■ ALLY PROACTIVE AI · UX CASE STUDY

DOUBLE DIAMOND · FINTECH · M&A
SELF-DIRECTED INDUSTRY CONCEPT
2026 · NEEL BHESANIYA

Designing trust into AI-driven deal intelligence.

How an investment banker comes to believe in, verify, and act on an AI-surfaced acquisition signal — in the 30 seconds between her coffee and her 8am pitch call.

Scenario

Deal Opportunity Detection

Surfaces

4 screens · dashboard → pipeline

Method

Double Diamond — 4 phases

Built in

Figma · interactive prototype

The Brief

02 / 17

The technology works. The open question is the experience.

Ally Proactive AI is a concept for surfacing deal opportunities, comparable transactions, and market signals to investment bankers. The question I set myself: design an end-to-end experience where bankers can trust, interpret, and act on AI-generated insights. I scoped four scenarios this kind of product has to handle — and chose the hardest on purpose.

01Deal Opportunity Detection"Company X may be seeking acquisition." The most generative scenario for trust + explainability.My pick

02Comparable Analysis EnhancementThree overlooked comparable deals.

03Market Risk AlertSector volatility impacts an active deal.

04Client Strategy InsightBuyback activity signals M&A liquidity.

Why this one: deal detection forces the design to answer the hardest question first — how does a senior professional come to believe a machine?

The Real Problem

03 / 17

Three hard problems hiding inside a simple-sounding brief.

I reframed one vague goal — "make AI trustworthy" — into three concrete design problems.

The trust problem

Bankers are paid to verify everything. An AI that says "trust me" loses instantly. Output has to be sourced, weighted, and easy to push back on.

The attention problem

A banker reads an insight headline in ~8 seconds. Thesis, confidence, urgency, and source-density must be readable at a glance — detail one click deep, not zero.

The action problem

Reading is expensive. Acting should be cheap. One click to pipeline, share, or request analysis — with provenance following automatically.

Decide what to do with an AI-surfaced signal in less time than it takes to read a Bloomberg headline.The objective, in one line

Process

04 / 17

A double-diamond approach — diverge, then converge, twice.

Discover broadly, define narrowly. Develop broadly again, deliver narrowly. The framework forces honest choices about what to leave out.

01 · Discover

Understand the space

User research — analysts, VPs, MDs
Competitor analysis — 6 tools
AI explainability patterns

02 · Define

Frame the right problem

HMW statement
3 design principles
Journey + IA scope

03 · Develop

Explore solutions

Information architecture
Confidence + trust UIs
Wireframes & flow tests

04 · Deliver

Ship one direction

4 connected screens
Custom tokens + components
Edge cases & transparency

01 · Discover · Findings

05 / 17

What bankers actually struggle with — five recurring frustrations.

In a real engagement: 8–10 interviews across analyst, VP, MD, and quant. For this concept, on a tight timebox: archetype-based reasoning from published banker interviews — each mapped to the design moment that solves it.

Trust

"Why this, why now?"

Alerts arrive with no reasoning chain. Analysts re-do the research to confirm — defeating the purpose.

→ weighted signals + confidence

Volume

Alert fatigue

Tools fire dozens of low-relevance alerts daily. Analysts learn to mute — then miss the important one.

→ ranked feed, low-conf demoted

Context

Isolated insights

An insight that doesn't connect to a mandate or pipeline feels orphaned and hard to action.

→ dashboard rails + pipeline link

Friction

Citation-hygiene tax

Sharing with an MD means manually copying sources and rebuilding the reasoning by hand.

→ provenance follows the action

Feedback

No way to push back

When the AI is wrong, there's nowhere to say why — so the model never learns from its best critics.

→ "Not Relevant" feedback control

If I can't see why the AI thinks this — and whether it's been right before — I'm not putting my name on it.Composite — M&A analyst, Industrials coverage

01 · Discover · Competitive Gap

06 / 17

Six tools bankers already use — and the gap none of them fill.

