At a public institution I worked with, the annual budget had two seasons. At the beginning of the year, each department asked for the largest allocation it could secure. If the funds ran out halfway through the year, spending commitments stopped and necessary purchases could not be made. If money remained in November, the incentive reversed: it had to be spent before year-end, because low budget execution made it harder to defend the same allocation the following year.
In companies, the operating budget is built bottom-up, usually in Excel. Business units state what they expect to earn or spend and defend those figures before management, which may reduce or increase them—sometimes at its discretion.
We call this financial planning, although much of the time goes into negotiation. When a number affects performance assessment and bonuses, the person proposing it is answering two questions: what will happen, and what do you want people to expect from me? Treasury is usually more candid out of necessity: cash availability and payment dates impose hard constraints.
In 2003, Michael Jensen of Harvard Business School gave this mechanism a rather undiplomatic title: Paying People to Lie. When pay depends on the gap to budget, he argued, managers and subordinates misrepresent the budget during its construction and game it during execution.
The company is left without the unbiased information it needs to coordinate its parts. Jensen himself acknowledged that the evidence on the cost of this problem was limited.
The criticism predates the article. Under Jan Wallander, Sweden’s Handelsbanken stopped using budgets in the early 1970s, seeing them as an unnecessary evil. In 1998, Jeremy Hope and Robin Fraser founded the “Beyond Budgeting” movement and made the bank its reference case. The proposal replaces the annual budget with rolling forecasts and gives greater autonomy to those making decisions.
Have these financial processes changed?
I thought this old world should have disappeared by now. Today there are cloud planning platforms and language models capable of commenting on a trial balance with more patience than many accountants. Yet surveys of financial-planning teams tell a different story.
In the Association for Financial Professionals’ 2025 survey, 96% of financial-planning professionals used spreadsheets for planning, regardless of company size. Of the same respondents, 71% also used enterprise-performance-management (EPM) platforms at least once per quarter.
Even so, spreadsheets still underpin planning: 82% use them to prepare data for the platform and 57% use them to work around it. The company bought the system, and Excel remained at work around it.
The FP&A Trends series has asked since 2017 how long it takes to produce a forecast. In 2017, 23% of organisations produced one in two days or less; in 2025, only 15% did. The average time rose from 5.75 to 6.09 days between 2017 and 2024, while the proportion rating their forecasts good or very good—42% in 2024—was virtually unchanged from the first year. Data collection and validation consume 46% of these teams’ time, the highest level in five years.
The 2026 figures do not change the picture. The AFP found a budget cycle of nearly nine weeks, unchanged for three years, and concluded that investment in planning technology has still not delivered the expected efficiency gains. In a survey of 314 organisations on finance-team technology, Gartner again found spreadsheets to be the most widely adopted technology and reported declining value and investment in extended-planning platforms.
These figures deserve careful reading. The FP&A Trends surveys are sponsored by planning-software vendors—SAP in 2024 and OneStream in 2025—and forecast quality is self-assessed. The samples are weighted towards larger companies: in 2025, only 8% of respondents worked for businesses with less than $25 million in revenue. Still, it is difficult to accuse companies selling planning platforms of pessimism.
The working days are being spent on something else. Software performs in seconds what once took an afternoon, but the forecasting calendar is still set by the people submitting numbers and by those discussing them. And the number is still negotiated: in 2025, 38% of organisations planned using “last year plus x%”—a negotiating method dressed up as a forecasting method.
What do elite finance teams do?
Some companies have shortened this cycle. Published cases come almost entirely from vendors and consulting firms: KeyBank says it reduced its forecast process from more than five days to one, in one of the largest Workday-based finance transformations in the United States. At Umicore, the finance director says a full forecast cycle used to take a working week and now takes less than a day.
What both describe is primarily the removal of manual work: data previously entered by hand now flows into the model automatically. The other feature that distinguishes the strongest teams is a driver-based model.
Rather than forecasting hundreds of account lines, they forecast volume, price, headcount and average collection days, then let the accounts follow. This has an advantage that is rarely obvious. In a driver-based model, anyone seeking to defend a number has to defend the underlying assumptions.
In FP&A Trends, 77% of organisations with fully or dynamically driver-based models rated their forecasts good or very good, compared with 27% of those using basic models or no model. It is a self-assessed correlation—capable teams are likely to have both better models and better forecasts—but the gap was already visible in 2024.
What forecasting competitions show
Statistical methods extend a series’ history according to a form chosen in advance. Exponential smoothing, for example, takes an average in which recent months carry more weight than earlier ones and adds trend and seasonality. ARIMA links each value to values and errors from previous periods. These methods have few parameters, work with limited history, and have existed for more than half a century.
