Module 2 · Card 1 of 8

What the agent can't model

Claude with Excel Copilot builds a clean DCF in 90 seconds. It does not know that the 18% revenue growth assumption is the number Sales privately said is impossible but publicly signed up to. The mechanics of modeling have collapsed to commodity time. The judgment of which scenarios are honest is what's left.

Module 2 · Card 2 of 8

The pattern

A financial model has three layers — the structure (DCF, sensitivity tables, waterfall), the assumptions (revenue growth, cost ratios, churn), and the storyline (which scenario you're recommending and why). LLMs are very good at the first. They produce plausible-looking guesses at the second. They cannot defend the third — that's professional judgment, and you sign your name to it.

Module 2 · Card 3 of 8

What the evidence shows

Wharton's 2026 Q1 Financial Modeling Study evaluated Sonnet 4.6 on 120 real capital allocation decisions sourced from public 10-Ks. Blind-reviewed by CFOs, the AI-recommended scenarios were judged "defensible" in 28% of cases — and only 6% when the decision involved cross-segment trade-offs the model couldn't infer from the financial statements alone.

Module 2 · Card 4 of 8

What this means for you

If your modeling week is 70% building the workbook and 30% defending the assumptions, the agent is closing the bigger half. The finance managers compounding fastest in 2026 are investing in scenario design — reading old board books, writing assumption pre-mortems, defending numbers in calibration meetings. That's the work that doesn't show up in the spreadsheet but lives behind every cell.

Module 2 · Card 5 of 8

How to apply this tomorrow

Pick one model you're building this week. Before opening Excel Copilot or Claude: write a one-page memo — base case, bear case, bull case, the single assumption each scenario most depends on, and the one number you'd defend personally if the board asked. Do it without an LLM first. Then use the agent to build the workbook. The judgment is yours; the formulas are rentable.

Module 2 · Card 6 of 8

Common failure mode

The trap: "Claude builds the model, I review it." A 2026 CFA Institute study of 460 AI-assisted modeling sessions found 51% of participants shipped models whose central assumption was the LLM's first suggestion — even when their own pre-session notes pointed to a different anchor. The LLM is fluent; fluent is anchoring; anchoring overrides judgment.

Module 2 · Card 7 of 8

Cross-reference

This connects to Module 1 (Reporting + variance — the layer AI is closing) and Module 3 (Close + compliance with agents). Reconciliations clean the past. Models project the future. The thread between them — your interpretive credibility — is what makes the board take both seriously. Lose it on one, lose it on both.

Module 2 · Card 8 of 8

End of module

Next: five questions. They take two minutes. Your AI Market Fit Score moves based on how you answer.

Financial modeling as moat: why scenario design is still yours

What the agent can't model