P / 01
Judgment sits on the RM alone.
No shared layer for reading client readiness. Every interpretation starts from scratch in one person's head.
Knowing the client
is the holy grail of private banking.
Most Relationship Managers never reach it.
Information flows, but meaning gets lost.
Holy Grail puts a layer between the data and the decision.
Scroll
01 · Key Insight · External Research
The industry knows where the bottleneck is.
33%
of an RM's time is core advisory work.
The rest is rebuilding context, pulling data, formatting handoffs.
79%
of HNWIs want their RM to help check decision biases.
65%
of HNWIs say wealth advice feels generic.
02 · The Friction · RM Reality
Information flows, but meaning gets lost.
There's no shared layer for who, when, what, why. The RM rebuilds it alone, in their head.
P / 02
Interpretation leaves no trace.
Decisions get made between Slack, CRM and memory. The reasoning behind them is rarely written, never indexable.
P / 03
Reads drift between RMs.
Two RMs covering one client produce two stories. Cover, handover and audit all suffer the gap.
03 · Where Meaning Leaks
The data is fine. What breaks is the rebuild step.
01 · COLLECT
Inputs arrive
Email, CRM, meeting notes, portfolio events, from disconnected systems.
→
02 · REBUILD
Context stitched by hand
The RM rebuilds the story alone, in their head, in minutes.
→
03 · TRANSLATE
Reasoning becomes advice
The story renders into a next-action call.
The leak is at rebuild: fragmented inputs, no time, no shared layer.
05 · The Gap · Why Between Them
The system sits in the gap between two people.
Ethan has eleven minutes. Dominic asks the same question three ways. Memory drifts. The record thins.
RM · Persona 01
Ethan Tan
Clients want to feel understood, but client context lives across multiple systems, so I keep rebuilding the story manually.
Goals
- Make clients feel understood
- Preserve trust in volatility
- Deliver outcomes with low compliance risk
Pain Points
- Admin overload
- Fragmented systems
- Soft cues not captured
- Tool switching
Behaviours
- Defaults to safe updates under pressure
- Misses weak signals when rushing
Needs
- Single scan view
- Pre-brief before meetings
- Compliance-safe phrasing
11 minaverage pre-meeting decision window
UHNW · Persona 02
Dominic Koh
I want to protect family wealth, but I delay decisions when choices feel complex or downside isn't clear. I expect discreet guidance with a clear next step.
Goals
- Preserve capital + keep liquidity
- Protect family interests
- Decide with confidence, not pressure
Pain Points
- Too many products + complexity
- Generic updates feel low value
- Fear of the wrong move in volatility
Behaviours
- Asks worst-case repeatedly
- Delays when info feels complex
- Moves faster with staged choices
Needs
- Two clear paths + trade-offs
- Downside boundaries first
- What changed recap
×3
asks the same question three different ways
memory drifts
·
consistency shifts
·
record thins
The system holds the structure. The RM holds the judgment.
04 · Research Foundation · Two-Axis Convergence
Holy Grail sits where two research lines cross.
Axis A · Human Judgment
Where do people drift?
- Behavioral economics — heuristics & biases
- Tversky & Kahneman, 1974
- Klein, 1998 — Naturalistic Decision Making
- Kahneman, 2011 — System 1 / System 2
Axis B · Regulatory
How must it be recorded?
- MAS Notice 626 · Banking Act §47
- PDPA §13 / §14
- GDPR Art 22 — automated decisions
- MiFID II Art 25 · EU AI Act Art 86
Convergence
The system assembles, the person judges, the firm records.
06 · Convergence · Interpretive Core Engine (ICE)
How the engine turns raw signals into one score.
Four steps from raw signal to one auditable Composite. Click each step to expand.
Step 01
Raw input ▾
Call transcripts, RM memos, trade timing, portfolio delta, market data.
MARKET
Bloomberg feeds · price delta
LIQUIDITY
Cash flow · exposure events
CREDIT
Rating changes · facility usage
REGULATORY
MAS notices · PDPA flags
BEHAVIORAL
RM memos · meeting tone
Step 02
Tokenization ▾
Five input categories shown as the main evidence gates for portfolio judgment.
For this scenario, five input categories are shown as the main evidence gates for portfolio judgment: Market · Liquidity · Credit · Regulatory · Behavioral. Each category is tokenized by Mistral and mapped to an ICE layer weight.
Step 03
Four layers ▾
Tokens fan out into four ICE layers. The L1 kernel is expressed as R = U + T − C − A, scaled asymmetrically (downside ×2.0 / upside ×1.2).
