For finance leaders
Three cases in a spreadsheet, each one a single path, each one assuming customers hold still and competitors do not exist. Counterfirm runs the same decision four hundred times with the uncertainty drawn from your own numbers, and returns a p10, a median and a p90 with the assumptions listed underneath.
Build a twinHow the ledger works
Every number in a run traces back to a line in a file you uploaded, or is labelled as an estimate or a default.
Best case, base case and worst case are three guesses wearing a suit.
They are built by taking the same model and nudging two inputs in the same direction, which means the worst case is not a probability, it is a mood. A distribution built by drawing every uncertain parameter independently, hundreds of times, is a different object: the p10 is a tenth of the runs, not the pessimistic version of one run.
The ledger
Each assumption in a run carries where it came from: read from a document with the line quoted, derived from something that was, estimated by a model, set by you, or an industry default. Nothing is blended into a single confident figure.
Each one also carries a band. Churn has a wide one, because churn is usually wide. Headcount has a narrow one, because you know your headcount. The bands are what the draws come from, so a run is not a sensitivity table bolted on afterwards, it is uncertainty carried the whole way through.
The finance questions
Gross margin moves in month one. Then the real question: how much you pass on, and what the pass through does to volume. The run carries both, with a pass through share you can set rather than assume.
Supplier raises pricesRevenue goes on a date. Cost does not. The run shows the gap between the two, the month cash gets uncomfortable, and how much of the recovery depends on a pipeline that has not been built yet.
Lose the largest customerMargin now against volume later, with a competitor who matches part of it, on a lag, and customers whose contracts end in different months. The sweep gives you the whole curve rather than one point on it.
Raise priceSeverance, the month the saving actually starts, morale, attrition that follows, and the revenue that was attached to the people who left. The saving is rarely the number in the paper.
Reduce headcountCash and covenants
A median that looks fine can sit above a distribution in which one run in seven breaks a covenant in month eight. The average never shows you that. The band does, and so does the count of runs that crossed the floor.
Cash on hand, the cash floor before a covenant bites and the months investors will fund a dip are all assumptions in the ledger with bands of their own. Set them to your actual facility terms and the run tells you not just whether you breach, but in which months and on the back of which event.
Concentration
The working method
Keep your model. It is the right tool for the general ledger, the tax line and the way your business actually books revenue, and no simulation is going to reproduce those. What the model cannot do is put a competitor and a customer with a mind in the same sheet.
The useful pattern is to run both and look at the gap. When the twin gives a materially different answer to your own plan, one of two things is true: the twin is standing on a default you can replace in a minute, or your plan is assuming something about behaviour that nobody has written down. Both are worth finding before the board meeting, not after.
Honestly
It is not an accounting system and it does not touch your ledger. It does not do revenue recognition, tax, consolidation, intercompany, currency translation or anything that has to be right to the cent. It operates in monthly aggregates and it rounds.
It will not audit. Nothing here is prepared to a standard, reviewed or signed. A run is a model output and it should be described as one in anything it appears in.
It is only as good as the bands, and the bands start as industry defaults. A twin built from one profit and loss gives you a wide band that is honest. Replacing defaults with your own history narrows it. Nothing narrows a band by asserting confidence.
The run
Every assumption has a relative half width in the dictionary. Churn carries a wide one, because monthly churn is genuinely uncertain in most records. Headcount carries a narrow one. Each replication draws its own value inside that band, so the spread you see is the spread implied by your own record rather than a fixed plus or minus fifteen percent.
When a document pins a number precisely, the band narrows. That is the direct reward for uploading more.
Yes. Runs keep their seed and their full assumption set. The same question with the same inputs gives the same answer, which matters when somebody asks in month six where the figure in the pack came from.
The simulation is arithmetic and runs with no API key. A language model is used optionally in three places: reading prose the parser could not handle, rewriting a finished brief into better English, and answering questions in Ask. Anything it extracts arrives in the ledger labelled as an estimate before it can reach a run.
Then the readiness score is low and the ledger is full of defaults, and both of those are stated plainly on the overview. The missing files are listed in the order it is worth fixing them. A wide honest band beats a narrow invented one.
Management accounts, a customer list and the question. If the band comes back wider than you would like, the ledger tells you exactly which document would narrow it.