What-if analysis: comparing manual vs AI approaches
Anyone who has spent a few quarters in finance, ops, or strategy has run a manual what-if. You open the model, you change a few inputs, you watch the bottom-line cell move, you write the new number down. It's fine. It's also slow and quietly misleading in ways most operators don't catch.
A what-if AI analysis tries to fix the same problem from a different direction. Same inputs, same outputs, but the work happens inside the engine and the answers come back in a different shape. This is a working comparison of the two approaches on a real-feeling problem.
The setup
Imagine you run a B2B SaaS company at $14M ARR. Your VP of Sales has just pitched you on a 30% expansion of the AE team, funded by trimming customer success. You have to take a position on it by Friday.
A manual what-if is going to look something like this: open the FP&A model, change the headcount line, change the CS retention assumption, watch ARR for the next four quarters move, repeat for a couple of scenarios. Maybe an hour, maybe an afternoon, depending on how clean the model is.
Where manual wins
Manual is good at three things and you should not let anyone talk you out of them.
It respects your judgment. Every cell you change is a cell you understand. The model isn't going to assume something on your behalf that you'd disagree with, because you typed every input.
The audit trail is trivial. When the CFO asks "where did the 14% churn assumption come from," you point at a cell and you say "I made it up, here's why." That's a clean conversation. An AI-generated assumption is harder to defend in the same room.
It forces specificity. You can't run a manual what-if without picking a number. The act of typing it makes you think about it. Engines that surface assumptions try to recover this property, but they don't quite get there.
Where the AI approach wins
Three places, also.
Branch coverage. A manual operator tries two or three scenarios, usually clustered around the answer they already suspected. An engine runs eight or ten, including ones the operator wouldn't have thought to model — the branch where the new hires don't ramp, the branch where the CS cut causes a churn cascade, the branch where the market shifts under both.
Sensitivity, automatically. Spreadsheets technically support sensitivity tables. Nobody uses them. The engine surfaces "your outcome is 4x more sensitive to ramp time than to headcount count" as part of its default output. That information was already in the model. The engine just bothers to look.
Speed when the problem is fresh. The first what-if on a new question is the most expensive in spreadsheet terms, because you have to model it. The engine has no setup cost — you describe the situation, you get the tree.
Where both fail
Both approaches assume the underlying numbers are right. If your churn data is wrong, both will be confidently wrong in the same direction. Neither tool catches that for you.
Both also struggle with novelty. A what-if about cutting CS depends on prior data about what happens when CS gets cut. If your situation has no analogue — say, a regulatory regime that doesn't exist yet — both approaches degrade into structured guessing.
A practical hybrid
The operators we see getting the most value run the engine first to scope the problem, then drop into the spreadsheet to defend the specific number. The engine tells them which assumption matters most; the spreadsheet lets them argue for or against it cell by cell. The engine is the wide net, the spreadsheet is the spear.
That order matters. If you start in the spreadsheet, you anchor on the first scenario you build and you never escape it. If you start in the engine, the spreadsheet becomes a tool of rigor rather than a tool of confirmation.