Full Banking Agent — Skill Library Build and Capstone
In Lessons 1 through 14, you progressed from understanding the three regulatory pillars to calculating ECL, capital ratios, and liquidity metrics, then connected the pillars through cross-pillar integration and validated data integrity through reconciliation. This capstone lesson brings everything together. You will build four banking skills from scratch, set up six scheduled operational tasks, execute a comprehensive cross-pillar scenario, and produce the deliverable that every banking risk function ultimately serves: the Board Risk Report.
This is a 90-minute lesson structured in three parts: skill library build (25 minutes), capstone scenario execution (50 minutes), and chapter summary (15 minutes).
Part 1: Skill Library Build — Exercise 11
Duration: 25 minutes
Building Skills Using Method A (Expert Interview)
Method A structures a skill by interviewing a domain expert — in this case, you are both the expert (having completed 14 lessons of banking regulation) and the builder. For each skill, you define the persona, the decision logic, and the validation criteria.
Build these four skills:
Skill 1: ifrs9-ecl
# SKILL.md structure
name: ifrs9-ecl
description: "Calculate Expected Credit Loss under IFRS 9. Handles
Stage 1 (12-month ECL), Stage 2 (lifetime ECL), and Stage 3
(lifetime ECL, credit-impaired). Supports individual facility
and portfolio-level calculation with macroeconomic scenario
weighting."
persona: "IFRS 9 credit risk analyst with expertise in ECL
modelling across retail and corporate portfolios"
key_decisions:
- "Stage classification drives measurement basis (12-month vs lifetime)"
- "PD source: point-in-time for ECL, through-the-cycle for regulatory EL"
- "LGD: match to collateral type and apply downturn adjustment under stress"
- "Scenario weighting: probability-weighted ECL is NOT the base case ECL"
validation:
- "ECL = PD x LGD x EAD (verify formula applied correctly)"
- "Stage 2 and 3 use lifetime PD (not 12-month)"
- "Probability-weighted ECL >= base case ECL (non-linearity check)"
Skill 2: ifrs9-staging
Define the skill for IFRS 9 stage classification. It should:
- Accept facility-level data (DPD, rating changes, qualitative indicators)
- Apply SICR triggers (quantitative: 30 DPD, 2-notch downgrade; qualitative: job loss, collateral decline)
- Return Stage 1, 2, or 3 with reasoning
- Flag edge cases where professional judgment is required
Skill 3: basel-rwa-credit
Define the skill for credit risk RWA calculation. It should:
- Support both Standardised Approach and IRB
- Apply correct risk weights by asset class
- Handle off-balance-sheet items with CCFs
- Apply the Basel IV output floor (72.5% of SA RWA)
Skill 4: aml-typologies
Define the skill for AML typology identification. It should:
- Recognise common typologies (structuring, layering, trade-based ML, PEP abuse)
- Accept transaction patterns as input
- Return identified typologies with confidence level and red flags
- Flag the agent boundary: typology identification is automated; SAR filing decision is human
Setting Up 6 Scheduled Tasks
After building the skills, configure these operational tasks:
| Task | Frequency | Skills Used | Output |
|---|---|---|---|
| 1. Staging monitor | Daily | ifrs9-staging | Flag any facilities where SICR indicators changed since last run |
| 2. ECL calculation | Quarterly (or on-demand) | ifrs9-ecl | Full portfolio ECL with scenario weighting |
| 3. Capital ratio | Daily | basel-capital, basel-rwa | CET1, Tier 1, Total Capital ratios |
| 4. LCR calculation | Daily | liquidity-lcr | LCR with HQLA breakdown |
| 5. AML alert prioritisation | Daily | aml-typologies | Prioritised alert queue with typology tags |
| 6. Sanctions batch screen | Daily | aml-cdd-edd | Screen all new payments against UK/EU/OFAC lists |
Validation: 11 Cross-Domain Queries
Test the skill library with these queries. Each should produce a correct, complete response:
- "Calculate ECL for a GBP 500M retail mortgage portfolio, Stage 1, PD 1.2%, LGD 25%"
- "This borrower has gone from A rating to BB with 45 DPD. What stage?"
