How AI helps with budgeting - and what budgeting is really about

If you are building the 2027 budget right now, you probably know how it ends. Ten business units send ten spreadsheets. Finance spends two weeks stitching them together. The board signs off in December. And in March, revenue is 8% below plan, costs are 3% above, and nobody can say exactly why - because the budget never contained the "why".

If you want to know how AI can actually help you optimize this process by as much as 40% - this article show a step by step guideline on how and were to use AI without a heavy tooling set-up, but with those tools you already have.

A budget is a file. Budgeting is a process.

A budget is the output. Budgeting is the process of deciding which drivers produce revenue, COGS, OPEX, CAPEX and EBITDA in each unit - and which activities will move those drivers next year. Most companies still build the file: in 2023 only 48% used driver-based models for budgeting or forecasting, and 96% of FP&A teams still plan in spreadsheets every week.

AI doesn't change that logic. It removes most of the manual work that makes companies skip it. Here is where it fits:

At every step AI proposes, checks or stitches - and finance makes the call.

Step 0: the data is the project

Be honest about where the time goes. In a 200m+ company with 8-12 units, history usually sits in two or three ERPs, a CRM nobody fully trusts and a stack of unit-level spreadsheets. In the AFP's 2025 survey, 61% of FP&A teams named data reliability as a challenge - ahead of any tool.

Two kinds of cleaning matter:

• Technical - one group chart of accounts, CRM clients matched to invoiced clients, consistent currencies, periods and cost centre hierarchies, mispostings moved to the right unit.
• Business - one-offs tagged and taken out of the baseline: the big one-time contract, the restructuring, the client lost in Q2, the acquisition in Q3.

The second one is where budgets quietly go wrong. Leave a one-time 2m contract in a unit's history and every driver built on it - revenue per client, conversion, cost per delivery hour - is inflated for next year.

What AI does here: proposes account mappings, matches records across systems and lists one-off candidates with the reason ("invoice 4x larger than this client's average"). What it doesn't do: decide. AI should not post mapping changes to the ledger, remove a one-off from the baseline, or reassign a misposting on its own. It also should not silently overwrite history or settle what counts as "one-off" - that is a business judgement. A controller accepts or rejects each item, and every change keeps a trail. The gains are real but modest.

Step 1: drivers and the activities behind them

A driver plan answers one question per unit: what has to happen for the number to come true? Here is an illustrative bridge for a service unit with a 24m revenue target:

Two things make this budgeting rather than a budget. First, the activities are in the plan - four reps and 190k of marketing are not overhead, they are what produces the 19 clients. Second, every line can be checked during the year. If the unit hits 95 qualified leads but wins 12 clients, you know the problem is conversion, not effort.

What AI does here: tests which drivers actually explained the unit's past results (did revenue follow client count, or deal size?), drafts the bridge from history, and builds a statistical baseline to compare the unit's own numbers against.

Step 2: don't model every unit the same way

The fastest way to kill driver-based budgeting is to build the bridge above for every unit and cost centre. The model gets too heavy to update, nobody checks it, and next year the team goes back to last year + 5%.

Decide the method before anyone types a number. A simple rule:

In a 10-unit group this often means three units with a full bridge - typically the ones carrying most of the revenue - four with a light model and three on simple growth. That is three serious models to defend in front of the board instead of ten half-built ones.

What AI does here: takes last year's P&L by unit, ranks units by share and volatility, proposes the method for each with a one-line reason, and drafts the input form each unit manager will fill in. You review a proposal instead of designing from a blank sheet.

Four AI workflows for this budget cycle

None of this needs a new planning system. It works on top of what you already have - Excel files from the units, exports from your EPM or ERP - with an AI assistant your company has approved for financial data.

The stress test is where most teams get the quickest win. A prompt as plain as this already works: "Here are the 2027 submissions from 10 units, the group assumptions and three years of actuals. List every assumption that differs from the group assumptions, every driver more than 20% above its 3-year trend, and every unit whose plan depends on one client or one hire." What comes back is typically things like:

• Unit C plans 12% wage inflation, the group assumption is 7%
• Unit F's conversion jumps from 14% to 22% with no new activity behind it
• Unit H's revenue growth needs two new sales reps who are not in its cost plan

None of these is hard to find. Nobody has the time to look for them across ten files in the week before the board meeting.

Worked example: how a PE investor reads your budget

In a PE-backed company the budget is not an internal plan. It is year one of the investment case. The investor reads it against the growth plan they underwrote and against the covenants on the acquisition debt - and they will stress-test it whether you do or not.

