AI agents Skills for underwriter in investment banking: What to Learn in 2026

AI is changing underwriting in investment banking in a very specific way: it is compressing the time spent on document review, covenant analysis, market comps, and first-pass risk screening. The underwriter who still spends hours manually pulling data from offering memoranda, credit memos, and issuer filings will get outpaced by someone who can use AI to do the first 80% and spend their time on judgment, structure, and exception handling.

The 5 Skills That Matter Most

1. •

   Prompting for financial document extraction

   Underwriters live in PDFs, decks, and filings. The useful skill is not “prompt engineering” in the abstract; it’s writing prompts that reliably extract debt terms, covenants, leverage ratios, maturity walls, and risk flags from messy deal documents. If you can turn a 200-page package into a structured summary in minutes, you become much faster at screening deals and preparing credit notes.

2. •

   Python for deal data cleanup and analysis

   You do not need to become a software engineer, but you do need enough Python to clean issuer data, normalize financial statements, and compare historical performance across deals. This matters because AI outputs are only useful if you can validate them against source data and build repeatable checks. A basic workflow with pandas is enough to automate boring reconciliation work that currently eats your day.

3. •

   Working with LLMs inside controlled workflows

   In banking, the question is not whether a model can answer a question. It’s whether the workflow is auditable, permissioned, and safe enough for internal use. Learn how retrieval-augmented generation works, how to ground answers in approved documents, and how to design prompts that force citations back to source material so your team can trust the output.

4. •

   Risk judgment with model output verification

   AI will be good at summarizing; it will still be weak at knowing when something matters. A strong underwriter needs to spot hallucinated terms, stale market references, missing exceptions in legal language, and inconsistent assumptions in financial models. Your value shifts toward being the person who catches what the model missed before it reaches committee.

5. •

   Workflow automation for recurring underwriting tasks

   The biggest productivity gains come from automating repeatable steps: intake triage, document classification, covenant extraction, peer comp gathering, and memo drafting. Learn how to connect tools like Excel, Python scripts, internal APIs, or low-code automation platforms so that routine work flows without manual copying and pasting. For an underwriter, this is where AI stops being a novelty and starts saving real hours every week.

Where to Learn

• •

  Coursera — Generative AI with Large Language Models

  Good starting point for understanding how LLMs work without getting buried in research papers. Pair this with your day job by testing prompts on anonymized underwriting documents over 2-3 weeks.

• •

  DeepLearning.AI — ChatGPT Prompt Engineering for Developers

  Short and practical. Use it to learn structured prompting patterns that help with extraction from term sheets, debt docs, and management presentations.

• •

  Python for Everybody by University of Michigan (Coursera)

  If you’re weak on coding fundamentals, this is a clean entry point. Spend 4-6 weeks getting comfortable enough to manipulate CSVs and financial datasets.

• •

  pandas documentation + Kaggle micro-courses

  pandas is the tool that will help you clean deal data fast. The Kaggle Python and Pandas micro-courses are short enough to finish in a weekend each.

• •

  Book: Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow by Aurélien Géron

  You do not need all of it immediately. Read the chapters on model evaluation and data pipelines so you understand what makes an output trustworthy before using AI-generated analysis in underwriting workflows.

How to Prove It

• •

  Build a covenant extractor for loan docs

  Take public credit agreements or sample term sheets and create a tool that extracts leverage ratios, interest coverage tests, baskets, maturity dates, and change-of-control clauses into a table. Show before-and-after time savings versus manual review.

• •

  Create an underwriting memo summarizer with citations

  Feed in public issuer filings or offering memoranda and generate a one-page summary that includes business overview, key risks, debt stack details, and direct citations to source pages. This proves you can use LLMs without turning them loose unsafely.

• •

  Automate peer comp gathering for new deals

  Build a script or spreadsheet workflow that pulls comparable company metrics from public sources or exported datasets into a standardized template. Underwriters spend too much time formatting comps; automating this shows practical value fast.

• •

  Design a red-flag checklist bot

  Create a simple internal tool that scans deal notes for missing items like stale financials, covenant breaches, unusual add-backs, concentrated customer exposure, or weak liquidity assumptions. This shows judgment plus automation: exactly what modern underwriting needs.

What NOT to Learn

• •

  Do not chase generic “AI strategy” content

  You do not need broad executive-level theory right now. If it does not help you review a deal faster or more accurately this quarter, it is noise.

• •

  Do not spend months on deep neural network theory

  Underwriting value comes from document intelligence, workflow automation, and risk validation—not building new foundation models from scratch.

• •

  Do not overfocus on flashy no-code demos

  A pretty chatbot that cannot cite sources or handle confidential deal data is useless in investment banking. Build tools that fit real control requirements and survive scrutiny from compliance and senior bankers.

A realistic timeline is 8-12 weeks of focused learning:

• •Weeks 1-2: prompt extraction + LLM basics
• •Weeks 3-5: Python + pandas
• •Weeks 6-8: workflow automation + validation patterns
• •Weeks 9-12: one portfolio project tied directly to underwriting work

If you want relevance in 2026 as an underwriter in investment banking, aim for this combination: faster document processing, stronger validation instincts, and enough automation skill to remove repetitive work from your desk without handing over judgment to the model.

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.