received · lock_in_ends

Productminimum · fee_components · exit_mechanism · shariah_status · source_url · verified_at

Rules enginedecision + reason codes

Model4 fixed blocks · explanation only

Recommendationrecommended_product · less_suited + reason codes · figures_shown · model_output · user_action (invested | asked | abandoned) · append only

Nothing ever overwrites a `Recommendation` row

That makes it the compliance record and the metrics table at the same time. Which reasoning pattern converts and which one loses people can be answered from one table.

`ProfileFact` appends, it never updates

Goals and income are stored with a timestamp, and every change is a new row. That is what makes “what would change this” real. You can show someone their answer moved and the recommendation moved with it.

Listing.match_attribute holds the single reason a listing came up for this user. It's the only thing the listing card shows.