What happens in the first hour
An analyst opening your model for the first time is not reading your forecast. They are testing whether the file can be trusted at all, and they have a routine for it that takes under an hour. Everything that follows in diligence depends on how that hour goes.
The routine is remarkably consistent across funds, because analysts learn it from each other. Knowing it is not gaming the process — it is the difference between a model that answers questions and one that generates them.
- Press F5 or open the name manager to find errors, external links and broken references.
- Trace precedents on three or four headline cells to see whether outputs derive from inputs.
- Check whether the balance sheet ties to zero, and whether a plug is doing the work.
- Recompute one or two metrics independently — usually net revenue retention and gross margin.
- Change a single driver and watch what moves, and more importantly what does not.
- Look for hardcoded constants inside formula cells.
- Compare the model's historical period against the accounts they already have.
- Push a driver to a plausible extreme and see whether the model still behaves.
Test one: does the output derive from the input?
The single fastest way to assess a model is to select the year-five revenue cell and trace precedents. If the chain leads back through a volume build, a price assumption and a retention curve, the model is doing work. If it leads to last year's revenue multiplied by a growth rate typed into a cell, the forecast is an assertion with a spreadsheet wrapped around it.
This is why driver-based builds matter beyond neatness. A driver-based model can be interrogated. When an investor disputes your growth rate, you can show them the three assumptions underneath it and argue about those instead — which is a much better conversation than defending a single number you chose.
Test two: the plug
A balance sheet that ties only because a line called 'other assets' or 'balancing figure' absorbs the difference is the most damaging finding in diligence, and it is common. It means something in the model does not reconcile and the author either did not know or chose not to resolve it.
The problem is not the plug itself; it is what it implies about everything downstream. If cash is derived from a balance sheet held together by a plug, the cash-out date — the number the entire investment decision rests on — is unreliable. Once an analyst finds a plug, they stop reading the model and start auditing it, and the timetable slips.
Test three: independent recomputation
Analysts recompute two or three metrics from the raw data rather than accepting the model's own calculation. Net revenue retention is nearly always one of them, because it is both heavily cited and easy to compute incorrectly.
The most frequent error: applying expansion and churn to total current revenue rather than to the prior-period cohort base. That overstates retention, and because retention compounds, the effect on year-five revenue is large — commonly 15% to 25%. When an analyst's recomputation comes out ten points below yours, every other metric in the file becomes suspect.
- Net revenue retention, computed on the prior-year cohort base only
- Gross margin, including the costs founders often exclude — support, hosting, inference, payment fees
- CAC payback, on fully loaded acquisition cost rather than paid media alone
- Burn multiple: net burn divided by net new ARR, which cannot be flattered by presentation
Test four: does it survive being pushed?
The last test is to break it. An analyst sets growth to zero, doubles churn, delays the raise by six months, or pushes a cost line to a plausible worst case, and watches what happens. Two failures are common. Either the model produces an absurd result — negative headcount, revenue from customers who have all churned — or nothing moves at all, revealing that the driver was never actually connected.
A downside case that still shows healthy profitability is treated as evidence of unseriousness rather than of resilience. Investors have seen hundreds of them. A downside case that shows the company surviving on specific, named actions — a hiring freeze at a stated month, a pricing change, a facility drawn — is the version that builds confidence.
What to do before the data room opens
Run the same routine on yourself. Trace precedents on your headline outputs. Confirm the balance sheet ties without a plug. Recompute retention on a clean cohort base. Remove every hardcoded constant from a formula cell. Push each major driver to an extreme and watch for nonsense.
Then have someone who did not build the model do it again, because the author is structurally the worst person to find their own errors. Two days of independent review before the data room opens is materially cheaper than two weeks of re-opened diligence after an analyst finds something — and a correction you disclose yourself is a housekeeping item rather than a credibility problem.