The engagement letter is signed on Thursday. On Friday the founder sends you a workbook called Budget 2026 v7 FINAL (2).xlsx, with a note: the board meets on Tuesday, can you present the forecast.
You did not build it. The founder did, over two years, with a previous CFO who left in March. It has 14 sheets, three of them called Sheet1, old and Copy of P&L. The cash line in month 9 is a number with no formula behind it.
You have four days to be able to defend it. And then, unlike a deal, you have to keep it alive every month after that.
This is a checklist for that week.
Why this is not a deal review
The transaction reflex is to read it like a data-room file: start at the output, trace the drivers, write a red-flag memo. That method is right, and I wrote it up for analysts who inherit a model on a deal.
A client model is a different job in three ways.
You will own it. A deal model is read once and argued over. A client budget comes back every month with new actuals, and every shortcut you accept in week 1 is a shortcut you run in month 6.
The builder is still in the room. The founder knows why row 47 is hardcoded, and may be attached to it. Your review is also a conversation with the person who pays you.
The numbers already have a history. The board has seen last quarter's forecast. Whatever you change has to be explainable against what they were shown, or the first thing you present looks like a restatement.
And you do this several times over. CFO Connect's 2026 benchmark of 516 finance professionals found that fractional CFOs typically serve 3 to 4 clients simultaneously (more in our fractional CFO statistics). Four clients, four models you did not design.
Day 1: freeze a copy and write down what it is for
Before opening a single formula, do two things.
Save the file as received, untouched, with the date in the name. This is your baseline. When the founder says in May that the forecast used to show a profitable Q4, you want to be able to open the version that said so.
Write 5 lines on what the model is for. Who built it, when, what decisions it has fed (a fundraise, a hiring plan, a bank covenant), which numbers the board or the investors have already seen, and what you are expected to produce from it next. Ask the founder for what you cannot answer. Twenty minutes on a call saves a day of reverse-engineering.
That list is the one people skip, and it tells you which cells matter.
Step 1: map it
Start at the outputs. Find the 5 to 10 numbers that leave the file: the cash runway, the EBITDA line the board tracks, the headcount plan, the covenant ratio if there is debt. Then trace each one back to where it comes from.
While you trace, build a sheet inventory, one line per sheet:
- Role: input, calculation, output, data (a ledger export, a bank statement), or unknown
- Read by: which other sheets reference it
- Time axis: monthly, yearly, both, none
The last column catches more than you would expect: a monthly P&L feeding a yearly cash sheet through a /12 is a timing assumption nobody decided.
Then list every sheet nothing reads. In a founder-built budget these are usually last year's version of a calculation, kept "just in case". They are noise, and you want them out of your head by day 2.
Step 2: find what is hardcoded
A deal model has usually been through a professional's hands. A founder budget has been edited by whoever was in a hurry that month, and a hardcode is the natural result: someone needed the forecast to show a number, and typed it.
It is also the most common way models fail. Across 85 operational spreadsheets examined through multi-day inspections, errors were found in 94% of them (Panko, EuSpRIG 2015; more figures in our spreadsheet errors and risks page). Not every error is a hardcode, but hardcodes are the ones that pass the eye test.
Three checks, in this order:
- Constants sweep. In Excel, select the calculation sheets and use Go To Special, then Constants, numbers only. Every literal outside the assumptions area is either an input in the wrong place or a hardcode. There is no third category.
- Pattern breaks. In a row where the same formula is dragged across 24 months, look for the one cell that differs. Month 9 overridden with a number is a typical find. Excel's green error triangle ("inconsistent formula") flags many of them.
- Perturbation. Move one driver by 10%, the one that should ripple furthest (volume, price, headcount), and read the outputs. Anything that does not move is disconnected. Then undo it.
That takes about 15 minutes and finds most of it. The full list of checks is in how to stop AI from hardcoding a financial model. It applies whether an AI or a founder at 11pm typed the hardcode.
Do not fix anything yet. Write each finding down with its location.
Step 3: find the drivers
Now you know where the model is live and where it is frozen. The next question is which few inputs actually carry it.
In most small-company budgets the answer is 3 to 6 lines: new customers or volume per month, price or average basket, gross margin, headcount by month with average cost, payment terms, and one or two business-specific items (churn, a large contract, a capex programme).
For each driver, find three things:
- Where it lives. One cell, or scattered across sheets? A growth rate typed in 4 places will be updated in 3.
- Where it came from. A historical average, a board target, the founder's gut. Ask. "Where does the 8% come from?" is a legitimate question in week 1 and an awkward one in month 4.
