How to build an annual budget with Claude

Written up from the video How and Why to Build a Budget with an AI Agent in Claude.

You can use Claude to build an annual budget by setting it up as an agent and taking it through six small steps: normalise last year's actuals, challenge your assumptions, then build revenue, people cost, the rest of the cost base and the phased P&L. Claude does the compiling, the schedules and the formulas. Leadership direction, the challenge of department managers and the sign-off stay with you.

In the video I build a full twelve-month budget for a fictional Dubai creative agency in about an hour. It will take you longer the first time. It will still be a fraction of the three to six weeks a budget usually takes in a small or medium-sized business.

The five steps of a budget, and which ones go to Claude

You cannot decide what to hand over until you see the whole process. In an SME, a budget is usually five steps:

1. Leadership sets the direction. A meeting, or several, between the owners and finance. This stays with you. Claude cannot sit in the meeting.
2. Department managers give their inputs, and you challenge them. A negotiation, because somebody has usually padded something. This stays with you. Claude can consolidate what you collect, once you have it.
3. Normalise last year's actuals. Strip out the one-offs to expose the real run rate and the real trend. Claude, with you supervising.
4. Finalise the assumptions and build the input sheets. Revenue, people cost and the rest of the cost base. Claude, with you supervising.
5. Consolidate into the budget. A twelve-month phased P&L. Claude, with you supervising.

Two steps stay with you. Three go to Claude, with you supervising every one of them. If someone tells you AI can produce a budget end to end from one clever prompt, they have probably never built a real budget with it.

Why an agent, not a one-off chat

A budget happens once a year, so the fair question is why not open a new chat. Three reasons.

• It holds your method. Every business budgets slightly differently. If that method lives only in your head, or in a chat you closed six months ago, you rebuild it every year.
• It works inside the boundaries you set. Most of what goes wrong with AI in finance goes wrong because nobody set the rules up front.
• You build it once and it keeps paying. Next year you are building inputs, not starting again. Something changes mid-year and you reflow it. Six months from now, when someone asks why the revenue assumptions are what they are, you ask the agent instead of digging through your own files.

What the agent needs

Give it the same things you would hand someone joining your team:

• The latest management pack, with the actuals, chart of accounts, general ledger extract and breakdowns. The output has to match its format.
• Last year's budget, if you have one, to show how the business phrased its plan and where that plan turned out to be wrong.
• A business context note: what the business does, how it makes money, its commercial terms.
• The strategic direction notes from leadership: expansion plans, hiring and leavers, what the business is trying to achieve.

Leave two things out on purpose. Anything that already contains the answer, such as a rolling forecast that carries a P&L into the budget year, because it will simply be copied; keep it back as an independent check at the end. And anything you have not checked yourself, because unverified data produces a confident, wrong answer.

The instructions that make it an agent

The instructions are what turn a folder of files into an agent you can rely on. Mine cover the scope (the P&L, not cash flow or scenarios unless asked), the reporting basis, and a strict order of sources: contracted data first, then closed actuals, then approved forecasts, then approved assumptions. Nothing outside that order counts, and where a figure does not exist it is recorded as a gap rather than estimated.

They also set the constraints. Claude does not supply strategic direction. It does not build towards a target: if the budget falls short, it quantifies the shortfall and sets out options instead of adjusting drivers to close it. It does not resolve a contradiction between files on its own. And it works in stages, presenting each one and waiting for sign-off before the next.

The six steps

1. Normalise the base year

Claude goes through last year's actuals and strips out everything that will not repeat, such as a legal settlement, an office move or a client that has gone, listing each item with its value, month and reason. Reported and normalised sit side by side. What to check: that the list is complete, and that nothing recurring has been removed. An agent that quietly strips the annual audit fee will flatter next year.

2. Challenge your assumptions

Write your assumptions first, from the trend analysis and your meetings with management and the department heads. Then ask Claude to interrogate them: which ones the base year does not support, which carry the most risk, and what you have not thought about. It challenges them; it does not answer them for you. What to check: that it actually argued with you. If it agreed with everything, ask again.

3. Revenue

Claude builds monthly revenue outward from what is contracted: the retainer book, price changes at each renewal month, new business at an agreed value and start month, and a stated churn provision. Where there is no pipeline, it uses a stated run rate and labels it as one. What to check: that the opening month ties back to the base year, and that anything called pipeline genuinely is pipeline.

4. People cost

Claude starts from the roster at year end, adds each hire in its start month and each leaver in theirs, applies the salary review from its effective month, and splits direct from indirect cost, broken into salaries, benefits and statutory charges. What to check: headcount by month agrees to your hiring plan, and the average cost per head is a number you recognise. People cost is usually the biggest line in a service business, and a hiring decision flows through everywhere.

5. The rest of the cost base

Claude builds every remaining line month by month: contracted where a contract exists, run rate where not, variable external cost as a stated percentage of the revenue it serves, and every one-off on its own line in its own month. What to check: that a one-off has not been smeared evenly across twelve months.

6. Build the budget model

Claude consolidates everything into a twelve-month phased P&L in exactly the same line order as the management pack, with no typed numbers: every figure links to a schedule, and a tie-out block proves each total agrees to the schedule behind it. What to check: click through a few cells. If any holds a typed number rather than a link, the model is not finished. The format rule matters more than it sounds: if the budget and the management reporting do not match, every variance conversation for the next twelve months becomes a translation exercise.

Where your judgement comes in

Look at what happened across those six steps. I did not ask for a budget in one go. I broke a large task into small pieces, gave the agent what it needed for each one, and checked each piece before moving on. That is how you control mistakes and get an output you would put in front of a board.

The time you get back goes into the part only you can do: talking to the people who own the numbers, testing what the plan actually depends on, and deciding what the business should commit to.

What good looks like

• One workbook, with a tab per schedule, in build order.
• Assumptions on their own visible tab, never buried inside a formula.
• The P&L in your management reporting format, line for line.
• Every figure on the P&L links to a schedule. Nothing typed.
• A visible tie-out showing each total agrees to the schedule behind it.
• Blue for inputs, black for formulas, green for links between tabs.