Machine learning reverses the logic: rather than the analyst selecting the form, the algorithm learns it from many examples. It can use data outside the series itself, such as price or a promotions calendar. In forecasting, many methods in this family combine hundreds of decision trees, each correcting the errors of those before it.
Deep learning is a branch of machine learning that uses neural networks with many layers—hence “deep”—and belongs to the same broad family as language models. The more flexible the method, the more data it requires and the harder it is to understand why it forecasts what it does.
To establish which methods forecast best, Spyros Makridakis has organised forecasting competitions since 1982. Organisers withhold the most recent values of real series and ask teams around the world to forecast them. The winner is the one that comes closest to what actually happened. The fourth edition, M4, ended in 2018 and had 61 methods forecast the same 100,000 series.
The result disappointed those expecting a victory for the machines: none of the six pure machine-learning methods beat a simple average of three statistical methods, while a hybrid combining both families won. Years later, Amazon researchers ran a pure neural network on the same series without special tuning and came close to the winner.
In the M5 competition, in 2020, the problem was different: 42,840 Walmart sales series, many of them similar to one another and accompanied by external data. Here, machine learning won.
For a mid-sized company, the lesson is practical. Its forecast may cover a few dozen line items with five years of monthly data, and in that setting simple methods tend to win, for now. Machine-learning methods pay off when there are thousands of similar series and variables that explain them. Complexity must earn its place against a simple alternative, and that can only be established by testing.
How should forecast effectiveness be assessed?
A forecasting system needs something Excel models generally do badly: it must remember what they forecast. If in April we forecast September, and in July we forecast September again, the difference between the two versions tells us more than a comparison with a budget prepared in October of the previous year. Version control also makes it possible to measure where the company is wrong and at which forecast horizon.
Suppose expected EBITDA for the year falls from €4.2 million to €3.7 million between two versions. The system decomposes the change: €360,000 comes from the margin of one deal and €140,000 from the delayed start of two customers. The movement breaches the materiality threshold set by the company and anticipates a cash shortfall five months ahead.
A large language model (LLM) can help write a paragraph explaining the cause of the change and the decision it requires: it compares versions and proposes an explanation. The forecast itself should remain with models that can be tested against what actually happened. In an implementation guide published in March 2026, Gartner describes AI cash forecasting as the use of traditional driver-based techniques that produce probabilistic, explainable forecasts for liquidity and scenario planning.
What can a mid-sized company do without spending millions?
The architecture that answers this question has six layers and can be based on SAF-T, the ERP, or a combination of both. No layer requires a planning platform. Most of the investment goes into model design and the first weeks of work; licences and infrastructure carry little weight.
1. Data sources
Each driver should come from the fastest source that can serve it. SAF-T has the advantage of being consistent across certified systems. The invoicing file contains sales line by line and is sent to the tax authorities by the fifth day of the following month. The accounting file contains every posting and must be capable of being produced during an inspection, although routine submission remains postponed.
A company with an integrated ERP can read directly from its database or API and gain daily updates on its order book, inventory and cost centres.
| Source | What it feeds | Frequency |
|---|---|---|
| Invoicing SAF-T | Sales by customer, product, quantity and price | Monthly, by the 5th, or real time |
| Accounting SAF-T or trial balance | Income statement and balance sheet by account | After close |
| ERP database or API | Orders, purchases, inventory and cost centres | Daily |
| Banks, through open banking | Balances and transactions | Daily |
| Payroll | People and cost per employee | Monthly |
| CRM | Opportunities and expected closing dates | Daily |
The limit of every accounting source is the close. Portugal’s December monthly VAT return is due by 20 February, and accounting that closes at that pace provides actuals a month and a half late. A written agreement with the accountant to close by the tenth working day is the least expensive line in the project.
2. Data infrastructure
Scheduled jobs read files and APIs and write everything to a small cloud database, in two zones: the data as received and the processed data. Between them sits the project’s most valuable component: a mapping table that connects each account and item to the driver that feeds it. Every month, processed data must reconcile to the trial balance. If it does not, the forecast begins from the wrong number with complete confidence.
Access is restricted by role because the accounting SAF-T contains everything and payroll contains personal data. The tools can be those the company already pays for, or a lightweight stack using a PostgreSQL database and Python routines.
