Study guide: AI and Corporate Finance

📋 Course Outline

1. Simulation Structure and Assessment
2. Kolumpi Business and Operations
3. Year Zero Diagnosis and Indicators
4. Variable-Cost Accounting
5. Forecasting and Profitability Thresholds
6. AI Finance Course Framework
7. Professional AI Risks and Verification
8. Professional AI Fluency
9. Prompting and Critical Review
10. Funding Strategy and Metrics
11. Financial Triangle and Assumptions
12. Building the Financial Model
13. Statements, Controls, and Decisions
14. Adversarial Audit and Sensitivity
15. Adversarial Audit of the Model
16. Sensitivity and Contradictory Sourcing
17. Investment Committee Preparation
18. Committee Presentation and Decision

📖 1. Simulation Structure and Assessment

★ Must-know

• The Kolumpi simulation runs across five sessions and six monthly decisions covering January to June of year 1.

• The assessment consists of 10% deadlines and participation, 35% six team decision reports, 40% the final synthesis and presentation, and 15% individual quizzes.

📌 The red AI regime prohibits AI during diagnostics, calculations, quizzes, decisions, and presentations; the orange regime allows conceptual tutoring without formulas or figures; and the green regime allows documented strategic criticism and report auditing outside class.

Further detail

📌 Every ratio in a submitted report must include its calculation components, otherwise it does not count.

💡 Memory Hook

Inform → Reflect → Decide → Simulate

📖 2. Kolumpi Business and Operations

★ Must-know

• The six products are:

  • Oliplas
  • Lavplas
  • Duoplas
  • Olibox
  • Lavbox
  • Duobox

• Production moves from olive or lavender soap base and packaging through the cutting workshop and conditioning workshop to finished products distributed through the website, independent or franchised perfumeries, marketplaces, and hypermarkets.

• Each machine provides 9,100 minutes per month, and the target workshop load is approximately 100% of theoretical consumption divided by workshop capacity.

📌 Raw-material stock is measured in days of production, whereas finished-product stock is measured in days of sales.

Further detail

📌 A standard single soap consumes 0.15 kg of soap base and its packaging, while a Duo consumes 0.15 kg of each fragrance.

💡 Memory Hook

A soap bar moves from raw materials through two workshops into four distribution channels.

📖 3. Year Zero Diagnosis and Indicators

📝 Essential Points

• The eight steering indicators are:

  • profitability
  • cash in days of revenue
  • raw-material stocks
  • finished-product stocks
  • workshops
  • human resources
  • customer satisfaction
  • innovation

• The cash-in-days-of-revenue indicator is maximal at 7 days or less, average at 20 days, and zero at 45 days or more; overdrafts incur agios at 100% per year.

• Ideal raw-material stock is 30 to 60 production days, while a finished-product stockout is penalized and finished-product overstock mainly consumes cash.

📐 Formula - The seasonality coefficient satisfies $\text{coefficient}=\frac{\text{observed sales}}{\text{trend sales}}$, so corrected sales satisfy $\text{corrected sales}=\frac{\text{observed sales}}{\text{coefficient}}$.

💡 Memory Hook

Seasonality → distorted comparisons → corrected forecasts and financing needs

📖 4. Variable-Cost Accounting

★ Must-know

📐 Formula - The analytical accounting cascade is $\text{revenue}-\text{variable costs}=\text{variable-cost margin}$, then variable-cost margin minus product-specific fixed costs equals product contribution margin, and contribution margins minus common fixed costs equal operating income.

📌 The variable-cost margin rate measures price sensitivity, whereas the contribution margin in euros identifies which product contributes most to covering fixed costs.

Further detail

• The variable costs include:

  • materials
  • machine minutes
  • discounts
  • 3% of the catalogue price

• Common fixed costs, including salaries, depreciation, and overheads, amount to €287,500 per month in year 0.

💡 Memory Hook

Revenue → variable costs → contribution margin → fixed costs → operating result

📖 5. Forecasting and Profitability Thresholds

★ Must-know

📐 Formula - For an action that adds fixed costs, the minimum additional revenue is $\Delta CA=\frac{\Delta CF}{m}$, where $m$ is the variable-cost margin rate.