Tool	Primary job	Explainability	Precedent / patterns	Feedback loop
Bloomberg Terminal	Real-time data + news	Low — alert only	None	None
AlphaSense	Doc search + summaries	Med — paragraph	None	Thumb up/down
Tegus	Expert call transcripts	Med — speaker	None	Thumb up/down
FactSet	Data, modeling, screening	Low — rule logic hidden	Manual comps	None
Affinity (CRM)	Relationship intelligence	Med — activity log	None	None
Ally Proactive AI — target	Deal-opportunity detection	High — confidence + pattern accuracy	Comparable historical patterns	Persistent, structured

The pattern: existing tools surface either data (Bloomberg, FactSet) or documents (AlphaSense, Tegus). None surface theses — synthesized hypotheses about what's likely to happen — backed by the precedent transactions that make them believable. That's the opening Ally Proactive AI can claim.

02 · Define

07 / 17

Four insights became three rules I held every decision to.

Confidence is a vector, not a number.

Users want to know how sure, compared to when, and whether it's rising or falling — direction is its own signal.

Counter-evidence is a feature, not a flaw.

Showing reasons the thesis might be wrong, at the same weight as confidence, is the most credibility-building move.

Precedent makes a thesis believable.

Bankers reason by comparison. "This pattern matched cases that led to acquisition" earns trust that a bare percent can't.

Provenance follows the action.

When an insight is added to the pipeline or shared, its origin tag travels with it — context never gets lost in handoff.

PRINCIPLE 01

The AI is a colleague, not an oracle.

Treat output like a junior analyst's memo — confident but defensible. Every claim weighted and easy to disagree with.

TEST · Can the user disagree without leaving the screen?

PRINCIPLE 02

Earn the click in 8 seconds.

Surface thesis, confidence, and freshness on the card. Full explainability lives one click deep — not zero, not two.

TEST · Can a banker triage this in under 10 seconds?

PRINCIPLE 03

Make the decision the easy part.

One click to pipeline, share, or queue analysis — each carrying its provenance automatically.

TEST · Does the origin follow the action untouched?

02 · Define · Journey

08 / 17

From morning coffee to decided.

One analyst, one morning, across four connected surfaces. A fast path for high-trust insights, a deeper path for ones worth verifying — both end in a captured decision.

07:42 · Dashboard

Scans the day

Ranked feed + signal breakdown + market pulse while coffee brews.

"What's worth my attention today?"

07:43 · Feed

Spots Mosaic

Top of feed. 87%, confidence rising. Reads the thesis line.

"Plausible. Is the AI making this up?"

07:44 · Insight

Opens the panel

4 trust metrics + comparable historical patterns + counter-evidence.

"OK — has this pattern been right before?"

Fast path

Add to Pipeline

Lands in IDENTIFIED, tagged "AI Insight." Provenance attached.

Trust path

Share with MD

Note + supporting signals. MD sees the same evidence.

Three key UX moments: triage (thesis + confidence + freshness on the card) → conviction (counter-evidence + precedent at the same scale as confidence) → action (one click, provenance follows).

03 · Develop · The Core Trade-off

09 / 17

How do you show "the AI is 87% sure"? — four directions I tested.

The hardest single-element decision in the design. A flat percent felt like a marketing claim; the goal was a visual that reads as honest.

Option A

Just the percent

87%

Rejected

Opaque. 87% of what? Reads like a marketing claim — no way to tell if it's high.

Option B

Tier-only badge

HIGH

Rejected

Hides the precision bankers want. Equates 75% and 95% — a 20-point gap that matters to a deal.

Option C

Circular dial

87%

Rejected

Marketing-y, hard to scan at row level, eats vertical space. Wrong for dense feeds.

Option D · Shipped

Tier + percent + trajectory

87%↗ Confidence Rising

Shipped

Glanceable as a tier ("High Confidence · 87%"), precise as a number, and paired with direction. Reads honestly.

