DE

The 5x Controller: What AI Really Changes in Controlling (2026)

By Martin Faust, Head of AI, Valsight

A 5x controller is not a controller who does the same work five times faster. No study shows that: in a large Danish survey, people who use AI at work report saving 2.8 percent of their working hours on average. The real lever for AI in controlling and FP&A lies elsewhere, in the roughly 45 percent of time that goes into collecting and validating data. And it only works if the AI does not hand out numbers but writes formulas that can be checked.

Where the 10x engineer comes from

Software has its “10x engineer”, sales has the rep who sells three times as much with AI. Both images contain a truth, just not the one the name suggests.

The 10x figure goes back to a 1968 study by Sackman, Erikson and Grant. It compared a handful of programmers and found huge differences between the single best and the single worst. Lutz Prechelt reanalysed the data in 1999 and called the legend “misleading”: between the fastest and the slowest quarter the factor is closer to 5. The number describes differences between people, not the effect of a tool.

What is more interesting is where the time goes. According to Atlassian’s State of Developer Experience 2025, developers spend only 16 percent of their time writing code. According to Salesforce’s State of Sales 2023, sales reps spend only 28 percent of their week actually selling. The rest is searching, aligning and administration. Where AI does deliver a multiplier, it shrinks that rest, not the core work.

Even there, caution is due. In a randomised study by METR (2025), experienced developers were 19 percent slower with AI, yet afterwards believed they had been about 20 percent faster. Perceived and measured productivity diverge.

Where controllers spend their time

Two solid sources exist for controlling, and they count differently.

Source Sample Finding
FP&A Trends Survey 2024 383 FP&A professionals, international 45% collecting and validating data, 35% generating insights and driving actions, the rest producing reports. Stable for four years
FP&A Trends Survey 2025 international Still 46% on data collection and validation
WHU Controller Panel Controllers and CFOs in Germany, Austria and Switzerland Reporting is the largest block at 21 to 23%, management advisory sits at around 14%. The split “has hardly changed between 2011 and 2020”
WHU Controller Panel by company size The annual budget takes eight weeks on average in smaller companies and sixteen in large ones

Two things stand out. First, almost half the time goes into groundwork before anyone thinks about the numbers. Second, a decade of digitalisation has barely moved that split. New tools alone did not free up the time. An honest caveat: there is no separate data-versus-analysis split for mid-sized companies in the German-speaking region. The 45 percent comes from international samples with many large corporations.

What studies show about AI in knowledge work

The evidence is more mixed than vendor promises suggest.

For modelling, forecasting and variance analysis there is no controlled study yet. Anyone promising a factor there is guessing.

The five levers: using AI in controlling

The 5 in “5x controller” does not stand for speed here but for five levers, one per block of time. Each comes with a method, a state of evidence and a first step.

Lever Time block What the AI does Evidence
1. Data preparation and reconciliation about 45% Writes reusable scripts and queries plausible
2. Reporting and commentary about 21 to 23% Drafts commentary, the controller checks the facts proven for writing
3. Planning and driver models budgeting and planning Builds the driver tree and an Excel model with formulas not studied, can be shown live
4. Analysis and scenarios part of the 35% Answers questions through the model, not from memory risky without method
5. Context and knowledge across everything Keeps the business model, assumptions and agreements foundation for 1 to 4

Lever 1: Data preparation and reconciliation

The largest block of time is also the one where a chat window helps least. Paste data into a chat and you get a one-off result nobody can trace. Better: the AI writes a script or a Power Query that handles the import and the reconciliation. The script runs again next month, the same way, and can be read. First step: take a reconciliation you do by hand every month and ask for a query instead of a result.

Lever 2: Reporting and commentary

This is where the evidence is strongest: drafts of variance commentary and management summaries come together much faster. Responsibility for every number in the text stays with the controller, because language models phrase wrong statements just as convincingly. First step: have the commentary for the monthly report drafted, then check every figure against the reporting.

Lever 3: Planning and driver models

This lever shows whether AI holds up in controlling. An example from our webinar: Tavenor Systemtechnik, a fictional special-purpose machinery maker with 820 employees and €178m revenue in 2025, in the middle of its 2027 planning round. Management wants €191m in revenue and says: “I don’t want to hear that it can’t be done. I want to hear what has to be true for it.”

Instead of a question in a chat window, there is a project: a folder with the annual report, the actuals, the notes from the planning kick-off and one page of working instructions. The instructions fix the method: driver tree first, numbers second. Actuals are reproduced exactly, every assumption gets a source, nothing is calculated in the text. Then come three instructions to an AI coding tool:

  1. Understand: “Read the context and propose a driver tree for revenue and EBIT. No numbers yet.”
  2. Build: “Turn it into an Excel model with formulas: actuals 2024 and 2025, plan 2026 and 2027. Check that the actuals are matched exactly.”
  3. Ask: “What has to be true for us to reach €191m in 2027? Base case plus the measures from the kick-off, as a one-page report.”