What they look for:

• Fit with the growth plan. Does this year put the company on the path to the exit EBITDA? Revenue growth, margin expansion and the timing of each initiative are compared with the plan.
• Quality of EBITDA. How much is recurring, how much is one-off, how big are the add-backs.
• Driver credibility. Which assumptions are above the historic trend, and what activity pays for the step change.
• Cash. EBITDA-to-cash conversion, working capital and CAPEX - cash, not EBITDA, services the debt.
• Track record. How close management's last two or three budgets came to actuals.

How they stress-test it: they flex the top three drivers by 10-20%, push new initiatives back by one or two quarters, strip out anything not yet proven, and check EBITDA, cash and leverage under each case.

Take a group whose growth plan needs 30m EBITDA in 2027, with a leverage covenant of 4.5x net debt / EBITDA:

The base case beats the plan. But a third of the growth depends on initiatives starting on time, and the combined downside breaches the covenant. The investor will find this in a week. Showing it first - with the cost levers you would pull and when - turns the board meeting from a defence into a discussion.

What AI does here: reads the growth plan, the investment case and past board packs, compares every budget assumption with them, and writes the investor's question list before the investor does. It also proposes which drivers to flex, based on which ones moved results most in the past. The scenarios themselves run through your financial model.

What AI won't do

• It won't fix the politics. Unit heads sandbag because their bonus depends on beating the plan. AI can show that Unit F's conversion is 30% below its trend - it can't make the conversation with the Unit F head any easier. What it changes is the starting point: you negotiate from evidence, not from last year + 5%.
• It won't sign anything. Every mapping, every one-off, every covenant definition has an owner on your team. If nobody is named as checking an AI output, it shouldn't go into the budget.
• It shouldn't do the arithmetic. Language models read, compare and draft well. Totals that must tie to the cent, and covenant ratios the bank will test, belong in a deterministic model.
• It needs the right setup for your data. Loan agreements and unit P&Ls are confidential. Use an enterprise deployment your IT and legal have approved, with access limited to the finance team - not a personal account.
• It won't make bad data good. If the history is wrong, AI will find patterns in the errors. Step 0 is not optional.

From our work: devices sold, leased and rented

One of the hardest budgets we have built was for a company that sells, leases and rents electronic devices. The same device can follow three paths, each with different revenue timing, margin and capital need. We budgeted each business unit and sales category - and the financing, which was the hardest part. Every leased or rented device has to be bought first.

We worked in five steps:

1. Core assumptions and drivers. We started with the drivers that actually run the business - such as devices placed per channel, the mix of sale, lease and rental, and contract terms.
2. Structure and rules. We decided how to structure the business: which units and entities exist, what the rules between them are and how they affect each other. That let us build one financial model where every driver connects to the right result.
3. Assumptions with the business. We sat down with management, sales and marketing and set the assumptions based on history and on their actual clients - and calculated how much financing each assumption required.
4. Financing. From device volumes we derived the capital need, looked for ways to reduce it and combined everything into one plan.
5. Costs. Finally we added the core OPEX.

The budget did two jobs. It was the basis for investor relations - the plan behind the next funding round and the financing for buying devices. And it became the core KPI set for sales, marketing, delivery and customer support.

During the year, the review runs on three levels - activities, KPIs, results - and each mismatch points to a different fix:

Looking back, the steps that took the most time - aligning history across units, testing assumptions with each team, and recalculating the financing for every change in the sale-lease-rent mix - are exactly the ones where AI helps most today.

Budgeting is not a file

Back to March. In a company that does budgeting, the meeting sounds different. Not "revenue is 8% below budget", but "conversion in Unit C is 18% instead of 22%, and the Q2 campaign didn't run". That is a conversation about action, not about numbers.

If you are mid-way through the 2027 budget, you don't need to rebuild everything. Three things are realistic before the board meeting:

1. Run the stress test on the unit submissions you already have.
2. Check the consolidated draft against your covenants - on a downside case, not just the base case.
3. Pick the two or three units that carry most of the revenue and turn their plans into a driver bridge with activities.

The rest can wait for the 2028 cycle. The file is the output. The drivers are the plan.

Sources

• FP&A Trends Digest - driver-based model adoption, 48% in 2023
• AFP 2025 FP&A Benchmarking Survey, via Rework - spreadsheets, data reliability
• Lionpoint - chart of accounts mapping with NLP
• Learnsignal - AI and covenant compliance