It's October. If your financial year ends in December, someone in your company is about to open last year's budget, add 5%, divide by twelve and call it a plan.
In May I wrote about why that budget is lying to you. A real budget is built from drivers, rebuilt from zero where the money is, and phased to the months things actually happen. Owners agreed with the argument. The most common reply was: "Fine, but who has two months to build that?"
You don't need two months any more. Most of the time in a budget goes on legwork: pulling actuals, sorting lines, building the calendar, laying out scenarios. That's the half a model like Claude does well. The other half, deciding what the business will try to do next year, is the half that makes it a budget. That half stays with you.
Here's the method, week by week, with the line drawn clearly between the two halves.
Before You Start: Set Up the Room
- Use a commercial account. Claude for Teams or Enterprise, not someone's personal subscription. Your numbers shouldn't sit anywhere whose terms allow training on them.
- Open one Project for the budget. Load the inputs once: 24 months of monthly P&L, last year's budget, the payroll list (anonymised if you prefer) and your contracts and renewal dates.
- Keep the spreadsheet as the source of truth. Claude writes and checks the formulas; the spreadsheet does the arithmetic. A number should never live only in a chat.
Week One: Build the Base
1. Sort the actuals. Ask Claude to tag every P&L line as fixed, variable or one-off, and to flag anything that looks odd: a month that jumps, a cost that appeared once, a supplier that changed. It's slow, tedious work for a person and a fast first pass for a model.
Your call: confirm the one-offs. Claude can see that March had an unusual insurance payment. Only you know whether it happens again. Whatever you confirm as a one-off comes out of the base before anything is projected forward.
2. Find the drivers. Give Claude the cleaned actuals and ask what moves revenue and cost: headcount, projects per month, average fee, covers per day, ad spend. It will usually suggest more than you need.
Your call: pick the three to five that really move the business, and drop the rest. A budget with twelve drivers is a model nobody updates. A budget with three is one you'll still be using in March.
3. Phase it to the calendar. Ask Claude to lay out the year month by month: GST quarters, CPF, annual insurance, licence renewals, the 13th month and bonuses, Chinese New Year, the slow weeks in your sector.
Your call: check it against reality. Last year's bank statements are the test. If the calendar says the bonus lands in December and your cash says January, the cash wins.
Week Two: Make the Calls
4. Rebuild the big lines from zero. Not every line deserves a fresh look. Payroll, marketing, software and rent usually do, because that's where the money is. Have Claude draft a one-page justification for each: what we spent, what it bought, and what we'd lose without it. Each budget owner fills theirs in.
Your call: challenge the answers. "We spent it last year" is not a reason.
5. Run three scenarios. Base, downside and stretch. Claude builds the spreadsheet logic so that each scenario changes only the drivers, never the formulas.
Your call: set the triggers. Decide now what you'll do if revenue is 15% behind by the end of March: which hire waits, which spend stops. A downside case without a trigger is just a sad number.
6. Red-team it. Ask Claude to argue against your budget the way a sceptical banker or investor would. Where is revenue optimistic? Which costs are too low for the growth you've assumed? What's missing altogether?
Your call: decide which objections count. Some will be fair, some will be generic. Telling them apart is judgement, not arithmetic.
7. Lock it on one page. Revenue by driver, costs as a percentage of revenue, the cash low point and when it lands, and the triggers. That page goes to the leadership meeting. The model stays behind it.
What I Don't Let Claude Decide
There's one line I hold in every budget: Claude can inform any decision, but it doesn't make one. Each of the five calls below stays with the owner. What changes is how well informed that call is when you make it.
1. Setting the targets. Ambition is the owner's job. A model will happily extrapolate last year; it can't decide what the business should try to become. What it can do is show you where your ambition sits against the market. Ask it for growth rates in your sector, typical margins for a business your size, and what comparable firms have done over the last three years, from listed peers, competitors' published accounts and industry reports. Then you set the target knowing whether you're aiming with the pack or above it.
Check: every benchmark needs a source you can open. A number without one is a guess with good grammar.