📐 Formula - For a single product, the minimum additional quantity is $\Delta Q=\frac{\Delta CF}{\text{unit variable-cost margin}}$.

• For Duobox, reducing the price to €7.20 with a 10% decrease lowers unit variable-cost margin from €5.125 to €4.429 and requires 28,929 units, a 15.7% increase, to preserve the former total margin.

Further detail

• A €8,000 monthly advertising cost requires at least €13,333 of additional revenue to be covered.

💡 Memory Hook

Observed sales are seasonal, whereas corrected sales reveal the underlying trend.

📖 6. AI Finance Course Framework

🔑 Key Concepts & Definitions

• Large language model: trained on billions of texts to predict the most plausible continuation of a text rather than to determine what is true

📝 Essential Points

• The four dimensions are (Anthropic):
  • Delegation
  • Description
  • Discernment
  • Diligence

📌 Delegation means deciding what to entrust to AI and what to retain, including retaining hypothesis choices, decisions, and fieldwork.

• Professional AI use requires describing the task, evaluating the output, iterating or rejecting it, and verifying, declaring, and assuming responsibility for the final use.

💡 Memory Hook

4D: Delegation, Description, Discernment, Diligence

📖 7. Professional AI Risks and Verification

★ Must-know

• AI hallucination risk is especially high for precise figures, sources, citations, and URLs, so an unsupported figure must be removed from the dossier.

📌 Sycophancy makes an AI validate the user's assumptions, whereas an adversarial prompt asks it to identify and rank objections.

📌 No real personal data may be entered into AI tools, and interview quotations must be anonymized before use.

Further detail

• Deloitte Australia reported in 2025 that a 440,000 AUD government report contained invented academic references and a fabricated judgment citation, and Deloitte reimbursed approximately 290,000 AUD.

• The course links responsible AI use to the GDPR and the European AI Act, with traceability becoming a legal expectation for higher-risk uses.

💡 Memory Hook

Plausible output without verification → hallucination, poor work, and bad decisions

📖 8. Professional AI Fluency

★ Must-know

📌 Professional AI use delegates formatting, ideas, and criticism to the AI while keeping hypotheses, decisions, and fieldwork with the human.

📌 The course requires that no real personal data be entered into an AI and that interview verbatims be anonymized beforehand.

• The framework has four dimensions:
  • Delegation
  • Description
  • Discernment
  • Diligence

Further detail

• The course links responsible AI use to the European GDPR and AI Act, including traceability of AI use in high-risk applications.

💡 Memory Hook

The 4D loop: Delegation, Description, Discernment, Diligence

📖 9. Prompting and Critical Review

★ Must-know

• A reliable AI workflow consists of describing the task and documents, examining the output as potentially false, discerning its accuracy, relevance, reasoning, and form, then re-describing the request for two to four further rounds.

📌 Effective prompts specify the deliverable, context and purpose, examples, reasoning process, structure, output format, constraints, and a request for self-criticism.

📌 A prompt should instruct the AI to state when it lacks knowledge or a reliable source instead of inventing an answer.

Further detail

• Short targeted iterations, persistent workspaces containing the team documents, and context engineering have replaced formulaic prompt tricks as the useful practices.

💡 Memory Hook

Describe → generate → discern → re-describe

📖 10. Funding Strategy and Metrics

★ Must-know

• The three paths are:

  • Raise investment
  • Build to sell
  • Remain profitable

• Core investor metrics include burn, runway, gross margin, CAC, LTV, LTV/CAC, and CAC payback, with a target LTV/CAC greater than 3.

📐 Formula - Runway is calculated as $\text{cash balance} \div \text{burn rate}$, CAC as $\text{sales and marketing spending} \div \text{new customers}$, and LTV/CAC as $\text{customer value} \div \text{customer acquisition cost}$.

