02 — Financial Services
Origination, servicing, risk and treasury modelled as one value engine — with agents that act inside model risk management, not around it.
North Star
Risk-adjusted return per client relationship
The value engine, end to end
The same six moves we run everywhere, expressed in the physics of this industry.
Where is margin lost between pricing and servicing?
Value pool map by product
What drives conversion, loss and cost-to-serve?
Causal + credit driver model
What happens to P&L if we shift policy?
Portfolio and journey twin
Who executes decisions at volume?
Governed decision agents
Can the regulator replay it?
SR 11-7 / DORA control file
Does the lift survive the cycle?
Outcome contract + backtest
Where the value is
Before a twin is built, we agree where the money sits and what moves it.
Value pool
Origination
Driver
Decision speed, pull-through, pricing precision
Measured as
Approved value per applicant
Value pool
Credit risk
Driver
Early-warning accuracy, collections timing
Measured as
Basis points of loss avoided
Value pool
Cost-to-serve
Driver
Contact deflection, straight-through processing
Measured as
Cost per serviced account
Value pool
Financial crime
Driver
False-positive rate, investigator throughput
Measured as
Alerts cleared per FTE
Use cases
Each one is a number somebody already owns, a twin that models it, and an agent that moves it inside a defined authority. Filter by function.
17 use cases
Assembles the capital, liquidity and stress-test pack from source ledgers, explains every movement, and flags where a ratio drifts toward its limit before the submission window closes.
Turns the quarter's performance data into a defensible written narrative: what moved, why it moved, and what the next three levers are — each sentence traceable to a figure.
Reads identity documents and proof-of-address scans, checks them against the register, and flags expiry, mismatch or tampering — so onboarding stalls only where a human judgement is genuinely required.
Pre-investigates transaction-monitoring alerts, resolves entities across systems, and drafts the suspicious-activity narrative with evidence attached. Humans decide; agents assemble.
Builds the underwriting file from statements, filings and bureau data, tests affordability under stress, and drafts a reason-coded recommendation an underwriter can accept, edit or reject.
Extends credit access to applicants without a conventional history using cash-flow and behavioural signals, with fairness testing built into the model card rather than bolted on afterwards.
Prices sub-minute credit decisions on thin signal, with a loss curve that is monitored per cohort and a kill-switch when the cohort drifts.
Predicts who repays early, who rolls, and who is about to miss — so treasury forecasts and collections treatments are set on behaviour rather than on averages.
Chooses the cheapest treatment that cures the account — message, plan, hardship route or escalation — and learns from what actually cured, not from what was policy.
Forecasts cash demand per location and per day, then sets replenishment so idle cash and emergency runs both fall.
Moves audit from quarterly sampling to full-population testing, with exceptions surfaced the week they occur and evidence packaged for the auditor.
Scores quote competitiveness in real time against observed trades and predicts where the price will be missed, so traders see the error before the market does.
Defines the objective, imposes the constraints, then simulates the strategy across regimes — including the regime nobody wants to model.
Screens a security universe for factor exposures and candidate signals, with decay and crowding tested before anything reaches a portfolio.
Ranks which companies are likely to transact in a given window from filings, ownership shifts and news, giving coverage teams a shortlist rather than a universe.
Extracts the claims, the guidance and the hedge from calls, filings and broadcast media, and cites the timestamp behind each one.
Puts the right product in front of the right customer at the moment of intent, and suppresses the offers that would have been declined anyway.
Outcome twins
Each twin owns a set of levers. Each feature moves one lever. Each agent executes inside a defined authority.
Simulate policy, pricing and channel mix against booked, risk-adjusted volume.
Levers
Features
Multi-agent layer
Underwriting Assist Agent
Assembles the credit file and drafts the rationale.
Pricing Agent
Proposes price within board-approved bands.
See deterioration early and act at the cheapest point in the curve.
Levers
Features
Multi-agent layer
Early Warning Agent
Flags and routes deteriorating exposures daily.
Cut false positives without moving detection risk.
Levers
Features
Multi-agent layer
Alert Triage Agent
Pre-investigates and packages alerts for review.
Reduce cost-to-serve while lifting resolution quality.
Levers
Features
Multi-agent layer
Servicing Agent
Resolves standard cases end to end with audit trail.
Unit economics
Set your volume, switch features on or off, and move each impact to your own evidence. Everything is expressed per account.
Accounts serviced per year
450,000
Contribution per account
$190
Per-account contribution; substitute per-loan or per-trade economics as needed.
Annual program investment
$3.5m
Origination Twin
Volume — Booked volume
3.0% — Finds approvable applicants inside current risk appetite.
$2.6m
annual value
Servicing Twin
Unit cost — Cost per account
6.0% — Removes manual handling from standard journeys.
$11.6m
annual value
Financial Crime Twin
Unit cost — Investigation cost
4.0% — False positives cleared before a human opens them.
$7.7m
annual value
Risk Twin
Leakage — Credit loss leakage
20.0% — Acts earlier where cure is cheapest.
$5.0m
annual value
Origination Twin
Unit revenue — Revenue per account
2.0% — Price precision within approved bands.
$5.6m
annual value
Value by feature
Annual value contributed by each active feature at current settings.
Value and ROI vs volume
How engineered value and ROI move as accounts serviced per year change. The marker is your current setting.
Cumulative value vs investment
Where the curve crosses the investment line is payback.
Value mix by lever
Which lever the engineered value is actually coming from.
Baseline contribution
$85.5m
Engineered annual value
$32.5m
ROI on program
829%
Payback
1.3 months
Directional model. In an engagement every number here is replaced by your measured baseline, signed off by finance, and written into the outcome contract before any agent goes live.
Scale
Champion/challenger against the incumbent policy.
Same twin, new exposures; model risk artefacts reused.
Shared runtime, jurisdictional guardrails, single risk register.
Integrations
Vendor-agnostic by design. Connectors are added per engagement — no platform lock-in.
Core & origination
Risk & crime
Data & cloud
CRM & servicing
Regulations and standards
Compliance is a design input, not a review gate. Every agent action is logged, attributable and reversible.
EU AI Act
Creditworthiness scoring is high-risk: oversight, logging, accuracy evidence.
DORA
ICT resilience, third-party register, incident reporting for AI services.
SR 11-7 / SS1/23
Model risk management: development, validation, monitoring.
GDPR Art. 22
Automated decision rights, explanation and human review.
AML / KYC
FATF-aligned detection with defensible thresholds.
PCI DSS
Cardholder data isolation across agent tooling.
Standards supported
Next industry
Manufacturing & Industrials
