AI Costing Prompts · By task
Build a driver-based forecast with scenarios
Give an AI model your operating drivers and get a clean four-quarter projection plus three Q4 scenarios that vary churn and new customer inflow. The prompt builds the forecast from drivers, not from a guessed top line, and tells you which lever actually moves the number.
In short
Feed the model your starting customer base, average revenue per customer, churn rate, new customers per quarter and variable cost percentage. It rolls the customer count forward quarter by quarter, converts that into revenue and contribution, then runs a conservative, base and aggressive scenario for Q4 by changing only the churn and new customer assumptions. The output shows you the spread between scenarios and identifies the more powerful driver over the range you tested.
What the prompt is doing
This is driver-based planning, the discipline at the heart of good profit-driven budgeting. Instead of starting with last year's revenue and adding a percentage, you build the forecast from the operational mechanics that actually generate it: how many customers you keep, how many you add, what each is worth, and what proportion of revenue survives as contribution. Every line in the forecast traces back to a driver you can argue about and change.
The scenarios are where this becomes a decision tool rather than a spreadsheet. By moving churn and new customer inflow independently, you see which lever your forecast is most sensitive to. That tells you where to focus management attention and how to read the risk in your own plan. A forecast that swings wildly on churn needs a retention answer before it needs an ambitious sales target. See profitability forecasting for how this connects to margin, not just revenue.
The prompt
You are an FP&A analyst who works driver-based. Work only from the data and drivers I give you. Do not invent any numbers. If a driver is missing, ask for it or flag it rather than guessing. Drivers: Starting active customers: 400 Average revenue per customer per quarter: 1,500 EUR Quarterly gross churn: 6% New customers per quarter: 50 Variable cost: 55% of revenue Do the following: 1. Project active customers at the end of each of the four quarters (start - churn + new). 2. Calculate revenue per quarter (average customers in the quarter x revenue per customer). 3. Calculate contribution per quarter (revenue x (1 - variable cost %)). 4. Build three Q4 scenarios, changing only churn and new customers: - Conservative: churn 8%, new 35 - Base: as given above - Aggressive: churn 4%, new 70 Show Q4 revenue for each. 5. State which driver moves the forecast most over these ranges. 6. List your assumptions.
A worked example
| Quarter | End customers | Revenue (EUR) | Contribution at 45% (EUR) |
|---|---|---|---|
| Q1 | 426 | 619,500 | 278,775 |
| Q2 | 450 | 657,330 | 295,799 |
| Q3 | 473 | 692,820 | 311,769 |
| Q4 | 495 | 726,225 | 326,801 |
Full-year revenue is approximately 2,695,875 EUR. Customers are projected by applying churn to the opening base each quarter and adding 50 new.
| Q4 scenario | Churn | New | Q4 revenue (EUR) |
|---|---|---|---|
| Conservative | 8% | 35 | 707,925 |
| Base | 6% | 50 | 726,225 |
| Aggressive | 4% | 70 | 748,425 |
The spread between conservative and aggressive Q4 revenue is roughly 40,500 EUR.
Over these ranges, new customers added is the more powerful driver, moving Q4 revenue around 1.8 times more than churn does. That is because the new inflow currently exceeds the number being churned, so each extra new customer has more weight than each point of retention. The model correctly notes the caveat: churn compounds over a longer horizon, so over two or three years the retention lever can overtake acquisition.
Assumption flagged: "average customers" is taken as the mean of opening and closing balances. The revenue basis changes everything; if you measure on end-of-quarter customers instead, every revenue figure rises.
What it costs you to run
The driver set above is roughly 300 tokens. A full run with the projection, the three scenarios and the commentary is a fraction of a cent on any current model. Re-running it with different driver values to explore your own sensitivities costs almost nothing, which is the point: scenario work should be cheap enough to do often.
Token figures are approximate and vary by model and tokeniser.
The guardrail that matters
Work only from the data and drivers I give you. Do not invent any numbers. If a driver is missing, ask for it or flag it rather than guessing.
Forecasting is where models are most tempted to be helpful by filling gaps. The instruction to ask for a missing driver rather than assume one keeps the forecast honest and keeps you in control of the assumptions. See how to stop AI inventing numbers.
When you need the real model
A driver-based forecast is only as good as the drivers behind it, and the hardest driver to get right is the true margin on each customer once cost to serve is loaded in. When the forecast needs to hold up in front of a board, the numbers should come from a model that traces cost properly rather than a flat variable percentage.