RAG systems Skills for underwriter in investment banking: What to Learn in 2026

AI is changing underwriting in investment banking in one specific way: it is compressing the time between raw deal materials and a decision-ready credit view. The underwriter who can still read a CIM, reconcile financials, extract covenants, and defend assumptions will stay valuable; the one who only knows manual document review will get squeezed.

The winning move in 2026 is not “learn AI” in the abstract. It is learning how to work with retrieval-augmented generation, document pipelines, and validation workflows so you can review more deals faster without losing judgment.

The 5 Skills That Matter Most

1. •

   Financial-document retrieval and search design

   Underwriters spend too much time hunting through offering memoranda, lender presentations, debt schedules, audited statements, and legal docs. RAG systems are only useful if they can pull the right source paragraph fast and reliably, so you need to understand chunking, metadata, and document indexing.

   For an underwriter in investment banking, this means building systems that answer questions like: “Where is the leverage covenant defined?” or “Which filing contains the adjusted EBITDA bridge?” If the retrieval layer is weak, everything downstream is garbage.

2. •

   Prompting for structured credit analysis

   You do not want chatty summaries. You want outputs that map to underwriting work: leverage ratios, liquidity headroom, covenant triggers, sources of repayment, and key risks by deal section.

   Learn how to force LLMs into structured templates such as JSON or tables. That lets you turn unstructured deal books into repeatable analysis instead of relying on free-form prose that no one can audit.

3. •

   Source verification and citation discipline

   In banking, hallucinated facts are not a nuisance; they are a control failure. You need to know how to make every answer trace back to source documents with page numbers or section references.

   This matters because underwriters have to defend conclusions in credit committees and model reviews. A RAG system that cannot show evidence from the CIM or 10-K is not usable in a regulated workflow.

4. •

   Workflow automation around deal intake

   The real value is not a chatbot sitting on top of PDFs. It is automating repetitive steps around intake: classifying documents, extracting key fields, flagging missing items, and routing exceptions for human review.

   For an underwriter in investment banking, this reduces the time spent on first-pass screening and frees you up for judgment calls on structure, sponsor quality, downside cases, and intercreditor issues. Learn enough Python or low-code automation to stitch these steps together.

5. •

   Risk controls for AI-assisted decisions

   Underwriting requires accountability. You need to understand evaluation methods: precision/recall for extraction tasks, citation accuracy, false positive rates on risk flags, and when to force human override.

   This skill separates toy demos from production use. If you can explain where the system fails and how it is monitored, you become the person who can safely bring AI into a bank’s underwriting process.

Where to Learn

• •

  DeepLearning.AI — Retrieval Augmented Generation (RAG) course

  • •Best for understanding retrieval pipelines, embeddings, chunking, and practical RAG architecture.
  • •Take this first if you want the mechanics behind document Q&A.
• •

  OpenAI Cookbook

  • •Useful for structured outputs, function calling patterns, evaluation ideas, and prompt design.
  • •Good reference when you want your model output to look like an underwriting memo instead of a conversation.
• •

  LangChain documentation

  • •Strong for building document loaders, retrievers, agents, and workflow glue.
  • •You do not need every feature; focus on document ingestion and retrieval chains.
• •

  Hugging Face course

  • •Helpful for understanding models, tokenization, embeddings basics, and open-source tooling.
  • •Worth doing if your bank prefers self-hosted or vendor-neutral options.
• •

  Book: Designing Machine Learning Systems by Chip Huyen

  • •Not RAG-specific, but excellent for thinking about evaluation, monitoring, data quality, and deployment tradeoffs.
  • •This is the book that helps you avoid building something clever but unusable in production.

A realistic timeline:

• •Weeks 1–2: Learn RAG basics and build simple document search over sample deal docs.
• •Weeks 3–4: Add structured extraction for financial metrics and covenant language.
• •Weeks 5–6: Add citations, evaluation checks, and exception handling.
• •Weeks 7–8: Package one end-to-end underwriting workflow demo with human review points.

How to Prove It

• •

  Deal book Q&A assistant

  • •Build a tool that answers underwriting questions from a set of CIMs or lender decks with citations.
  • •Example queries: leverage ratio definition, debt maturity schedule, sponsor ownership changes.
  • •This proves retrieval quality and source discipline.
• •

  Covenant extraction pipeline

  • •Create a workflow that pulls covenant thresholds from credit agreements into a table.
  • •Include fields like test dates، baskets، cure rights، springing covenants.
  • •This shows you can convert legal text into usable underwriting inputs.
• •

  First-pass credit memo generator

  • •Feed in company filings or offering materials and produce a draft memo with sections like business overview، capital structure، risks، repayment sources.
  • •Keep it clearly marked as draft and cite every claim.
  • •This demonstrates structured prompting plus auditability.
• •

  Missing-data checker for deal intake

  • •Build a checklist tool that compares received documents against required underwriting artifacts.
  • •Flag missing audited statements، debt schedules، management projections، lien searches.
  • •This proves you understand operational friction in real deal flow.

What NOT to Learn

• •

  Generic chatbot building with no underwriting use case

  A Slack bot that answers random finance trivia does not help your career. Banks care about controlled workflows tied to actual deal work.

• •

  Overfocusing on model training

  Most underwriters do not need to train foundation models. You need retrieval quality، structured outputs، evaluation، and controls far more than GPU-heavy experimentation.

• •

  Vague “AI strategy” content

  Reading endless think pieces about transformation will not make you useful on a live transaction. Spend your time on document pipelines，citation checks，and exception handling instead.

If you are an underwriter in investment banking trying to stay relevant in 2026，your edge is simple: combine domain judgment with enough RAG skill to make AI trustworthy around real deal documents. That is a six-to-eight-week learning path if you stay focused on practical systems instead of theory-heavy detours.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.