Financial institutions have long relied on structured, third-normal-form reference models to tame complexity across customers, accounts, transactions, instruments, and risk. Those models still matter, but today’s landscape adds real-time decision-making, multimodal data (including documents, chat, and voice), and AI systems that learn, reason, and act. This article reframes classic financial-services data patterns into an AI-first blueprint: one that treats data models not just as schemas, but as living, governed “data products” that power analytics, machine learning, and agentic workflows.
Principles for an AI-First Data Model
Canonical domains, productized
Retain the canonical domains Party/Customer, Account, Relationship, Instrument, Order/Trade, Position, Transaction, Collateral, Risk Exposure—but publish them as data products with owners, SLAs, contracts, and versioned interfaces. Consumers (dashboards, feature pipelines, LLM retrieval, event processors) integrate via contracts rather than point-to-point ETL.
Separation of concerns: facts, metrics, features, and narratives
Facts: gold, immutable records (e.g., posted transactions).
Metrics: curated business logic (e.g., delinquency rate, VAR).
Features: ML-ready signals with lineage (e.g., rolling average spend, merchant entropy).
Narratives: LLM-generated artifacts with sources and citations (e.g., KYC case summaries). The model explicitly represents each layer and its dependencies with other layers.
Event-centric and real-time by default
Model high-value business moments (PaymentInitiated, TradeExecuted, AlertRaised, ConsentUpdated). Events unlock streaming features, low-latency fraud/risk decisions, and accurate replay.
Governance built-in
Every artifact is tagged with its owner, purpose, PII classification, consent policy, lineage, and retention . Models capture compliance posture (GDPR, AML/KYC) as first-class references, not afterthoughts.
Target Architecture: Lakehouse + Streams + Graph + Vectors
Lakehouse for durable facts and batch analytics.
Streaming bus (CDC from cores, event ingestion) for near-real-time features and alerts.
Knowledge graph (entities + relationships across customers, devices, merchants, counterparties) for AML and credit explainability.
Vector indexes for LLM retrieval over documents, policies, procedures, tickets, and customer communications.
Feature store with offline/online parity; Model Registry and Prompt/Tool Registry for LLMs and agents.
Semantic layer that defines governed metrics consumed by BI and agents alike.
AI-Native Representations to Add
Augment the traditional entities with AI-ready ones.
FeatureDefinition (id, name, owner, input_contracts, transform_code_ref, pii_flags, bias_notes, last_backfill)
FeatureValue (entity_key, feature_id, ts, value, offline/online_source)
EmbeddingIndex (index_id, domain, modality: doc/chat/call, vector_dim, partition_keys)
EmbeddedChunk (index_id, chunk_id, vector, text_ref, source_uri, pii_masking_policy)
PromptTemplate (id, purpose, inputs_schema, guardrails_ref, evaluation_suite)
ToolDefinition (id, name, contract, rate_limits, privacy_scope)
ModelCard (model_id, version, training_data_refs, risks, intended_use, approval_status)
ModelRun (run_id, model_id, dataset_ref, metrics, fairness_audit, lineage_hash)
Policy/Consent (policy_id, data_class, allowed_uses, legal_basis, expiry)
Case/Alert (case_id, entity, reason_codes, evidence_refs, disposition, narrative_ref).
These entities enable you to track how inputs become features, how features inform models or LLM prompts, and how decisions are explained and audited.
Modernized Canonical Domains (Condensed)
Party / Customer / Counterparty: unify identification, KYC attributes, risk scores, consents, communication preferences, and linkages to digital identities/devices.
Account / Wallet / Contract: balances, limits, entitlements, product terms, pricing, and regulatory classification.
Transaction / Payment / Transfer: immutable ledger events with channels, devices, geolocation, counterparties; link each transaction to graph nodes.
Instrument / Position / Order / Trade: full lifecycle with market data refs, exposures, haircuts, netting sets; align with risk and collateral data products.
Risk Exposure / Limit / Breach: Unify credit, market, and liquidity exposures; stream breaches as events to downstream controls.
Document / Interaction: contracts, statements, tickets, emails, chats, call transcripts indexed into EmbeddingIndex for retrieval and summarization.
Case / Alert / Investigation: structured workflow entity connected to evidence (transactions, graph paths, documents) and narrative outputs from LLMs.
From Reference Model to AI Feature Space
A good rule: one “feature view” per decision. Examples.
Card-present fraud: velocity features, merchant diversity score, device trust, graph-distance to known bad nodes, and typical spend embedding similarity.
AML monitoring: counterparty risk score, transaction pattern anomaly, graph motifs (fan-in/fan-out), jurisdictional risk, LLM-extracted purpose-of-payment signals.
Credit underwriting: income/expense ratios, utilization, cash-flow stability features, employment signals derived from unstructured docs, and explainability anchors.
Each feature references its upstream contracts and bias/PII flags; each model run writes explainability artifacts (feature importances, reason codes, SHAP slices) and policy checks (fair lending, disparate impact).
