Estrella Zijlstra
← Work / Case study 02 · Candescent / AI Product Strategy · Fintech

Intelligent banking at platform scale.

Personalization strategy for a platform serving 1,300 financial institutions and 30 million end users, and the system that made it buildable rather than aspirational.

Huge Inc. · 2025–2026
AI-generated summary

AI product strategy for personalizing digital banking across 1,300 financial institutions and 30 million users. The Modes framework resolves user signals into adaptive product responses, eight financial modes prioritized across 88+ features and five surfaces, turning personalization from a feature wishlist into a board-level strategic system on the road to a Q4 2026 V1 launch.

Client

Candescent, the largest non-core digital banking provider in the US. B2B2C, white-label.

Engagement

Discovery & Vision at Huge Inc. Fourteen weeks, extended as the scope grew.

My role

Product strategy lead for the strategy layer: user strategy, personalization, prioritization, measurement, within a co-delivered experience vision.

Scale

1,300 financial institutions. 30 million end users. Five product surfaces.

01
The problem

A platform grown through acquisition into a disconnected ecosystem, failing its FI clients and, through them, millions of consumers. One-size-fits-all in a market where the leaders had gone adaptive.

FI clients described feeling at the mercy of the vendor, with 6–12 month implementation timelines. Internally, three separate mobile teams shipped against three separate roadmaps. But underneath the surface symptoms sat a harder structural problem: every product decision, from what a dashboard shows to how onboarding flows, was made without live intelligence about what the person on the other side of the screen actually needed in that moment.

And the standard fix, a handful of personas, could not work here. Candescent doesn't serve one bank's customers; it serves 1,300 institutions whose users range from a first-time checking customer at a rural credit union to the CFO of a growing SMB. A single "right" experience does not exist.

The platform knew an enormous amount about its users. None of that knowledge reached the experience. It needed a system for producing a million right experiences.

The strategic question

How do you build user intelligence into the product itself, so that what each person sees is composed from what they need right now, and it works for every institution without 1,300 bespoke builds?

02
What I built

Research that earned the right to a point of view

A multi-stream discovery: a survey of banking professionals across C-level, VP, and IT roles; in-depth interviews with consumers, SMBs, and FI stakeholders; a competitive audit spanning 20+ players from Chase to Mercury; and an AI-assisted synthesis of 84,000+ consumer conversations. I owned the desk research and led the synthesis into the Discovery Report that became the reference document for everything downstream. The operating principle: learn before launching. Every hypothesis was stress-tested while decisions could still change.

Segmentation built for behavior, not demographics

Deep segmentation research surfaced a load-bearing bifurcation: Pragmatists who want simplicity and reliability versus Power Users who want control and data. It grew into a full B2B2C segmentation covering consumers, businesses, bankers, and administrators. Each segment carries a distinct product opportunity. This became the platform's user strategy.

Modes × Segments
Same modes. Different lives.
Consumer
SMB
Pragmatist
40–50%
Power User
25–35%
Wealth Builder
15–20%
Credit Builder
10–15%
Solo Operator
35–40%
Growing SMB
40–45%
Maintain & Protect
Optimize & Analyze
Build & Grow
Learn & Plan
Control & Delegate
Recover & Stabilize
Transact & Move
Secure & Verify
Primary
Secondary
Occasional
Rare
Control & Delegate separates the Solo Operator from the Growing SMB; recovery risk is shared across both. No segment needs everything, and no mode belongs to one segment.

The Modes framework, the central invention

Instead of asking "who is this user," the system asks "what is this user trying to do right now." Eight adaptive modes were designed, shared across consumers and SMBs and weighted differently by segment, each named as an action pair and derived from a jobs-to-be-done inventory. Each mode specifies the jobs it serves, the signal that detects it, and the product response: hero content, tone, and what to de-emphasize.

01
Maintain & Protect
02
Optimize & Analyze
03
Build & Grow
04
Learn & Plan
05
Control & Delegate
06
Recover & Stabilize
07
Transact & Move
08
Secure & Verify

Signal → Mode → Response

A framework is only as good as its operability. Every mode is paired with an explicit detection chain: four signal categories, weighted by context, that trigger the mode and drive the response. This turned personalization from an aspiration into an engineering specification.

01 · Signals
Behavioral
Navigation patterns, feature usage, dwell
Transactional
Money movement, balances, account activity
Configuration
Settings, alerts, delegated permissions
Conversational
Support inquiries, search, assistant intents
02 · Mode detection
Signals are weighted by context and scored against the mode library.
Active mode
Optimize & Analyze
One of 8 adaptive financial modes
Each mode defines the jobs it serves, its detection threshold, and its response.
03 · Response
Hero content
What leads the page for this job, right now
Tone
How the product speaks: reassure, guide, or empower
De-emphasis
What steps aside so the current job stays in focus
The question shifts from "who is the user" to "what is the user trying to do right now." Detection is an engineering specification, not an aspiration.

