info@marzallabs.ai
Modern skyscrapers in London’s financial district

Financial Services

Document intelligence that
survives supervisory scrutiny

Built against FCA AI principles, PRA SS1/23 model risk expectations and UK GDPR — in one framework rather than three projects.

  • Asset management
  • Banking
  • Capital markets
  • Lending
  • FCA-regulated
  • PRA-supervised

Consumer Duty raised the evidential bar on how decisions affecting customers are made and recorded. PRA SS1/23 set expectations on model risk that most firms are still building towards. The EU AI Act adds a risk classification obligation on top of both, and the ICO is clear that automated processing carries its own transparency duties. The awkward part is that these arrive as four separate frameworks with four different vocabularies for what is largely the same underlying control.

We map all four into a single register, then design the controls once. Explainability, audit trails, human review thresholds and model inventory are specified into the architecture rather than retrofitted before an inspection — which is both cheaper and, in our experience, the only version that survives contact with a supervisor.

Capabilities

What we build for financial services

Each capability below is delivered through named assets that already exist. Figures are delivered results unless marked otherwise.

Regulatory document intelligence

Extraction, comparison and classification across policies, contracts and regulatory correspondence.

  • Clause-level comparison with severity rating on every finding
  • Every output cited back to source page and section
  • Deployed inside your own tenancy, documents never leave it

Model risk management

Model inventory, tiering, validation expectations and monitoring built against PRA SS1/23.

  • Model inventory with named ownership per tier
  • Validation and monitoring expectations set per risk tier
  • Written for a supervisory review rather than an internal one

Explainability and audit trails

XAI on every automated output, with the decision record a supervisor will actually ask for.

  • 100% XAI trace coverage in our production systems
  • Every source, weighting and human decision timestamped
  • Override capture on every human review

Governed data platforms

Medallion lakehouse foundations with lineage, quality testing and a unified semantic layer.

  • 14 source systems into 1 at a comparable engagement
  • 180+ automated tests on every pipeline refresh
  • Full lineage, so any figure traces to its source

Consumer Duty evidence

Decision records structured so outcomes for customers can be evidenced rather than asserted.

  • Decision rationale retained with each automated output
  • Human review documented against the decision it relates to
  • Reporting formatted for a board pack from the outset

Value measurement and ROI

Baselines, KPIs and attribution defined before deployment, not reconstructed after it.

  • Baselines captured pre-deployment as standard
  • Time, accuracy and efficiency gains attributed explicitly
  • 3-week fixed-scope framework engagement

Named assets

The inventory we draw on

Twenty-one products, diagnostics, agent patterns, platform components and governance frameworks. Filter by category.

  • PolicyCompareIQClause-level policy and slip comparison
  • ClassifierIQAsset classification with a firm-specific memory loop
  • D&O AnalyzerIQEight-vector D&O underwriting research
  • TriageIQSubmission intake scoring and routing — 2026
  • AI-Ready BlueprintThree-week fixed-scope data estate diagnostic
  • AI Readiness AssessmentData, technology, capability and governance scoring
  • Data & AI Roadmap DesignSequenced, costed initiative plan with owners
  • AI Operating Model DesignReviewer roles, escalation paths, HITL workflow design
  • Technology Stack AdvisoryPlatform selection against your estate and constraints
  • AI Agent Architecture & DesignTask boundaries, tool permissions, failure modes
  • Document Processing AgentsMulti-format ingestion across email, PDF and ACORD
  • Underwriting & Triage AgentsAppetite scoring, capacity checks, context packs
  • Claims Processing AgentsClassification, routing and loss-narrative extraction
  • RAG Knowledge BasesRetrieval over your wordings, guidance and precedent
  • Modern Data Platform DesignMedallion architecture on Fabric, Databricks or Snowflake
  • Data Governance FrameworkOwnership, lineage and automated quality rules
  • Data Quality Programmes180+ automated tests running on every pipeline refresh
  • Underwriting & Claims AnalyticsSemantic layer and agreed portfolio definitions
  • Databricks Partnership ServicesLakehouse delivery under the Bronze partnership
  • AI Compliance & GovernanceFCA, Lloyd’s, EU AI Act and GDPR in one register
  • ROI & Value FrameworkBaselines, KPIs and attribution set before build

Partnerships

The platforms behind the delivery

Databricks Bronze partner, with production delivery on every platform named here.

  • Microsoft AzureCloud, HITL pipelines, Azure OpenAI
  • Azure AI FoundryModel catalogue and agent tooling
  • DatabricksBronze partner · Delta Lake and ML
  • Google CloudCloud platform and data services
  • Anthropic ClaudeDocument reasoning
  • OpenAIExtraction and analysis
  • Microsoft FabricUnified analytics and OneLake
  • SnowflakeData cloud analytics
  • LangChainAgent orchestration
  • dbtTransformation and testing

Meet four overlapping regulatory obligations with one governance framework and one evidence base.

Talk now