3. Models
At the centre is a driver-based financial model: 15 to 25 drivers, each with an owner, source, method and frequency; integrated financial statements; cash derived from payment terms; and a rolling horizon of 12 to 18 months. What changes from driver to driver is the forecasting method, and the choice is made through a competition.
| Family | When it fits | Investment |
|---|---|---|
| Rules | Known relationships: salaries, contracts and payment terms | Design time |
| Statistical methods: ETS, ARIMA, Theta and combinations | Monthly series with no external variables; these are the baseline | Free libraries |
| Machine learning: combined decision trees | Many similar series with explanatory variables, such as sales by customer or product with price and pipeline data | Weeks of development and clean data |
| Pre-trained deep learning | An additional candidate, without bespoke training, using models such as Amazon’s Chronos | Low cost to use; testing is the cost |
Training neural networks from scratch is rarely justified at this scale because they require large volumes of data. Pre-trained models change the equation: in the authors’ own test, Chronos outperformed local statistical models on series it had never seen. That result must be confirmed on each company’s own series. The rule is the same for every candidate. Each enters a backtest against a seasonal naïve forecast; in production, the best method—or a combination—remains. Combinations won M4; machine learning won M5, which had many series and explanatory variables.
4. Versions and measurement
Each cycle records the baseline and every adjustment made on top of it, including author and rationale; nothing is overwritten. This table makes it possible to measure error by forecast horizon and bias, and to apply the test Michael Gilliland calls forecast value added: every adjustment must improve on the naïve forecast, and an adjustment that does not improve it is removed. There are good reasons to do this.
Across more than 60,000 forecasts from four companies, small adjustments often worsened the result, and upward adjustments were more often wrong in direction, although adjustments helped on average in three of the four companies. Before launch, reconstruct the previous 24 to 36 months of forecasts using only information available on each date. Compare them with the naïve forecast and the old budget. It is the first evidence to show management.
5. Explanation
The decomposition happens in the financial model: between two versions, each driver receives the share of the variance attributable to it, in euros. Materiality rules, set by the CEO and CFO, determine when the information reaches someone—for example, when annual EBITDA moves by more than €100,000 or 2%, or projected cash falls below the minimum within the next six months.
When a rule triggers, the system prepares a small file containing the two versions, drivers ranked by impact, owners’ comments and the name of the decision-maker. An LLM writes a short paragraph from that file. An automated check confirms that every figure in the text exists in the file, and the controller approves it before the paragraph is sent. Leading platforms already sell this as automated variance analysis.
Retrieval-augmented generation (RAG) generates explanations supported by retrieval. It enters the process when an explanation needs documents: contracts, commercial notes, meeting minutes or emails from business units. Those documents are indexed, and retrieval returns source-labelled excerpts that the model uses to provide context. The figures continue to come from the financial model.
In a mid-sized company, one month’s context often fits into a single prompt, and RAG can wait. The monthly commentary costs less than a coffee.
6. Delivery and governance
The CEO receives the paragraph when a rule triggers and one page per month—there may be no need to open a dashboard. The forecast sits outside variable compensation and separately from the target, so that those producing it have no reason to bias it.
The CFO owns the model, the controller runs the monthly cycle, and every driver has someone accountable for it. The model logic itself is versioned, with the date and rationale for every change.
Investment and timeline
For a company with €20 million to €50 million in annual revenue, the first version requires around nine weeks of development. The timeline has four phases: two weeks for system-design decisions; three for data sources, infrastructure and the model; two for method selection and version control; and two for explanation.
Three cycles then run in parallel with the old process before it is switched off. Licences and cloud infrastructure are a small part of the cost structure. The investment is in model design and in whoever leads the conversations with management. A team with a CFO able to design drivers and someone able to connect the data can assemble this architecture on its own.
Limits
This architecture accelerates number collection and forecast production. It does not, however, solve two things: late accounting close and constructing the forecast around the desired result rather than a productive discussion of assumptions. In the latter case, it becomes a negotiation again, however good the model may be.
These are organisational problems, and the surveys suggest they weigh more heavily than technology in large companies. In 2024, 31% of organisations reported conflicts of interest between headquarters and business units, while 13% acknowledged that plans were politically influenced. In 2025, the lack of integrated processes between departments was cited as a barrier twice as often as the lack of a common system: 42% versus 21%.
A mid-sized company, with a finance director and a controller, has fewer people negotiating the number and fewer political layers. It may therefore derive more value from this architecture than some multinationals have derived from an elite platform. It is only a hypothesis, but one worth testing.
Building the architecture above takes the right mix of finance and data skills. If you want a second opinion on your setup, let’s talk.[Get in touch →]
