# AI in Finance — Templates & Checklists

Working templates from the *Agentic AI in Finance* track at Agentic Academy (agentic-academy.dev).
**Educational use only — not financial, investment, tax, or legal advice.**

> These pair with the track's lessons. The point is not to let AI do the work unchecked — it's to go faster *while staying honest about what's verified*.

---

## 1. The VERIFY Checklist (use on every AI output before it leaves your hands)

Treat every AI output as a draft until each box is ticked:

- [ ] **V — Values recomputed.** Every derived number (ratios, totals, multiples) recalculated by hand or in the model, not taken from the AI.
- [ ] **E — Every citation resolves.** Each source named by the AI actually exists and says what's claimed. Unfindable source → delete the claim.
- [ ] **R — Reconciled to primary source.** Headline figures traced back to the filing / transcript / data provider, not the AI's paraphrase.
- [ ] **I — Internally consistent.** Numbers agree with each other (e.g., EV / EBITDA matches the stated multiple); the conclusion follows from the data.
- [ ] **F — Flagged what's unverified.** Anything you couldn't confirm is explicitly marked UNVERIFIED, not buried.
- [ ] **Y — Yours to own.** A qualified human (you) signs off and is accountable for the judgment — the AI is not.

---

## 2. Per-Claim Verification Log (attach to any AI-assisted deliverable)

| # | Claim / figure | Source cited | Status | Notes |
|---|----------------|--------------|--------|-------|
| 1 | | | Verified / Unverified / Could-not-confirm | |
| 2 | | | | |
| 3 | | | | |

*Rule of thumb: if the recommendation only holds up on rows marked "Unverified," it isn't ready.*

---

## 3. DCF Audit Prompt (paste into Claude / ChatGPT alongside the model)

```
Audit this DCF for me. Show your work as a labelled list:
1) Rebuild the WACC from its inputs (cost of equity via CAPM, cost of
   debt, weights, tax rate) so I can check it.
2) State terminal value as a % of enterprise value, and re-derive it with
   BOTH perpetuity-growth and exit-multiple methods; compare.
3) Trace the cash-flow build: revenue -> EBIT -> unlevered FCF, listing
   each key assumption (growth, margin, working capital, capex).
4) List every HARD-CODED number that should be a live formula.
5) Produce a WACC x terminal-growth sensitivity table.
Do not change the model — just report. Flag anything you cannot verify
from the inputs provided.
```

---

## 4. Investment Committee (IC) Memo Skeleton

```
RECOMMENDATION: [Proceed / Pass / Proceed with conditions]
ONE-LINE THESIS: [...]

1. The opportunity (2-3 sentences)
2. Key financials — each figure tagged [verified] or [unverified]
3. Quality of earnings — adjustments and what drives them
4. Risks (ranked) and mitigants
5. Valuation — method, key assumptions, sensitivity range (not a single number)
6. What we could NOT verify in the time available
7. Conditions / next diligence steps

DISCIPLINE: the thesis must survive when every [unverified] line is removed.
```

---

## 5. "Spot the Hallucination" self-test

Before trusting any AI valuation summary, ask:
- Does the arithmetic agree with itself? (Recompute the multiple from EV and EBITDA.)
- Is every cited report real and findable? (Watch for plausible but fake sources.)
- Does the conclusion actually follow from the numbers, or contradict them?
- Are projected/scenario figures clearly separated from cited facts?

---

## 6. Data-Handling Rules (non-negotiable)

- [ ] No MNPI or client data in consumer chatbots — approved enterprise tools only.
- [ ] Prompts and outputs for supervised/client-facing work are retained (FINRA/SEC recordkeeping applies to AI too).
- [ ] Least-privilege access for any connected tool or MCP connector.
- [ ] A human owns every decision; AI assists, it does not approve.

---

*© Agentic Academy — share freely with attribution. Regulations and tool capabilities change; verify current rules and data before relying on them.*