Building a simple financial model

0:00 Building a simple financial model

An investor asks you a simple question. If your customer acquisition cost goes up by half, when do you run out of money? If you have to say, I'll get back to you, you don't have a model. You have a spreadsheet of hopes. In this lecture, you'll learn to build a driver-based financial model, the kind where every number traces back to an assumption you can explain and test. You'll build six months of forecasts, add scenarios, include AI costs properly, and use AI assistants to check your work without handing over control.

0:41 Why it matters

Why does this matter? A model isn't about predicting the future perfectly. Nobody can. It's about thinking clearly. It shows which assumptions matter most, how long your money lasts, and what has to be true for the plan to work. It helps you decide when to hire, how much to spend on marketing and what price you need. And it makes conversations with partners, lenders and investors far easier. Here's the key idea. Every number in your forecast should trace back to a driver you can explain and test.

1:20 Recipe, not photograph

Here's an analogy. A good model is like a recipe, not a photograph of a cake. A photograph shows the result, but you can't change anything. A recipe shows the ingredients and the method, so if you want a bigger cake or less sugar, you change the ingredients and the result follows. Top-down forecasts, like we'll capture one percent of a huge market, are photographs. Driver-based models are recipes. Marketing spend divided by CAC gives new customers. Last month's customers, minus churn, plus new customers, gives this month's customers. Customers times price gives revenue. Change an ingredient, and everything downstream updates.

2:04 Three tabs

Now the structure. Three parts. First, an Inputs tab: one assumption per row, with a label, a unit and a source, like CAC two thousand rupees, from a three-week ad test. Second, a Calcs tab with months across the columns and the formulas: new customers, customers at the end of the month, revenue, variable costs, contribution, marketing, fixed costs and operating result. Third, an Outputs tab: a monthly summary, cumulative cash need, key metrics and a chart. The golden rule is never to type a raw number inside a formula. Always point to an input. That's what makes the model a recipe you can adjust.

2:50 Worked example 1: six months (illustrative, PKR)

A simple worked example, illustrative, in Pakistani rupees. Marketing spend is two hundred thousand a month and CAC is two thousand, so a hundred new customers a month. Churn is four percent. Price is three thousand a month, and variable cost six hundred, so contribution is two thousand four hundred per customer. Fixed costs are one point one million. Month one: a hundred customers, three hundred thousand revenue, a loss of about one point zero six million. By month six, with about five hundred and forty customers, the business roughly breaks even. Add up the losses before then and you get about three point one million. That's the minimum funding needed, before a buffer.

3:40 Scenarios

Next, scenarios, because the base case is never what happens. Build at least three. A base case with your most likely assumptions. A downside case with higher CAC, higher churn and slower sales. And an upside case. In our example, if CAC rises from two thousand to three thousand, new customers fall to about sixty-seven a month, and break-even moves much later. That downside case is what decides how much buffer you need. A simple way to build this is a scenario switch: a cell where you choose base, downside or upside, and the Inputs tab picks the matching column of assumptions.

4:24 AI costs as drivers

Now, if your product uses AI models, add AI costs as drivers, not as a fixed software line. The drivers are tasks per customer per month, cost per task, from the unit economics lesson, and human review minutes per task. AI cost of goods each month is customers times tasks per customer times the cost per task plus review. It flows straight into contribution. Then add two AI-specific scenarios. One where cost per task rises, because you need a better model for quality. And one where usage per customer doubles, because a heavy-user segment grows faster than you expected.

5:07 Worked example 2: AI bookkeeping (illustrative)

Now a realistic scenario, illustrative. Sara runs an AI-assisted bookkeeping service in Manchester. Her first model had one line for software at two hundred pounds a month. When she rebuilt it with drivers, she found each client generated about four hundred AI tasks a month, and each task needed about half a minute of human review. At her team's loaded cost, the review time cost far more than the model usage. Her margin wasn't driven by technology at all. It was driven by review time. So she targeted her improvement work at reducing review, with better extraction and exception flags, instead of chasing a cheaper model.

5:53 Watch me: core formulas

Watch me build the core formulas. In the Calcs tab, column C is month two. New customers equals marketing spend divided by CAC, both pointing to the Inputs tab with fixed references. Customers at the end equals last month's customers times one minus churn, plus this month's new customers. Revenue equals customers times price. Variable costs equal customers times variable cost. Contribution is revenue minus variable costs. Operating result is contribution minus marketing minus fixed costs. And a cumulative row adds each month to the last. Then I add a check row that turns red if customers ever go negative. The exact formulas are in the lesson text.

6:40 AI as a model checker

Now, can an AI assistant help? Yes, with care. Spreadsheet assistants in tools like Google Sheets and Excel, and general assistants like Claude or ChatGPT, can draft formulas, explain someone else's model and spot inconsistencies. But they can also produce formulas that look right and are subtly wrong, like referencing the wrong row. So ask the assistant to explain every formula in plain English and flag anything that mixes units or references the wrong row. Then check two rows by hand. And don't paste confidential financials into a tool whose terms allow training on your data. AI is good at syntax. It doesn't know your business.

7:26 Common mistakes

Let's run through common mistakes. Top-down revenue instead of drivers. Forgetting taxes, payment fees, refunds and hiring lags. Assuming new hires are productive from day one. Ignoring churn. Having no downside scenario. Hard-coding numbers inside formulas. Treating AI costs as a fixed software line. And trusting AI-drafted formulas without checking them. Finally, compare your forecast with actuals every month. The gaps are where your learning is. If CAC is always higher than you assumed, your model is telling you something important about your channels.

8:03 Recap and try this now

Let's recap. A good model is a recipe: inputs with sources, calculations built from drivers, and clear outputs. Build base, downside and upside scenarios, and let the downside set your buffer. If you use AI models, treat usage and review time as drivers of cost of goods. Use AI assistants to explain and check formulas, then verify by hand. Your try this now: build the three-tab, six-month model from the lesson for your idea, add a scenario switch, and find the month you break even in the downside case. Next, we'll turn that into a cash plan: runway and break-even.