The unlock: confidence isn't a single value to display — it's a direction over time. Every insight in the shipped design carries its tier, its percent, AND a "Confidence Rising / Stable" trajectory. That pairing is what makes the number feel earned, not asserted.

04 · Deliver · Screen 01 · Dashboard

10 / 17

The home screen treats AI as core workflow, not a notification.

Outcome metrics, up top

KPI tiles lead with what the tool is for — pipeline value, and "Avg. Analyst Hours Saved / Week."

A ranked feed, not a chronological one

"Ranked by your coverage, confidence, and time-sensitivity." Each card answers what / why-now / how-sure / how-fresh.

A transparency rail

"AI Signal Breakdown" shows the funnel (3 high / 8 medium / 36 low) — the AI shows its volume, not just its winners.

Grounded in the analyst's market

Market Pulse + Calendar localize to UK coverage — FTSE 100, BoE decisions, Companies House filings.

04 · Deliver · Screen 02 · Insights Feed

11 / 17

The full feed is built for 8-second triage.

The thesis is the headline

Not "AI detected unusual activity" — the actual hypothesis in plain English: "May be seeking acquisition."

Confidence tier + percent + trajectory

"High Confidence · 87%" with "Confidence Rising." A "Trend Watch / Low Confidence · 60%" item sits lower — surfaced, not suppressed.

Filter chips for power users

All / Deal Opportunities / Risk Alerts / Client Signals — refine the feed, don't just browse it.

Evidence in one line

Each card shows signal count, source count, and confidence direction — enough to triage without opening.

04 · Deliver · Screen 03 · The Hero

12 / 17

The insight panel earns conviction with precedent, not just a percent.

Four trust metrics, above the fold

Overall Confidence, Pattern Accuracy (73%), Time Sensitivity, and Counter Evidence — given equal visual weight.

Comparable Historical Patterns

"AI matched cases with similar signal clusters that led to acquisition within 18 months." Precedent is how bankers actually verify a thesis.

A two-tier action ladder

Lightweight: Save to Watchlist · Notify · Not Relevant. Committing: Share with MD · Request Deeper Analysis · Add to Pipeline.

Disagreement is one click

"Not Relevant" is a first-class control — the feedback channel sits beside the output, not buried in a menu.

Counter-evidence at the same scale as confidence is the single most credibility-building choice in the design.

04 · Deliver · Screen 04 · Pipeline

13 / 17

Where the decision lands — and the origin travels with it.

A real system of record

"Add to Pipeline" isn't a dead end — the deal appears as a card moving through Identified → Qualifying → Pitching → Mandate.

Provenance follows the deal

Cards keep their origin tag — "AI Insight," "New AI Signal" — so months later, anyone can see the AI sourced it.

Risk surfaced on the card

"Risk: Sector Volatility," "Risk: IX Spike," "Liquidity Signal" — the same trust signals ride into the deal view.

This is how trust scales

Beyond one analyst — into MD review, into pipeline records, into firm memory. Context never has to be rebuilt.

05 · Handoff · The Redline

14 / 17

Handing off so engineering never has to guess.

I redline the signature component — the insight card — down to the token. Every measurement references a variable, not a magic number, and the full anatomy is numbered so a developer can build it without pinging me ten times.

1AI InsightHigh Confidence · 87%⚐ Top of FeedSurfaced 2h ago

2Mosaic Industries MOSI · Specialty Chemicals

3May be seeking acquisition

4Debt-to-equity dropped 38% over 6 mo. CEO met with 3 boutique IB advisors. Dividend paused Q2. Cash reserves +22% YoY.