- "Calculate RWA for GBP 200M in corporate bonds rated BBB under SA"
- "What is the CET1 ratio if CET1 is GBP 400M and RWA is GBP 3.2B?"
- "Classify HQLA: GBP 500M in government bonds, GBP 100M in covered bonds"
- "A customer makes 6 cash deposits of GBP 9,800 in one week. What typology?"
- "Screen: payment from Ali Mohammed (UAE) to Steel Corp (Dubai), USD 150K"
- "Trace cascade: GBP 30M loan discovered as fraudulent. Stage migration? ECL impact? CET1 impact?"
- "Reconcile: ECL model says GBP 125M, GL says GBP 123M. What is the likely cause?"
- "Design an adverse macro scenario for a UK bank: GDP -3%, unemployment 8%"
- "Produce a monthly regulatory cycle workflow: which tasks run on which days?"
Part 2: Capstone Scenario — Exercise 10: Board Risk Report
Duration: 50 minutes
Bank Profile
| Metric | Value |
|---|---|
| Total loan book | GBP 4.2B |
| ECL (current) | GBP 124M (Stage 1: GBP 18M, Stage 2: GBP 62M, Stage 3: GBP 44M) |
| CET1 ratio | 11.2% (current), 10.8% (after latest ECL movement) |
| Leverage ratio | 4.6% |
| LCR | 142% |
| NSFR | 108% |
| AML metrics | 847 alerts in queue, 12 SARs filed this quarter, 2 MRAs outstanding |
| Key concentration | CRE: GBP 680M with 14% in Stage 2; Consumer: early DPD stress |
Macroeconomic Overlay Scenarios
| Scenario | Probability | GDP | Unemployment | Key Assumption |
|---|---|---|---|---|
| Base | 50% | +1.2% | 4.8% | Gradual recovery |
| Adverse | 35% | -0.8% | 6.5% | Mild recession, property correction |
| Severe | 15% | -3.0% | 9.1% | Deep recession, property crash |
Phase 1: ECL Movement Dashboard
Calculate the ECL movement from the prior quarter:
| Component | Prior Quarter | Current Quarter | Movement |
|---|---|---|---|
| Stage 1 ECL | GBP 15M | GBP 18M | +GBP 3M |
| Stage 2 ECL | GBP 55M | GBP 62M | +GBP 7M |
| Stage 3 ECL | GBP 40M | GBP 44M | +GBP 4M |
| Total ECL | GBP 110M | GBP 124M | +GBP 14M |
Analyse: What is driving the GBP 14M increase? Is it stage migration, PD deterioration, LGD changes, or new originations? What does the CRE concentration (GBP 680M with 14% in Stage 2) tell you about the portfolio trajectory?
Phase 2: Capital Dashboard
| Metric | Minimum | + Buffer | Current | Status |
|---|---|---|---|---|
| CET1 ratio | 4.5% | 7.0% (+ CCB) | 10.8% | Above buffer |
| Tier 1 ratio | 6.0% | 8.5% | ? (calculate) | |
| Total Capital ratio | 8.0% | 10.5% | ? (calculate) | |
| Leverage ratio | 3.0% | — | 4.6% | Above minimum |
The CET1 ratio has moved from 11.2% to 10.8% — a 40bp decline. What caused it? (The GBP 14M ECL increase, post-tax at 75%, reduces retained earnings by GBP 10.5M.)
Phase 3: Liquidity Dashboard
| Metric | Minimum | Current | Headroom |
|---|---|---|---|
| LCR | 100% | 142% | 42pp |
| NSFR | 100% | 108% | 8pp |
The NSFR headroom is thin at 8 percentage points. What would cause it to breach 100%? (Large wholesale funding maturity without replacement, or significant increase in long-term lending without matching stable funding.)
