Why and how to set up FP&A for a small team

# Why start before you feel ready

The usual answer is that you start once you can afford someone to do it. That gets it backwards: at fifty people you can afford the salary, but you cannot buy back the three years of history you never measured.

1. Start while it is still small. Ten cost centers can be restructured in an afternoon. Fifty take a quarter - and by then somebody is reporting on the old structure and somebody else on the new one.
2. Plan for growth instead of reacting to it. A plan built on what actually drives each line tells you what the next ten hires cost before you make them. A plan of typed-in numbers only tells you what you already assumed.
3. Give managers and investors one set of numbers. Not so that everyone agrees, but so the argument is about the decision rather than about whose spreadsheet is right.
4. Bad news arrives earlier. A monthly loop finds the problem in week six. The year-end close finds it in month fourteen, when the only options left are expensive.

# The value is the record

You already do FP&A. Somebody looks at the numbers, forms a view, and acts on it. What is missing is not the thinking. It is the record.

A number on its own tells you nothing. It only means something next to another number produced the same way - the same month last year, the plan you set in January, the same line three months ago. That comparison is what turns a figure into a measurement instead of an anecdote.

Which is why doing it brilliantly once is worth so little. A sharp analysis in March that never gets repeated leaves you with one number and no trend, and you cannot tell afterwards whether it was good or lucky.

It is also why this cannot live in one person's head. A head keeps no record, so there is nothing to compare this year against last. And the company cannot be handed to anyone else, cannot be sold without a painful review by the buyer, and does not run well while that person is away.

# Who does not need this

• More than one revenue stream.
• Stock.
• Debt with a covenant - a loan with conditions you have to keep meeting.
• People who own their own budgets.
• Or simply the point where nobody sees all of it at once.

In practice that is usually from about ten people upwards, but the test is better than the number.

Below that line it is genuinely not worth it. Take a six-person consultancy: one revenue stream, no stock, no debt, and customers who pay in fourteen days. One person can hold that business in their head and be right. Adding a process there would only add a meeting.

If you are past that line, the next question is who does the work - because the answer people assume is the expensive one.

# Why large companies hire an analyst

Start with the thing you are already better at than a large company: the time between spotting a problem and deciding what to do about it. Forty people beat four thousand, every time. The person who can act is already in the room. Large companies employ part of their finance team just to close that gap. You do not have the gap - what you do not have is anyone to do the repetitive half.

Large companies do employ FP&A analysts. Smaller ones often cannot afford one, and some could but would rather spend the salary elsewhere. Either way, people draw the wrong lesson from it.

Adding up forty companies, mapping twelve different sets of accounts, rebuilding a four-thousand-row report every month. That is what the headcount is for. The judgment half never grew with the company. It is still a few people asking which assumption was wrong, and it is still the part that changes what you do.

The monthly loop splits every stage in two: the part that can be prepared for you, and the part you have to judge. The prepared part is the entire reason a large company needs the headcount, and it is the only part software can take over. So the question is not how to afford an analyst. It is how to get the prepared part done without one.

# Who runs it, and for how long

What no system does, at any price:

• Decide whether to act on a variance, accept it, or ignore it - see ending with a decision.
• Choose which assumption to change when the plan misses.
• Decide what to leave out.
• Own the number in front of the board.

You need less than people expect. One person who can read a P&L, who will ask why a line moved, and who is senior enough to get an answer. That is the whole list. It is not a finance qualification.

# Where AI helps, and where it does not

This is the strongest objection to everything above. If a model can do the repetitive half on demand, you do not need a system - you need a good prompt.

And here is the strongest form of the argument, which is largely true: in a small company, the person who would run this does not exist. An imperfect answer today beats a perfect answer you never get.

So assume the best case. A model that never makes an arithmetic mistake and never invents a number. The problem below still applies, and that is how you know it is not a question of quality.

FP&A is not one person asking a question. Someone submits their numbers, someone consolidates them, someone reviews, someone presents a version to the board. Even with no cost centers and no formal departments, someone owns each number and has to explain it. All of those people have to be looking at the same numbers. A chat gives each of them their own answer.

The system also has to remember: the mapping, the structure, the drivers, and every month you have already closed. If you measured something one way in March, you have to measure it the same way in November, or the comparison means nothing. Ask a model the same question twice and you get two reasonable answers.

Then there are versions. The frozen plan, this month's re-forecast, the numbers you showed the bank, last year's actuals. All of them exist at once, and everyone has to know which one they are looking at. A chat gives you a number with no label on it.

None of this is a bet against models getting better. It is about what FP&A actually is: shared between people, kept in versions, and done the same way every month. A model works with one person at a time, remembers nothing, and answers differently every time you ask - which is exactly what makes it good at the other job.

Explaining is where a model earns its place. Once your system has produced the number, a model is good at four things: saying what moved, drafting the commentary, pointing at the line worth a look, and answering a question so nobody has to hunt for where the number is stored. That is real value and you can have it today - on top of a system, not instead of one.

# Where to start

The first month is mostly decisions, not work.

1. Fix the structure once. Design the statements properly - a day of work, and the only part that is expensive to change later.
2. Pick the date and pick the person. Both, before anything else.
3. Compare against last year first. You have no plan yet, and last year's numbers work from month one.
4. Agree how big a deviation has to be before you investigate it. Do it now, while nobody feels accused.
5. Add the plan once three months of the loop are running, not before - then build it on drivers.
6. Then run it in a month when nothing happened. That is the month that tells you whether you have a system.