5◳ 4 Signals▤ 11 Sources↗ Confidence RisingOpen Insight →

Component anatomy → token

1ConfidenceChip · variants ai · high · medium · low · color color/confidence-* · gap space-2

2Company title · type/display-serif-28 · color/ink

3Thesis · type/body-sb-15 · 1 line, truncate with ellipsis

4Evidence · type/body-13 · color/ink-mute · clamp 2 lines

5Meta row · type/mono-11 · gap space-4 · CTA color/green

card/pad space-5 · 20 card/radius radius-lg · 8 card/border color/line card/bg surface row/gap space-2.5 · 10

05 · Handoff · States, Contract & Process

15 / 17

The two things engineers can't build without — states and a data contract.

Every variant & state — specified, not assumed

ConfidenceHigh (green) · Medium (amber) · Low → demoted to “Trend Watch” Trajectory↗ Rising · → Stable · ↘ Falling CardDefault · Hover (border-strong + shadow) · Focus (2px green ring) LoadingSkeleton rows — never a blank panel Empty“No insights in your coverage yet” Error“Couldn’t load — retry” + cached timestamp A11yCard is one focusable element · tier announced as text, not color alone

The data contract — what the ML lead returns per insight

Insight { thesis: string // ≤120 chars, plain English confidence: 0–100 // tier derived from value trajectory: rising | stable | falling patternAccuracy: { rate, n } // backtest + sample size timeSensitivity: high | med | low signals: Signal[] // weighted sources: Source[] // count → card counterEvidence: int // equal weight comparables: Deal[] // precedent matches }

Annotate in Figma

Tokens + states baked into the file — dev mode reads them directly.

Spec sheet

This doc — anatomy, props, states, and the data contract in one place.

Walkthrough

30-min handoff call with front-end and the ML lead together.

Open thread

A Slack channel stays live through the sprint for edge-case calls.

Why it matters

A spec'd component ships without ten "what's the spacing here?" pings — and because the data contract is agreed up front, the front-end and the model get built in parallel instead of blocking on each other. On an AI product, that contract is the design–engineering relationship.

04 · Deliver · Edge Cases

16 / 17

Where the design has to keep its promises.

Six failure modes I designed for — several already visible in the shipped screens.

Low confidence <60%

Demoted to a "Trend Watch" tag and pushed down the feed — surfaced for completeness, never top-of-feed. Prevents alert fatigue.

Confidence stalls

Trajectory switches from "Rising" to "Stable" — the absence of momentum is itself information the analyst can act on.

User disagrees

"Not Relevant" with optional reason. Suppresses similar insights and feeds the model so this analyst's coverage sharpens.

Active-deal risk

Pipeline cards carry "Risk: Sector Volatility / IX Spike" tags so a known opportunity downgrades visibly, in context.

Thin precedent

When comparable patterns are weak, the panel leads with counter-evidence and softens the recommended action.

Client-data isolation

Insights scoped to coverage groups; no client data crosses the model boundary — a hard constraint stated in the brief.

Reflection & Outcomes

17 / 17

What I'd take into the next AI product I design.

Trust is a chain, not a feature.

It's won by preserving the reasoning chain — confidence, precedent, counter-evidence, provenance — through every screen. Lose one link and it collapses.

Show uncertainty louder than confidence.

Counter-evidence at equal weight was the single most credibility-building decision — counter-intuitive, but right for high stakes.

Bankers reason by precedent.

"This pattern matched deals that led to acquisition" did more for trust than any confidence number alone.

Provenance is the system's job.

The origin tag follows the deal into the pipeline automatically — that's how trust scales past one analyst.

If this shipped — the metrics I'd hold it to

38%

Insight → action rate (vs ~12% baseline)

2.4min

Insight → captured decision (−86%)

11.4hr

Analyst hours saved / wk (surfaced in-product)

<8%

Reported false-positives, into retraining

— Illustrative targets, co-defined with PM in a real engagement —

Three open questions I'd love to discuss

01 Is the right unit of feedback an insight, a signal, or a specific comparable?

02 How do we attribute a closed deal — credit to AI, analyst, or MD?

03 Should the confidence threshold be configurable per coverage area, or globally tuned?