Kumu / Finance & Securities / Stage 3
Stage 3 — AI-Assisted Build + Audit
models/builds/…-model.xlsx + analysis/…-build-audit.md
Generate a working workbook from your own Stage 2 spec — any AI tool, or by hand — then audit the result ruthlessly. Your graded skill isn't typing formulas; it's specifying precisely and auditing skeptically — the two things an analyst still owns when AI does the assembly.
What you'll learn
- How to drive an AI build from a spec — and why re-explaining mid-chat means your spec has a defect
- What a professional audit of AI output looks like: check, find, fix, document
- Auditable model construction: named ranges, formulas-only, color coding, visible checks
There is no starter template
Your spec is the template. The instructor workbook exists — as the grading key.
The build contract
The finished workbook must satisfy all seven, verifiable by inspection — and by the grading script — regardless of how it was produced:
| # | Requirement |
|---|---|
| 1 | All ten named ranges from the contract, each attached to the right cell |
| 2 | Formulas, never hard-coded values. Every calculated cell is a formula referencing named ranges — a pasted number where a formula belongs scores zero for that element (checked mechanically) |
| 3 | Cover page — scenario, author, date, data-provenance block (placeholders noted as indicative) |
| 4 | Legend/Key tab — Yellow = inputs · Blue = assumptions · Green = formulas · Gray = outputs, applied throughout |
| 5 | All three hedge families — forward; money market in its three explicit steps; put and call with premium cost in USD and proceeds as a function of S_T |
| 6 | Sensitivity table + chart — ±5% in 1% steps, formula-driven (no hand-typed rows) |
| 7 | Validation checks live in the workbook — the parity check and your spec's §7 check figures, computed, visible, passing |
Run the build — step by step
Tool guidance — the contract is graded, not the tool
Required reading before your first build session
The AI Tools Lab is the field guide this stage assumes: chat vs "code" tools, web vs desktop vs CLI, and how to give the AI your repo so it checks your conventions instead of guessing them. Ten minutes there decides whether the build is one clean session or an afternoon of flailing.
- Claude for Excel — if you have access; availability varies and don't buy anything
- Claude / ChatGPT on the web — paste or link your spec (GitHub URL is cleanest); ask for the workbook, or for the structure + formulas to assemble yourself
- Copilot in Excel
- Manual build directly from your spec — always allowed
Your spec goes in as-is
If you find yourself re-explaining the model in the chat, that's a spec defect — fix the
spec (commit the change), then regenerate. Log every prompt in prompt-log.md.
Feeding context: best is a GitHub link to your committed spec; medium is uploading the file; last resort is copy-paste. Version-controlled context is the professional pattern — reproducible and auditable.
The audit note — required, ≥3 findings
AI output is a draft, not a deliverable. Audit the generated workbook against your spec's validation rules and document at least 3 findings — things you checked and confirmed, or found broken and fixed. Real examples:
- A hardcoded value where a formula belongs
- A named range attached to the wrong cell
- A sign error in the put payoff
- A sensitivity row that doesn't recalculate when inputs change
- The parity check failing because of a rate-basis mismatch (ACT/360 vs ACT/365)
For each finding: what you checked, what you found, what you did.
"Everything was perfect" is a red flag
Not a good sign — the grader will be auditing the same workbook. An audit that found nothing usually means an audit that looked at nothing.
Submission checklist
- Workbook generated from the committed spec (prompt logged) or built by hand from it
- All ten named ranges present and attached to the right cells
- Every calculated cell is a formula referencing named ranges — no pasted results
- Cover page + Legend/Key tab; color convention applied throughout
- MM hedge in three explicit steps; put and call payoffs vary with
S_T - Sensitivity table ±5% in 1% steps, formula-driven, with chart
- Parity check and spec §7 checks computed in the workbook and passing
- Audit note with ≥3 substantive findings; fixes committed individually
- Workbook at
models/builds/YYYY-MM-DD-{lastname}-{scenario-slug}-model.xlsx; audit atanalysis/YYYY-MM-DD-{lastname}-build-audit.md prompt-log.mdupdated
Rubric
| Criterion | Weight | Strong work looks like |
|---|---|---|
| Contract compliance | 50% | Named ranges complete and correct; formulas-only (mechanically checked); all hedges + sensitivity present and computing correctly |
| Structure & presentation | 25% | Cover, legend/key, color convention, auditable layout |
| Audit note | 25% | ≥3 substantive findings with evidence; fixes committed |