2. Choosing the drivers, and the numbers on them. Claude can measure how much each driver actually moved last year's P&L: what a 10% change in projects per month does to profit, against a 10% change in the average fee. That tells you where effort pays. It also arms you for the budget meeting. If your sales director proposes 5% growth in a sector that grew 20%, that isn't prudence. It's either sandbagging or a pipeline problem, and either way it's worth a conversation. If they propose 40%, Claude can lay out what has to be true for it: how many more projects, at what fee, delivered by how many more people. The meeting becomes a discussion of assumptions instead of adjectives.
Your call: which drivers you'll run the business by, and which targets you'll accept from the people who own them.
3. Deciding headcount. Hiring, and not hiring, are decisions about people, and they stay with people. The economics behind them are modelling, and Claude models well. Have it compare each planned hire with the alternatives: contractors or overtime as a variable cost against a permanent salary as a fixed one, the volume at which the hire pays for itself, and what happens to margin if the work doesn't arrive. Then run the same exercise on the people you already have. A replacement costs a recruiter's fee (commonly 15 to 25% of the first year's salary), months of ramp-up and the output lost while the seat is empty. Set that against the cost of a retention bonus or a pay adjustment for the two or three people you can't afford to lose.
Your call: who you hire, who you keep, and when.
4. Being the calculator. Claude writes and checks the formulas; the spreadsheet does the arithmetic. A number that only exists in a chat window isn't in the budget. Where Claude earns its keep here is as an auditor. Ask it to trace every formula in the model, find hard-coded numbers hiding inside formulas, confirm the scenarios change only the driver cells, and flag any total that doesn't add across and down.
5. Signing off a number that doesn't tie to the ledger. Every figure in the final budget traces back to the actuals. Claude can do the reconciliation legwork: tie the base year to the trial balance, list every difference and propose an explanation for each. You sign off when the list is empty, or when every item left on it has an explanation you accept.
Where This Breaks
- Messy books. If the month-end close isn't reliable, neither are the actuals, and Claude will sort the wrong numbers very quickly. Fix the close first. I've written about how long it should take.
- Less than a year of history. With no seasonality to learn from, build from the drivers alone and treat the calendar as a first guess you'll correct by June.
- Long conversations. Two weeks of back-and-forth is a lot for any model to hold, and it can lose track of which version of a number is current. That's manageable, and the next section is how.
Managing a Long Build
Claude only works from what's in its context window: everything in the current conversation, up to a limit. A two-week budget build fills that window, and when it's full the oldest detail gets summarised or dropped. That's how a model ends up quoting February's draft of a number in March. Five habits stop it.
- Use the biggest window you can. For a build this size I work in Claude Code, which can hold around a million tokens of context on the models that offer it: enough for two years of monthly actuals, the drivers, the calendar and weeks of back-and-forth. Typing
/contextshows how full the window is. - Compact at half, not at the limit. Compacting replaces the conversation so far with Claude's own summary of it, freeing room to carry on. Left alone, it happens automatically when the window is nearly full, and you don't choose what survives. I compact at about 50%, roughly 500,000 tokens, and tell it what to keep:
/compact keep the confirmed one-offs, the three drivers and their values, and every trigger. - Keep a decision log outside the chat. One plain file in the Project (in Claude Code, a scratchpad file in the folder) where every confirmed decision goes the moment it's made: the one-offs excluded, the drivers chosen, each assumption and who agreed it. Have Claude add to the log at the end of every step and read it at the start of the next. The log survives any compaction. The conversation doesn't.
- Write the standing rules down once. The financial year, the currency, how the chart of accounts maps to budget lines, "never hard-code a number": put them in the Project's instructions, or a
CLAUDE.mdfile in Claude Code, so they load every session instead of being repeated and half-remembered. - Read the file, don't recall it. Start each step by having Claude open the current spreadsheet and the log. When the chat and the file disagree, the file wins.
A Budget Is a Conversation
The point of a budget was never the spreadsheet. It's the conversation the spreadsheet forces: what are we trying to do next year, and what will we do when it doesn't go to plan? Most SMEs never get to that conversation, because the legwork eats the time.
Give the legwork away. Keep the conversation.
Tested on Claude, October 2026.