Deutsche Bank

At Deutsche Bank Research, analysts responsible for delivering original economic and financial analysis faced a structurally time-consuming workflow before any publishable insight could reach market. Every report began with manual collection and cross-referencing of financial statements, regulatory filings (including SEC documents), and industry reports — followed by synthesis across internal research archives and the broader global context. The volume of unstructured document processing created a consistent bottleneck, restricting both the number of topics analysts could cover and the depth achievable within any given report. For a global investment bank operating across Research, Investment Bank Origination & Advisory, and Fixed Income & Currencies, the cumulative drag — and growing dependence on outsourcing to supplement analyst capacity — carried real competitive and operational cost.

Deutsche Bank built DB Lumina, an internally developed AI research agent deployed on Google Cloud in September 2024. The platform delivers three integrated capabilities: a Gemini-powered conversational interface supporting document upload, summarization, translation, proofreading, and content drafting; a prompt template system that standardizes high-volume document processing workflows — extraction, summarization, earnings note creation — without requiring analysts to engineer prompts individually; and a RAG architecture grounded in enterprise knowledge sources including internal research, SEC filings, and proprietary document repositories, with inline citations and source viewers for verification. The infrastructure spans GKE for microservice orchestration, Vertex AI for multimodal AI inference, Cloud SQL with pgvector for vector storage, Dataflow for document ingestion and embedding, and the Discovery Engine API for retrieval. Guardrails, audit logging via Cloud Storage and BigQuery, and Azure AD–integrated authentication enforce compliance end-to-end.

DB Lumina reached approximately 5,000 analysts across Deutsche Bank Research, Investment Bank Origination & Advisory, and Fixed Income & Currencies within its first year of production, with a roadmap to expand beyond 10,000 users. Measured time savings upon rollout:

• Up to 2 hours saved per full research report
• 30–45 minutes saved per earnings note template

Beyond speed, one analyst expanded an earnings report's depth by 50% by adding regional sections and forecast summaries that would have been impractical under the prior manual workflow. Editorial and grammatical accuracy across analyst notes improved noticeably post-launch, and Deutsche Bank's Global COO for Investment Research described the adoption as the foundation of a broader AI-in-research program.

• RAG grounded in authoritative enterprise sources — SEC filings, proprietary research archives — is non-negotiable in regulated financial environments; citation precision directly determines analyst trust and compliance defensibility.
• Prompt templates bridge the gap between raw LLM capability and consistent team output; standardizing them across roles reduces quality variance without constraining individual analyst workflows.
• Evaluating generative AI in financial services requires domain-specific metrics beyond standard benchmarks — citation recall, false rejection rates, and verbosity control matter as much as accuracy.
• Enterprise rollout in a regulated institution requires identity integration, centralized authorization, and complete audit logging architected from day one, not retrofitted post-launch.