- How much it moves the answer. A crude sensitivity on runway is enough. If a 10% change in new customers moves runway by 2 months and payment terms by 2 weeks, you know where to look first.
Step 4: tie it to actuals
This step separates taking over a model from reading one. Before touching any forecast assumption, check that the model starts from reality.
Take the last 3 closed months and compare them line by line with the accounting: revenue, gross margin, payroll, total opex, closing cash. Then check that the opening cash of the first forecast month equals the bank balance at that date.
You will find gaps. The usual ones:
- Actuals typed in by hand, from a P&L export, with one month copied from the wrong column
- Accounts mapped differently from the accounting: the model's "Marketing" includes tools that the ledger books as software
- Timing: the model books annual subscriptions when invoiced, the accounts spread them
- Opening cash off because a financing inflow was forecast and never updated to what actually arrived
A forecast that does not start from the real opening cash is wrong in every month after, however good the drivers are. The monthly version of this problem, and why it eats hours, is in budget vs actuals.
Write down the account mapping you end up with. It is the most reusable thing you will produce this week.
Step 5: decide what to keep
You now have a map, a hardcode list, the drivers and a gap list. The temptation is to rebuild. Resist it for 2 weeks.
A simple triage, section by section:
| Section | Ties to actuals? | Hardcodes? | Decision |
|---|---|---|---|
| Revenue build | Yes | 2, known | Keep, fix the 2 |
| Payroll | Yes | None | Keep |
| Opex | Partly | Several | Fix the mapping, replace hardcodes with drivers |
| Cash | No | Month 9 | Rebuild from the P&L and working capital |
Rebuild only what you cannot defend. Every number the board has already seen must stay reconcilable to whatever you hand them next. If you change the structure, prepare the bridge: last forecast, what changed, new forecast, one line per reason.
What you send the client at the end of week 1
One page. What the model is for, what you trust, what you do not trust yet and why, what you are changing before Tuesday, and what you will change over the next month. The founder should be able to read it in 5 minutes.
It sets the scope of what you will defend in front of the board, and it records, with a date, the state of the model the day you took it over.
Making month 2 cheaper than month 1
Engagements are getting longer. Heidrick & Struggles' 2026 survey of 3,810 independent professionals found that 42% of projects now last longer than six months, up from 27% in 2021 (across all functions, not CFOs specifically). The takeover week is a one-off. The monthly run is the job.
What makes month 2 cheap is keeping three things from week 1: the account mapping, the list of known hardcodes and what replaced them, and the version of the model you took over. If they live in your head or in an email thread, you will redo part of this checklist every month, for every client. I went through what has to survive between working sessions separately; it is the same list, whether the session is yours or Claude's.
Where Layerz fits
Everything above works in Excel alone. If you take over one model a year, do it that way.
If you take over several, steps 1 to 3 are the mechanical part, and that is the part Layerz now does for you. Layerz is a structured spreadsheet, ready for AI, and since September 2026 it imports an Excel model through a single Import file button.
The import is deterministic, with no AI in it. It reads the workbook and turns its formulas into the model's live structure: a dragged formula becomes a line that computes, an input row becomes an assumption. What does not come across is not hidden. You get:
- A fidelity score: the share of the workbook's numbers that now live as formulas and values
- Kept as values: the lines frozen to their Excel value, grouped by cause (a cell that broke its row's pattern, a formula with no equivalent). When a whole line is frozen, its original Excel formula stays attached to it as a comment
- Not imported: what was left out and why, for example a sheet whose periods could not be dated
That report is the takeover map from steps 1 and 2, in the time it takes to upload the file. When the model totals a ledger sheet with SUMIFS per account and per period, that sheet comes in as a data source and those formulas become mapping rules: half of step 4's account mapping, written down for you. Not every ledger shape is read yet; the ones that are not show up in the report instead of disappearing. On any line, ⌘I opens the inspector: its properties, its notes, and what uses it.
Two limits, plainly. A high fidelity score means the model came across, not that it is right: a wrong assumption imports perfectly. And steps 4 and 5 are judgement. The import gives you somewhere to aim, it does not aim for you. When the client wants a file back, the model exports to Excel with live formulas. The verification cost does not disappear, it shrinks to the part only you can do.
The takeaway
The founder's workbook is not the problem. It holds two years of reasonable decisions about the business. None of them are written down, and you have four days.
Freeze it, map it, find the hardcodes by behaviour, find the few drivers that carry it, tie it to the ledger, and change only what you cannot defend. Then keep the map. Month 6 on this client, and week 1 on the next one, both start from it.
Top comments (0)