Further detail

• The metrics vary by category:

  • SaaS: MRR, NRR, ACV, pipeline
  • Marketplace: GMV, take-rate, contribution margin, concentration
  • B2B fintech: volumes, fee structure, gross margin, pipeline
  • Apps and B2C: growth, organic acquisition, DAU/MAU, retention

• The indicative funding levels are:

  • Personal contributions and love money: 1,000–30,000 euros
  • Pépite or student grants: 1,000–5,000 euros
  • Public subsidies: 5,000–30,000 euros or more
  • Bank and honor loans: 10,000–100,000 euros
  • Business angels: 50,000–500,000 euros
  • Seed funds and venture capital: 300,000–3 million euros or more

💡 Memory Hook

Raise, resell, or bootstrap: growth, an identifiable asset, or rapid profitability

📖 11. Financial Triangle and Assumptions

★ Must-know

📌 The balance sheet shows what the company owns and owes at a date, the income statement shows products minus charges over a period, and cash flow shows money actually entering and leaving the bank account.

📌 The financial triangle requires every decision to be analyzed through the balance sheet, income statement, and cash flow statement.

• Amortization is an income-statement charge and reduces the balance-sheet value of an asset, but it is never a cash movement after the original purchase.

Further detail

• For a 700-euro loan repayment containing 80 euros of interest and 620 euros of principal, only the 80 euros of interest is an income-statement charge, while the 620 euros reduces the balance-sheet debt.

💡 Memory Hook

The balance sheet is a photo, the income statement a film, and cash flow the bank account

📖 12. Building the Financial Model

★ Must-know

• The seven blocks are:
  • Revenue by category
  • Variable costs
  • Fixed costs
  • Payroll
  • Investments and amortization
  • Capital and subsidies
  • Borrowing

📐 Formula - Each revenue category is derived as $\text{volume} \times \text{price} = \text{revenue}$ from explicit drivers such as actions, audience, conversion, quantities, price, and seasonality.

• Variable costs are expressed per unit multiplied by quantities sold, while purchases may affect cash when bought and inventory when not yet sold.

📐 Formula - For a loan, monthly principal repayment is $\text{loan amount} \div \text{duration}$, monthly interest is $\text{annual rate} \times \text{remaining principal} \div 12$, and the monthly payment is principal repayment plus interest.

Further detail

• The model must include a founder salary no later than month 12 and must justify each revenue amount through a one-line step plan showing the required sales activity.

💡 Memory Hook

Revenue drivers → costs → salaries → investments → funding → debt

📖 13. Statements, Controls, and Decisions

★ Must-know

📐 Formula - The income statement cascade is $\text{revenue} - \text{variable costs} = \text{gross margin}$, followed by fixed costs, payroll, amortization, financial charges, and corporate tax to obtain net income.

📌 Corporate tax is 25% of the year's cumulative positive pre-tax income, recorded as an expense in year N, shown as a liability on the year-N balance sheet, and paid in March of year N+1.

• The balance sheet must balance through formulas using assets such as net fixed assets, inventory, receivables, and December cash, and liabilities such as capital, cumulative result, remaining debt, and unpaid tax.

Further detail

📌 If cumulative pre-tax income is negative, corporate tax is zero in the course model and no deficit carryforward is used.

📌 The four spreadsheet rules are explicit titled units, no hidden hard-coded numbers in formulas, no merged cells, and no duplicate manual entry of an existing value.

💡 Memory Hook

Bad assumptions distort statements; linked formulas and controls expose the errors

📖 14. Adversarial Audit and Sensitivity

★ Must-know

• An adversarial AI audit asks the AI to find ten errors and classify them by severity, after which the team verifies every accusation and records each verdict and justification.

📌 An audit finding can be founded, unfounded, or still to be verified, as illustrated by a genuine formula-column offset, an incorrect claim about monthly tax payment, and an unsupported conversion-rate criticism.

• Sensitivity analysis varies four or five key variables by plus or minus 20% and examines the resulting cumulative net income and lowest cash balance.

📌 Field evidence should be dated and named because AI produces plausible text, whereas fieldwork produces facts from reality.

Further detail

• The most important sensitivity result is the interpretation of which variable damages the model first and at what threshold it becomes critical.

💡 Memory Hook

The AI may identify a real error, a cosmetic issue, or a false accusation

📖 15. Adversarial Audit of the Model

★ Must-know

📌 After the model is built manually, the AI becomes an adversarial auditor whose accusations must be classified as founded, unfounded, or requiring verification.

