Gains:
- Ability to clearly separate the assumption, driver and output layers of a financial model with AI
- Ability to set up optimistic/base/pessimistic scenarios and univariate sensitivity tables with AI support
- Ability to verify AI's model logic in terms of circular referencing, formula and assumption consistency
A financial model is a structured accounting framework that simulates the future of a company or project with numbers. At its simplest, it's a "put in a few assumptions, get results" setup: how do profits, cash, and value change when you change the growth rate, cost rate, and investment amount? A good model consists of three clear layers: assumptions (the numbers you input), drivers (the logic that connects those numbers to the output), and outputs (profit, cash, rates). AI speeds up the modeling process in many ways: it proposes the model skeleton, establishes formula logic, replicates scenarios, and prepares sensitivity tables. But the model's most dangerous errors are the invisible ones: an incorrect formula, a hidden circular reference, or an inconsistent assumption silently makes the model wrong. The essence of this unit is to check the model logic that the AI builds.
In this unit, we will learn to separate the three layers of the model, set up optimistic/baseline/pessimistic scenarios and sentiment table, and validate the model logic of AI.
Concepts: Assumption: The number you manually enter into the model (e.g. growth 10%). Driver: Formula logic that connects the assumption to the output. Scenario: Complete change of a set of assumptions (optimistic/base/pessimistic). Sensitivity analysis: Changing a single assumption step by step and seeing the effect on the output. Circular reference: Direct/indirect connection of a cell to itself.
Three Layers of the Model
The most important disciplinary rule of a model is to separate assumptions from calculation. Assumptions are collected in one place, the account is tied to them; so that when you change an assumption, the entire model is updated consistently. Enforce this distinction when having AI build models.
layer
Content
Good design rule
Assumptions
Growth, margin, maturity, discount, investment
One place, open label
Drivers
Formulas such as revenue = price × volume
Referring to the assumption, there is no fixed embedded number
Outputs
Profit, cash, rates, value
It is derived only from calculation, it is not written by hand.
Tip: Never embed a "fixed" number into the formula in the model (e.g. =Turnover*0.55). Instead, put 0.55 in the assumption cell and reference it. This way, when the assumption changes, the model is updated from a single point and you can easily capture the hidden constants that the AI has embedded.
Script and Sensibility
Scenario and sensitivity are different things and can be confused. The scenario changes multiple assumptions together (in the pessimistic world, both growth decreases and costs increase). Sensitivity isolates a single assumption and measures its impact on output (how the value changes if only the discount rate changes). Together they show where the model is fragile.
Model skeleton prompt: "Set up a simple 3-year revenue-profit model for a SaaS company. Separate the layers: (1) ASSUMPTIONS (starting customers 500, monthly growth 4%, monthly subscription 1,200 TL, gross margin 78%, monthly fixed expense 900 thousand TL), (2) DRIVERS (number of customers, revenue, expense formulas), (3) OUTPUTS (monthly revenue, gross profit, operating profit) Don't put constant numbers in formulas, just assume it all."
Scenario prompt: "Add three scenarios to this model. BASE: given assumptions. OPTIMISTIC: monthly growth 6%, margin 80%. PESIMIST: monthly growth 2%, margin 72%, fixed expense 10% more. Table the 36th month operating profit of the three scenarios side by side."
Sensitivity table prompt: "Show the sensitivity of the 36th month operating profit to the monthly growth rate. Change the growth from 1% to 7% in 1 point steps, keep other assumptions constant and record the profit at each step in the table. Summarize how sensitive the profit is to growth in one sentence."
Model audit prompt: "In the model you have built: (1) are there fixed numbers embedded in the formulas that are not based on assumptions? (2) is there any circular reference risk? (3) are the units consistent (monthly/annual, TL/thousand TL)? Report each risk you find by itemizing it."
Weak Prompt / Strong Prompt
WEAK: "Make me a 3-year financial model." (Result: a calculation with mixed layers, embedded constants, difficult to audit.) STRONG: "Build a 3-year model with separate assumption/driver/output layers, no fixed embedded in the formulas, growth and margin assumptions that can be changed in one place; then add a growth sensitivity table and check the risks of cyclical reference and unit inconsistency in the model."
Mini Cases
Case 1 — Cost of embedded constant. An analyst reduces margin from 78% to 75% in the AI-built model, but operating profit changes less than expected. When he runs the control prompt, he sees that the cost formula is based on the embedded 0.22 constant, not the assumption. Even if the margin cell changed, the formula used the old value. The model responds correctly when based on a fixed assumption.
Case 2 — Fragile value. In an investment decision model, the base scenario looks attractive. When the analyst builds a sensitivity table, he sees that operating profit is halved when growth decreases from 5% to 3%. So the whole appeal of the model rests on a single optimistic growth assumption. This awareness changes the decision from “definite yes” to “cautious before growth is confirmed.”
Case 3 — Cyclic trap. In a borrowing model, interest expense depends on the debt balance, the debt balance depends on the cash need, and the cash need depends on the interest. While installing AI, a circular reference occurs and Excel gives a warning. Instead of closing the warning and continuing, the analyst revises the logic and breaks the chain to calculate the interest on the previous period balance. Thus, the result is both accurate and traceable.
Caution: A model is never justified by a single scenario that looks good. The sentence "It is very profitable in the base scenario" cannot be a justification for a decision without knowing what happens in the pessimistic scenario. In each model, ask for the least optimistic/base/pessimistic triad and sensitivity to the critical assumption.
Four Questions of Model Checking
Ask each model the AI builds these four questions:
- Is there any embedded fixed? Is each number based on an assumption or embedded in the formula?
- Are the units consistent? Are monthly and annual, TL and thousand TL mixed?
- Is there a circular reference? Is a cell connecting to itself?
- Does the output react correctly to the assumption? When you change an assumption, does the output change in the expected direction?
Common mistakes
- Not separating assumption from calculation. When constants are embedded in the formula, the model cannot be controlled and reacts incorrectly.
- Deciding with a single scenario. A decision made without seeing a pessimistic situation is risky.
- Turn off the cyclic warning. A warning is a sign of a mistake, not an annoyance to be silenced.
- Skipping sensitivity. The robustness of the model cannot be measured without knowing which assumption the result depends on.
- Unit confusion. Monthly/annual and TL/thousand TL discrepancies create silent but large errors.
In summary
A financial model is an accounting framework that simulates the future with assumptions, and its power comes from the clarity of the layers. The AI framework quickly builds scenarios and sentiment tables; but it can introduce invisible errors such as embedded constant, circular reference and unit inconsistency. Separate the assumption from the calculation, add the optimist/base/pessimist triad and sensitivity to the critical assumption to each model, ask the four audit questions at a time. A single good scenario cannot justify a decision.
Application task
Have the AI build a simple 3-year revenue-profit model divided into three layers (assumption/driver/output) and ask for no constants embedded in the formulas. Add three optimistic/baseline/pessimistic scenarios and put the latest operating profit side by side. Then ask for a sensitivity table for the most critical assumption (e.g. growth). Finally, run the model check prompt and check for embedded constants, cyclic references, and unit inconsistencies and note the findings.
checklist
- [ ] I separated the assumption, driver and output layers.
- [ ] I verified that there are no constants embedded in the formulas.
- [ ] I set up at least three optimistic/some/pessimistic scenarios.
- [ ] I added a sensitivity table for the critical assumption.
- [ ] I checked for circular reference.
- [ ] I verified unit consistency (monthly/yearly, TL/thousand TL).
- [ ] I changed an assumption and tested that the output responded correctly.