Page composition by mode

To make the system buildable at platform scale, modes were mapped down to the component level. Every module on every page is typed as Fixed, Flex, Adaptive, or Optional, so the same design system serves every institution while each user's page composes itself around their current mode.

Page Composition by Mode
One design system. A million right experiences.
Every module is typed
Adaptive
AI-driven; composed per mode at runtime
Fixed
Core structure; identical for everyone
Flex
FI can reorder within guardrails
Optional
On/off toggle per institution
The same skeleton serves 1,300 institutions. The mode decides what each person sees, in what order, and with what emphasis.
Mode: Maintain & Protect
Header & Navigation
Account Summary
Adaptive hero
Safe-to-Spend Widget
Daily allowance after upcoming bills
Bills & Upcoming (next 7 days)
Adaptive · P2
Recent Activity
Flex · P3
Spending Snapshot · Optional
Bottom Nav: Home · Pay · Insights · More
Mode: Recover & Stabilize
Header & Navigation (simplified)
Account Summary
Adaptive hero
Priority Alert
Immediate actions, available funds
Breathing Room Meter
Adaptive · P3
Recent Activity (minimal)
Flex
Upsells & Investment Teaser · Hidden
Bottom Nav: Home · Pay · Help
Same components, same institution, two different people: one page composes for confidence, the other for crisis. De-emphasis is a design decision, not an absence.

The persona simulator

To pressure-test the logic before anything was built, I created a Persona → Mode Calculator: a working tool that scores modes from persona inputs and shows how any given user profile would experience the composed page. It let the team and the client interrogate the system with evidence, not intuition, while decisions could still change.

AI as an intelligence layer, not a feature

Co-created with our AI strategist: intelligence as architecture woven through the experience: personalized insights, tips, and recommendations fed by the same signal chain, governed by confidence thresholds and reasoning transparency, invisible by default. The research showed a real tension: an executive AI mandate on one side, users who hadn't asked for AI on the other. The Modes system resolved it. AI became the engine serving the user's current job, not a chatbot bolted onto a legacy app.

A sequenced path, not a poster

I built the prioritization methodology and the wave plan, from V1 Functional and V1 Sticky & Magnetic through V2 and Vision, with a documented rationale for every deferral, plus the measurement framework to track the transformation against outcomes. The roadmap treats personalization as runtime infrastructure: the intelligence ships with the product, so the experience keeps adapting after launch.

Path to Intelligent Banking
Each phase unlocks the next. The intelligence ships with the product.
Foundation
MLP
Post-MLP
Vision
01
Designed Intelligence
Smart by structure, not by data. Mode-aware layouts, unified IA.
02
Calculated Intelligence
Insights from data the platform already holds. Safe-to-spend, forecasts.
03
Connected Intelligence
The full picture. Aggregation and behavioral signal collection begin.
04
Predictive Intelligence
Automatic mode detection. The system anticipates, guides, and keeps adapting.
You can't have prediction without behavioral signals, signals without calculation, or calculation without structure. Sequencing is the strategy: the sooner signals are collected, the sooner the models learn, and the experience keeps adapting after launch.
03
The judgment calls

Two decisions defined the engagement more than any deliverable.

1

Taking user skepticism about AI seriously instead of designing around it

FI clients were excited about AI; end users were wary. We designed for the skeptics: intelligence woven through the experience rather than a feature demanding attention. Then we put the concepts back in front of those same users. The skeptics became enthusiasts once recommendations felt smart and never intrusive. That's what learning before launching buys: an AI story that was bold, evidence-backed, and already validated with the hardest audience.

2

Refusing to let personalization stay conceptual

Most personalization strategies die as principles on a slide. Forcing every mode down to a detection signal, a page composition, and a computable score is what made this one buildable.

04
Outcome

The vision was adopted as one of Candescent's top three board-level strategic priorities. Intelligent Banking has since been publicly announced, with V1 set to launch in Q4 2026.

Top 3
Board-level strategic priority
8
Adaptive financial modes
88+
Features & components across five surfaces
Q4 '26
V1 application launch

"The articulated strategy and early design thinking is really on point."

Chief Design Officer, Candescent

The work opened the door to follow-on engagements, where I helped define the MLP scope for the new app.

05
Why it matters

The hardest problems in financial services share a shape: the organization knows more about its customers than almost any industry, and almost none of that intelligence reaches the moment of decision. This engagement closed that gap twice. Before launch, by learning first. And after launch, by building the intelligence into the runtime itself, so the experience keeps adapting for every user, at every institution, at once.

You might also like Product Operations in the age of AI
Point of view · 7 min →
Index ← All work
Next case study SharkNinja
Product Lifecycle · Commerce →