• The adverse audit consists of submitting the model with a request to find 10 errors, classifying each claim by severity, verifying it independently, and recording the verdict and justification in the final file.

📌 A false accusation exposed by the team is as valuable as a real error corrected because resisting an AI that sounds certain is a rare competence.

Further detail

• The course gives three audit outcomes: the rule that corporate income tax for year N is paid in N+1 is unfounded when the AI assumes monthly payment, a one-column reference shift from March 2028 is founded, and a 35% conversion rate is to be verified through a sector benchmark and field evidence.

💡 Memory Hook

The AI accuses; the team verifies and judges.

📖 16. Sensitivity and Contradictory Sourcing

★ Must-know

• Sensitivity analysis varies each of four or five key variables by plus or minus 20% and measures the resulting cumulative net income and lowest cash position.

📌 Sensitivity analysis turns a business plan from a prediction into a map of bets by identifying which variable kills the model first and at what threshold.

• Contradictory sourcing requires taking the five most important market figures, opening the primary source cited by the AI, recording the real source and discrepancy, and removing or relabelling an untraceable figure as a hypothesis to test.

📌 A chatbot answers questions, whereas an agent acts by chaining steps, using tools, and pursuing an objective within a defined scope and audited deliverable.

Further detail

• If the model dies when conversion falls from 10% to 7%, conversion becomes the priority field test before any money is spent.

💡 Memory Hook

A variable moves → the model reveals which field must be tested.

📖 17. Investment Committee Preparation

📝 Essential Points

• The investment committee preparation consists of an eight-minute investor-format pitch, a public cross-team audit draw, five written and ranked questions with a one-page investment opinion, and an individual three-minute oral examination.

• The Livrable 3 dossier is due the day before the committee at 17:00 and must contain an Excel workbook with seven tabs, a memo of no more than four pages, an AI logbook, four field proofs, and usage declarations.

📌 The pitch must connect the problem, solution, sourced market, field evidence and traction, team, key model figures including cash, and the requested path of the money.

💡 Memory Hook

Pitch → audit questions → individual oral → final dossier.

📖 18. Committee Presentation and Decision

★ Must-know

• The committee session includes installation and voting rules, a 10-minute pitch, questions from the audit team, an individual oral examination, deliberation and voting, and a final debrief.

• The handwritten opinion contains:

  • An investment decision
  • Three precise strengths
  • Three risks ordered by severity
  • Two conditions before signing
  • The best and weakest answers to difficult questions

📌 A market risk means that nobody buys, whereas an execution risk means that the team cannot deliver the project.

• The workflow contains:
  • Divergent ideation
  • Hostile investor criticism
  • Socratic Excel tutoring
  • Adverse model auditing
  • Contradictory sourcing
  • Meta-prompt improvement

📌 The Socratic Excel tutor must never provide a complete formula, ask what the student has tried, explain concepts through different examples, describe symptoms instead of corrections, refuse to do the work, and end every response with a question.

📌 The sourcing prompt requires primary sources, competing figures, explanations of why the figure could vary from single to double, a plausible interval instead of a point estimate, and the explicit label “not found” when no solid source exists.

• The rules require:

  • One account per student
  • Prompts drafted before submission
  • Changing tools when a limit is reached
  • Never using real personal data
  • Changing training settings immediately

• The declaration includes:

  • Tools and dates
  • AI usages
  • Rejected or removed outputs and reasons
  • Work completed without AI and why
  • A final responsibility statement

Further detail

• Each team receives 100 k€ of fictional capital to allocate across no more than three projects other than its own, creating a class allocation market.

📌 The adverse audit prompt instructs the AI to report structural errors, reasoning errors, and fragilities, estimate their cash impact, explain how the team can verify them in under five minutes, and state what it could not inspect.

• The recommended tools are:
  • Claude
  • Gemini
  • NotebookLM
  • Le Chat
  • ChatGPT

💡 Memory Hook

Founders during the pitch; investors during the rest of the session.

📊 Synthesis Tables

Kolumpi AI regimes

The financial triangle