Phase 4: AML Dashboard
| Metric | Value | Trend | Concern Level |
|---|---|---|---|
| Open alerts | 847 | Up 12% QoQ | Medium |
| SARs filed | 12 | Up from 8 last quarter | Medium-High |
| MRAs outstanding | 2 | Unchanged | High |
| Average alert age | 4.2 days | Up from 3.8 days | Medium |
The 2 outstanding MRAs (Matters Requiring Attention from regulators) are the highest-priority item. What are typical MRA themes for UK banks? (Transaction monitoring effectiveness, PEP screening gaps, beneficial ownership documentation.)
Phase 5: Integrated Stress Test
Apply the adverse scenario (GDP -0.8%, unemployment 6.5%, probability 35%) combined with an AML fine of GBP 50M:
| Impact | Calculation | Result |
|---|---|---|
| ECL increase under adverse | ECL rises from GBP 124M to approximately GBP 180M | +GBP 56M |
| Post-tax CET1 impact (ECL) | GBP 56M x 75% = GBP 42M | -GBP 42M |
| AML fine (post-tax) | GBP 50M x 75% = GBP 37.5M | -GBP 37.5M |
| Combined CET1 impact | -GBP 79.5M | |
| Stressed CET1 | Starting CET1 minus GBP 79.5M | Calculate |
| Stressed CET1 ratio | Stressed CET1 / RWA (assume RWA increases 5% under stress) | Calculate |
Does the bank breach the combined buffer under the adverse + AML fine scenario? What management actions are available?
Phase 6: Board Risk Report — 10 Slides
Assemble your analysis into a 10-slide Board Risk Report:
| Slide | Content |
|---|---|
| 1 | Executive Summary: Key metrics, traffic-light status, top 3 risks |
| 2 | ECL Movement: Dashboard from Phase 1 with trend chart |
| 3 | Stage Migration Analysis: Where is credit quality deteriorating? |
| 4 | Capital Adequacy: Dashboard from Phase 2, buffer headroom |
| 5 | Liquidity Position: LCR and NSFR from Phase 3, NSFR concern |
| 6 | AML and Financial Crime: Dashboard from Phase 4, MRA status |
| 7 | Integrated Stress Test: Results from Phase 5, survival assessment |
| 8 | Concentration Risk: CRE GBP 680M deep-dive, consumer early stress |
| 9 | Recommendations: 5 specific actions with owners and timelines |
| 10 | Appendix: Methodology notes, data sources, model limitations |
Part 3: Chapter Summary — Five Principles of Banking Domain AI
This chapter covered three regulatory pillars, 14 exercises, and a full cross-pillar capstone. Five principles emerged from this work.
Principle 1: Model governance applies to AI models. The same regulatory framework (SR 11-7 in the US, SS1/23 in the UK) that governs traditional risk models applies to AI models used in banking. An AI agent that calculates ECL or stages assets is a model, and it requires validation, back-testing, documentation, and approval just like a statistical model.
Principle 2: SICR assessment is irreducibly human. The AI can compute PDs, identify DPD thresholds, and flag qualitative indicators. But the decision of whether a significant increase in credit risk has occurred — particularly for facilities like A004 (job loss, no DPD) and A006 (property decline, no rating change) — requires professional judgment that cannot be fully automated. The agent assists; the human decides.
Principle 3: AML is a legal obligation, not a data exercise. Filing a SAR is a legal act with criminal consequences for failure. The tipping-off prohibition is a criminal offence. The agent can gather data, screen transactions, and identify typologies, but it must never file a SAR, communicate suspicion to customers, or make risk acceptance decisions. These boundaries are not optional features — they are legal requirements.
Principle 4: The Basel IV output floor reshapes the industry. The 72.5% output floor means that IRB banks can no longer achieve dramatically lower RWA than SA banks for the same portfolio. This compresses the capital advantage of IRB models, changes the economics of low-risk lending (mortgages, sovereigns), and means that the capital calculation you did in Exercise 3 becomes increasingly relevant even for IRB banks.
