Tier 04 · Cognitive
ICAAP & ORSA, in days.
Glass-Box AI that reads a full ICAAP package, applies regulatory frameworks, reasons across 12 dimensions, and produces a defensible rating — inside a cockpit where every case, citation and judgment lives on one screen.
Risk trend — 12 months
Every citation copied verbatim from law
The Supervisory Cockpit
One screen. Every case. Every citation.
A sovereign, real-time AI cockpit where intake triage, background screening, fit-&-proper decisions and market-conduct analysis run side by side — every AI finding grounded in a verbatim statutory citation, every run logged, every prompt visible.
Beneficial-owner threshold requires enhanced due diligence.
Banking Act · s.14(3) — verbatim
Source of funds narrative incomplete.
NO_GROUNDING confidence halved · human review
Self-sufficient
Runs entirely on the regulator's own infrastructure — local LLM inference, local embeddings, local retrieval corpus. No cloud, no API keys, no data egress.
Intelligent
Not a viewer — an analyst. The cockpit triages cases, screens adverse media, drafts fit-&-proper determinations and classifies conduct risk, each with confidence scores and human confirm/override gates.
Real-time & accountable
A live “Recent runs & metrics” panel shows every AI run — UI- or API-triggered — with its evaluated metrics. Glass Box is not a slogan; it is a panel on the landing page.
Eight capability blocks
Not a dashboard. An operating environment.
Each block maps 1:1 to a working cockpit service.
Intelligent Case Triage & Routing
Drop a licence-application export → risk tier + routing + findings, every finding citing an exact statutory provision copied verbatim, never invented.
/cockpit/licensingAdverse Media & Background Screening
Automated adverse-media entity linkage + AI-assisted review with a supervisor-configurable threshold slider.
/cockpit/screeningFit & Proper Decision Support
Per-regulator determinations; the deterministic screening layer stays instant while the AI writes the narrative.
/cockpit/fit-properMulti-Regulator Segregation
One subject, two regulator-scoped assessments with attribute-based access control — a scoped view can never expose the other regulator's data. Twin Peaks-ready.
/cockpit/multi-regulatorMarket Conduct AI (advisory)
Three-layer governance: AI suggests, analyst confirms or overrides, the recorded value is the analyst's.
/cockpit/mcs-conduct-aiAsk-the-Corpus Assistant
A grounded natural-language assistant that answers “why was this entity flagged?” using the corpus plus the current page's result.
RAG chat · every pageSource Finder
Type a query, see the exact retrieved passages — retrieval transparency as a first-class feature.
/cockpit/sourcesSupervisor-owned AI
Supervisors can open, read and edit the system prompt behind every AI workflow — “the AI works for you, on your instructions.”
Editable prompt registryCapabilities
What’s included in Tier 4
RAG-augmented reading
Retrieval-augmented generation ingests the full regulatory submission, parses every page, and grounds every AI judgment in the institution's actual data with cited source paragraphs.
12-dimension reasoning
Multi-framework scoring engine analyses risk across credit, market, operational, liquidity, and eight additional dimensions. Each score includes a structured rationale and evidence chain.
Glass-Box explainability
Every risk judgment is fully traceable. Source citations, reasoning chains, prompt audit trails, and human-in-the-loop gates at every stage. No black boxes in supervisory decisions.
Traceability isn't a policy here — it's enforced by the pipeline itself. ↓The differentiating technical story
The Citation Guardrail
A citation is a property of a retrieved document, never model text. A finding the corpus cannot support is structurally impossible to cite.
Step 01
Section-level indexing
Each statute is chunked one subsection per passage; the citation label (act · section · subsection) is composed from the document's own metadata.
Step 02
Hybrid retrieval
Dense + keyword retrieval surfaces only real passages from the regulator's own corpus — the model may reference nothing else.
Step 03
Grounded generation
The finding is written against retrieved passages; every claim carries the passage's own citation label — not one written by the AI.
Step 04 · Guardrail
Citation guardrail
Unsupportable finding → marked, confidence halved, routed to human review.
NO_GROUNDING
A hallucinated legal citation isn’t caught by review — it’s prevented by architecture.
Assessment lifecycle
25 banks, one ICAAP pipeline
- 1
Batch ingestion
25 submissions parsed — structured data + narrative, OCR
- 2
12-dimension risk scoring
RAG-augmented, explainable — flags qualitative weakness
- 3
Supervisory review board
Multi-assessor review · human sign-off on every rating
- 4
Assessment report
Composite rating + intervention ladder + full audit trail
Time compression — canonical claim
Day in the life
How Tier 4 transforms daily work
Anna Kowalska
Head of prudential supervision
“An ICAAP assessment that consumed our team for three months now produces a preliminary supervisory report in two days. The AI reads every page, and I can trace every risk score back to the source paragraph.”
9:00 AM — Cockpit triage
Opens the Cockpit. Overnight triage has already routed three intakes and flagged one NO_GROUNDING finding for human review.
11:00 AM — Dimension review
The 12-dimension scoring run for the pending ICAAP is on screen, every judgment carrying its source citation.
2:00 PM — Corpus chat
Asks the corpus chat a follow-up — “why was this entity flagged?” — and gets an answer grounded in law plus the result on screen, ready for the board pack.
4:00 PM — Review board
Review board convenes. The evidence chain — findings, citations, overrides, prompts — exports straight from the Cockpit.
Outcome
Full ICAAP supervisory assessment completed in days instead of months, with 100% explainable AI and complete audit trail for every risk judgment. And between assessments, nothing goes dark — the Cockpit keeps triaging, screening and logging every day in between.
Outcomes
Measurable impact from day one
Architecture
Where Tier 4 sits
Use cases in action
See Tier 4 at work
Platform tiers
Explore other tiers
Core Regulatory Functions
Entity registry, 12-dimension risk engine, and supervisor workspace.
Intelligent Process Automation
Intelligent intake, completeness verification, and analyst triage.
Data Governance & BI-Driven Supervision
Peer benchmarks, sector heatmaps, and early-warning triggers.
Frequently asked questions
Can supervisors edit the AI's instructions in CORA?
Yes. CORA Tier 4 exposes an editable prompt registry: supervisors can open, read, and edit the system prompt behind every AI workflow. Advisory outputs pass a confirm-or-override gate, and the recorded value is always the analyst's — the AI works for the supervisor, on the supervisor's instructions.
Book a Tier 4 walkthrough
See how CORA Tier 4 automates ICAAP and ORSA assessment with sovereign, explainable Glass-Box AI.
Schedule a demoEvery insight starts from governed data → Tier 3