The output is a driver tree (order intake via order backlog to new-machine revenue, installed base times service penetration to service revenue), an Excel model in which every number is a formula, and a one-page report.

The key sentence: the language model does not calculate. It writes what calculates. You describe the logic, the AI writes the formula the way you would write an Excel formula yourself, and the formula calculates the same way every time. This is more than a matter of style. Language models are unreliable at arithmetic in free text: in one study (Dziri et al., NeurIPS 2023), GPT-4 got only 59 percent of three-digit by three-digit multiplications right. Let the model write code that does the calculating instead, and accuracy on benchmarks with arithmetic and financial tasks rises by around 12 percent on average (Chen et al., “Program of Thoughts”, 2023).

That does not make it error-free. Even the best models fully solve only 34.9 percent of tasks in a recent benchmark built from real business workbooks (SpreadsheetBench 2, 2026). That is why the reconciliation with actuals belongs in the model and the controller’s review belongs at the end. First step: set up a folder with context for one planning topic and ask only for a driver tree first.

Lever 4: Analysis and scenarios

With a model made of formulas, every question becomes a change to a driver: what happens if service penetration rises by two points? The answer comes from the model and can be recalculated. Without a model, the AI answers the same question from memory, and that is exactly where the BCG consultants were 19 percentage points off. First step: ask scenario questions only of a model, never of an empty chat window.

Lever 5: Context and knowledge

The chat window forgets everything. The project folder does not: the business model, assumptions, sources and agreements stay in place, every question builds on them, and a colleague can look up why an assumption was set the way it was. On its own this lever saves little time, but it is what makes the other four repeatable. First step: write one page describing your business model and your planning logic. It is the most important file in the folder.

The path to the 5x controller: four stages

Stage Way of working What is missing
1. Chat window Individual questions, results to copy Context, repeatability
2. Project with context Folder with documents and working instructions Verifiable calculation
3. AI writes formulas and code Models and queries that do the calculating Collaboration in the team
4. Shared model Several teams, versions, approvals -

Most controllers who use AI are at stage 1. The big jump is between stages 2 and 3. At stage 4 the question changes: as soon as several teams feed their assumptions into the same model, the traceability that stage 3 brought is lost without software. Who changed what, which version is valid? That is where planning software such as Valsight comes in.

The trap: speed without traceability

In controlling, a number is only worth something once someone can stand behind it. Whoever signs off the plan, whether management, an advisory board or a bank, has to be able to recalculate it. In our conversations with finance leaders at mid-sized companies, four out of five raised this problem unprompted, even though we had not asked about AI.

A fast forecast that nobody can explain is therefore not a productivity gain but a risk. The research backs this up: AI helps with writing but makes analyses of spreadsheets and notes worse when people follow it unchecked. And feeling faster is not the same as being faster.

What 5x can realistically mean

A worked example, not a measurement: today time splits roughly into 45 percent data, 20 percent reports and 35 percent insights. If scripts halve the data work and drafts save a third of the reporting time, almost two thirds of the time is left for analysis and advice instead of just over a third. That is not five times the speed, but it is a different role.

Where a real factor appears, it is in individual tasks: a first draft of a driver model in one session instead of days. Whether that holds in your day-to-day work only a measurement can tell. Record how long a forecast, a monthly commentary or a budget model takes today, and compare afterwards. Feeling faster is not enough, as the METR study showed.

The coming years will depend less on which model is ahead at the moment than on whether controlling teams have a method: context instead of chat, drivers instead of trends, formulas instead of answers.

Frequently asked questions

Do I need to know how to code to build a driver model with AI?

No. You describe the logic in words, the AI writes formulas and code. You still have to check the result, the same way you would check a colleague’s Excel file.

Which AI tool do I need?

An AI coding tool that works with a project folder, for example Claude Code, Codex or Cowork. These usually require a subscription. The tool is interchangeable, the method is not.

Which data may I give to an AI?

Settle that with your IT and data protection team before real company figures go into any tool. Published financial statements or an anonymised example are good for first attempts.

Does a self-built model replace planning software?

For individual analyses and first models, often yes. Once several teams contribute assumptions, versions get approved and the model has to run across the company, you need software that keeps track of changes, versions and ownership.

How do actuals get into the model?

In the example, as an Excel export and the annual report in the project folder. In daily work it pays to automate the export once with a script (lever 1) instead of copying it again every month.

Sources