Principle 5: Pillar interaction is where insight lives. A bank that reports ECL, capital, and AML separately gets three answers. A bank that traces the cascade — provision increase reduces CET1, which tightens the buffer, which constrains the dividend, which signals market concern, which increases funding costs, which compresses NII — gets one integrated answer. The banking plugin's router exists specifically to make this integration automatic rather than manual.
Try With AI
Use these prompts in Claude or your preferred AI assistant to complete the capstone.
Prompt 1: Building the Skill Library
Help me build 4 banking skills using Method A (expert interview):
1. ifrs9-ecl: Calculate Expected Credit Loss. Must handle Stage 1
(12-month), Stage 2 (lifetime), Stage 3 (lifetime impaired).
Support scenario weighting. Key validation: weighted ECL >= base.
2. ifrs9-staging: Stage classification using SICR triggers. Must
handle quantitative (30 DPD, 2-notch downgrade) and qualitative
(job loss, collateral decline) triggers. Flag edge cases.
3. basel-rwa-credit: Credit risk RWA under SA and IRB. Must handle
all SA risk weights, off-balance-sheet CCFs, and the 72.5%
output floor.
4. aml-typologies: Identify money laundering typologies from
transaction patterns. Must recognize structuring, layering,
trade-based ML, PEP abuse. Must NOT recommend SAR filing.
For each skill, generate the SKILL.md with: persona, key
decisions, input/output specification, and validation criteria.
What you are learning: Building skills from your own expertise is Method A — the expert interview approach from Chapter 15. By structuring what you learned in Lessons 3-10 into reusable SKILL.md files, you transform knowledge into operational capability. The skill does not replace your judgment — it packages your analytical framework so that routine calculations are automated and edge cases are flagged for human review.
Prompt 2: Executing the Capstone Scenario
Execute a Board Risk Report for this bank:
Loan book: GBP 4.2B
ECL: GBP 124M (S1 GBP 18M, S2 GBP 62M, S3 GBP 44M)
Prior quarter ECL: GBP 110M (S1 GBP 15M, S2 GBP 55M, S3 GBP 40M)
CET1 ratio: 10.8% (down from 11.2%)
Leverage: 4.6%, LCR: 142%, NSFR: 108%
AML: 847 alerts, 12 SARs, 2 MRAs
CRE concentration: GBP 680M, 14% Stage 2
Macro scenarios: Base (GDP +1.2%, 50%), Adverse (GDP -0.8%, 35%),
Severe (GDP -3.0%, 15%)
Run all 5 phases:
1. ECL Movement (what drove the GBP 14M increase?)
2. Capital Dashboard (all ratios vs minimums + buffers)
3. Liquidity (LCR/NSFR headroom analysis)
4. AML Dashboard (alert trends, MRA priorities)
5. Integrated Stress (adverse scenario + GBP 50M AML fine)
Then produce a 10-slide Board Risk Report outline with the key
message for each slide.
What you are learning: The Board Risk Report is the ultimate cross-pillar deliverable. It forces you to integrate ECL, capital, liquidity, and AML into a single narrative that a non-technical board member can understand. The integrated stress test (adverse macro combined with AML fine) demonstrates why cross-pillar analysis matters — a bank that stress-tests ECL and AML separately misses the compounding effect of simultaneous shocks.
Prompt 3: Validating the Five Principles
For each of the five principles of banking domain AI, provide:
1. A one-sentence explanation of the principle
2. A specific example from the Chapter 21 exercises that
demonstrates it
3. What goes wrong if the principle is violated
The five principles:
1. Model governance applies to AI models (SR 11-7)
2. SICR assessment is irreducibly human
3. AML is a legal obligation, not a data exercise
4. Basel IV output floor reshapes the industry
5. Pillar interaction is where insight lives
For principle 5, trace a specific cascade: a GBP 50M fraud
discovery through all three pillars, quantifying each impact.
What you are learning: The five principles are the lasting takeaways from this chapter. Technical skills — ECL calculation, capital ratios, AML screening — can be looked up and re-learned. But the principles shape how you think about banking AI: what must be automated, what must remain human, and where the integration of multiple regulatory views creates insight that no single pillar can provide.