LLM & Agent Patterns that Belong in the Model
RAG for regulated answers: connect PromptTemplate → Retrieval plan → EmbeddingIndex → source attestation; store citations and confidence.
Case narrative generation: inputs = Case, Evidence (transactions, graph paths, documents), Policy; output = Narrative with footnotes and redaction mask.
Agentic workflows with tools: O rchestrate tools (risk limits API, pricing, sanctions screen) with policies that constrain data scope and logging. Persist the ToolCall trail for audit.
Governance, Risk, and Compliance as Data
Represent controls as data, too.
DataContract (producer, schema_hash, SLAs, PII flags, test suite)
ControlTest (id, scope, cadence, last_run, status, evidence_uri)
PIIMap (column→classification, masking_rule, consent_policy)
Red-team Findings (model_id, vector_index, prompt_template, issue_type, severity, remediation_ref)
This makes governance queryable and automatable: “show all features used in production credit models that depend on unexpired consent and passed fairness last month.”
Four End-to-End Exemplars
1. Digital Onboarding & KYC
Capture identity documents and interactions in Document/Interaction; embed into EmbeddingIndex.
Extract entities (sanctions status, address quality), write to Party with confidence scores.
The decision utilizes features such as “document consistency,” “device trust,” and “watchlist proximity.”
All prompts, evidence, and decisions are logged under the Case with explainability artifacts.
2. AML Transaction Monitoring
Stream Transaction events; run graph analytics to surface unusual motifs.
LLM generates factual narratives from evidence with citations; analysts review in Case UI.
Dispositions close the loop and feed ModelRun performance, as well as the next retraining window.
3. Credit & Pricing
Combine cash-flow features from Transaction, bureau data, and employment signals extracted from documents.
Keep a Policy map for what can/can’t be used in a jurisdiction.
Write reason codes and SHAP summaries to ModelRun for adverse-action letters.
4. Customer 360 & Next Best Action
Merge metrics (CLV, churn risk) with LLM-summarized tickets and emails.
RAG answers for agents use Ethe mbeddingIndex restricted by Consent and Policy.
Actions and offers are events joined back to outcomes for uplift measurement.
Example Minimal Schemas (Illustrative)
// Transaction (fact)
{ "tx_id":"...", "account_id":"...", "posted_ts":"...", "amount":123.45, "currency":"USD",
"channel":"POS", "merchant_id":"...", "device_id":"...", "counterparty_id":"...", "geo":{"lat":...,"lon":...} }
// EmbeddedChunk (for RAG)
{ "index_id":"aml_docs_v1", "chunk_id":"...", "vector":[...], "text_ref":"s3://docs/...",
"source_uri":"...", "pii_masking_policy":"MASK_NAMES" }
// FeatureValue (online)
{ "entity_key":"acct:123", "feature_id":"v2_spend_velocity_24h", "ts":"...", "value":7.0 }
// Case linking evidence
{ "case_id":"KYC-2025-000123", "entity":"party:987", "reason_codes":["ADDR_MISMATCH"],
"evidence_refs":["tx:...","doc:...","graphpath:..."], "narrative_ref":"nar:..." }
Operating the Model: MLOps + LLMOps
Contracts & tests at every interface (data, features, prompts, tools).
Lineage from raw facts to narratives; hash and store to make audits fast.
Monitoring: drift (features, embeddings), safety (prompt jailbreaks), and business KPIs.
Privacy: consent-aware retrieval, data minimization, field-level masking, and jurisdiction routing.
Responsible AI: fairness checks per model and segment; red-teaming for LLMs; human-in-the-loop for high-risk decisions.
Metrics That Matter
Time-to-signal (ingest→feature latency), coverage of contracts under test, model SLAs met, reduction of false positives in AML, fraud catch rate at fixed customer friction, credit decision turnaround, narrative accuracy compared to human benchmark, percentage of retrieval answers with citations, and audit cycle time.
Implementation Path (Pragmatic)
Phase 1 (0-90 days): stand up lakehouse, event ingestion for 1–2 journeys (e.g., payments, KYC), a small feature store, a single EmbeddingIndex for policy/KYC docs, and contracts/lineage.
Phase 2 (90-180 days): wire graph, expand online features, add Case/Narrative workflows, adopt model registry and prompt registry, embed governance queries into ops dashboards.
Phase 3 (beyond): Scale to cross-domain decisions, multi-jurisdiction consent routing, agentic processes with tool governance, and continuous model/prompt evaluation.
Conclusion
An AI-first financial-services data model extends familiar domains with events, features, vectors, graphs, and governance that is explicit and queryable. The destination is not a single monolithic schema, but a portfolio of well-owned data products that feed analytics, machine learning, and compliant LLM workflows, each one explainable, auditable, and ready for real-time decisions.
Join the conversation! Your thoughts help the community grow.