L1 · 0.40
Self-Motivation
Inferred from observable behavioral tokens through the UTCA kernel. The only layer that reads the client directly.
L2 · 0.33
Implicit Social
Market context: shared anchor across the client cohort. Read directly from external data, not from observation.
L3 · —
Explicit Social
Currently excluded. Canonical input tokens not yet implemented; remains a defined ICE layer.
L4 · 0.27
Environmental
Geopolitical, regulatory, and volatility shocks. Mistral classifies Bloomberg, MAS notices, and news into tokens.
Step 04
Two surfaces ▾
One ICE engine, two exposures: RM Dashboard uses the per-client briefing surface; Compliance Terminal exposes the same reasoning trail for weekly audit.
RM DASHBOARD
One client. One Composite. One trajectory line. One next-action sentence the RM can say out loud.
View surface →
COMPLIANCE TERMINAL
The CO's weekly view. Signals get tokenized, scored on L1/L2/L4, then rolled into one Composite. At this layer the question isn't the client; it's whether the book can carry the risk.
- WHO
- CO · Compliance Officer
- WHEN
- Days · weeks · quarter · post-hoc forensics
- WHAT
- Tokenization · ICE Composite · audit trail · book-wide pattern
- NEVER SHOWS
- Real names · live prices · emotion labels · execution buttons
Sources
07 / A · Axis A · Processing
Compliance Terminal. The CO's weekly book-level audit.
CT cuts raw signals into tokens, measures them across L1, L2, L4, then aggregates to one Composite. The question it asks is not what to do about one person. It asks whether the book can carry this level of risk.
- WHO
- CO · Compliance Officer
- WHEN
- Mon 08:00 SGT · weekly · T-7 / T-30 / T-90 windows
- WHAT
- Tokenization · ICE Composite · audit trail · book-wide pattern · escalation
- NEVER SHOWS
- Real names · live P&L · execution buttons · emotion labels (H1 lock)
CT · Carousel
07 / B · Axis B · Result
RM Dashboard. The 11-minute pre-meeting read.
Eleven minutes before a meeting, the RM lands here. One client. One Composite. One trajectory line. One next-action sentence the RM can say out loud without rephrasing.
- WHO
- RM · Ethan Tan
- WHEN
- 11 minutes before a meeting · hours · days
- WHAT
- One client's Composite · trajectory · portfolio · next action
- NEVER SHOWS
- L3 raw · audit log · emotion labels (Hesitant / Calm)
RM Dashboard · Carousel
08 · Scenario · Russia-Ukraine 2022
One client. Six weeks. The system in motion.
Dominic Koh. SGD 28M book, 22% in European energy. Russia invades on Feb 24. Mark-to-market drops 20% peak-to-trough over six weeks. RM Ethan Tan has eleven minutes pre-meeting to decide.
Client
Dominic Koh
APAC UHNW · 2nd-gen wealth
Exposure
SGD 28M · 22%
STOXX 600 ETF · small RUB position
Window
6 weeks
Feb 18 → Apr 5 · T0 baseline → T5 stabilization
T0 · 2022-02-18
−4.55
in 16 days.
+3.17
in 28 days.
Loss runs ×2.0. Recovery runs ×1.2.
The 72-hour hold turned a −20% sell into a −6.6% rebalance.
The 72-hour hold turned a −20% sell into a −6.6% rebalance.
Sources
Four things the system will never do.
They're locks: written into the files at design time, enforced in code.
2026-04-15 · LOCK 01
Emotion label display banned
Direct display of inferred emotion labels (Hesitant / Calm / Tense) in any client-facing or RM-facing surface is prohibited. Only objective behavioural observations are shown.
Scope · RM Dashboard · all screens
2026-04-15 · LOCK 02
Raw score UI exposure banned
L3 raw scores and intermediate ICE layer values are not surfaced in the RM Dashboard. Only the aggregated Composite is shown, with trajectory direction.
Scope · RM Dashboard · score display
2026-04-15 · LOCK 03
Record channel dual-track banned
AI-generated records and RM-authored records must not be stored in the same field. Every record entry must carry a source tag. No auto-merge without RM review.
Scope · Record system · all channels
2026-04-15 · LOCK 04
Auto-profiling of clients banned
The system may not automatically update a client's risk profile or segment without RM initiation. All profile changes require explicit RM confirmation before write.
Scope · Client profile · all writes