AI in controlling: what actually works in 2026

In spring the board asked about the AI strategy for finance. Since then the Copilot licences have been bought, a chatbot pilot is running on the management dashboard, and the controller has a notebook that comments on variances. In the management meeting someone asks the obvious question: "What is EBIT for Region West in August?" The chatbot answers – three times, with three different figures, depending on whether you ask for "EBIT", "operating result" or "result West". The CFO closes the laptop. That is how we usually encounter AI in controlling in 2026: plenty of tooling, little foundation.

AI in controlling: the numbers are more sobering than the demos

Gartner surveyed 204 finance leaders in March 2026. 45 % say their AI investments lean primarily toward productivity and efficiency; only 20 % name decision quality (Gartner 2026). Gartner calls the result a perception gap: finance reports progress on adoption, the board sees limited strategic impact. The machine does faster what was already running. Whether it was the right thing, it does not check.

In the German-speaking market the picture is more down to earth. The CPMC benchmarking study on digitalisation in controlling, commissioned by the International Association of Controllers (ICV), interviewed eighteen companies in 2024: four use AI in controlling at least marginally, mainly for revenue forecasting; nine use RPA, and manual work still dominates planning and reporting (Dillenberger/Spang 2025). The BARC Planning Survey 26, with 804 respondents, shows movement: the share of companies using AI for predictive planning or forecasting doubled within a year, from 11 to 27 %. What do they expect from it? 75 % name relief from manual work, 52 % validation of manual plans, 51 % higher forecast accuracy (BARC 2026). Again: relief first, assurance second, better numbers last.

What does the opening scene cost? First, licences that run per head and month while the pilot is no longer opened in the meeting. Second, trust: once a tool returns the same figure three different ways, the management team is back to negotiating the number instead of the decision. Third, and most expensive: a wrong figure nobody checked because it was fluently phrased. Fabian Meisser puts it plainly in his article for SwissAccounting: KPIs must not rely on approximate estimates from AI, but on exact values according to the company's definition (Meisser 2024, p. 34). The chatbot did not know the definition because it was written down nowhere. That is not an AI problem but a model problem that AI has made visible.

Our position follows from that: AI is only as good as the question you ask it and the data model it stands on. Both belong to controlling – not to IT, and not to the vendor.

The test grid: value, constant rules, validation

Meisser proposes an assessment matrix with two axes: value creation – what value or time saving does the use case bring, in francs? – and regularity – do the rules of the game stay constant, so that the past is an indicator of the future? (Meisser 2024, pp. 32 f.). In our reviews we have added a third question, because in practice it decides most cases: who validates the result before it goes into a meeting? If the honest answer is "nobody" or "the end user", the use case is not ready, however good the demo was.

With these three questions, the AI ideas clients bring us split into two groups.

AI use cases in controlling against the test grid

Use case	Value in CHF	Rules constant	Who validates	Verdict
Draft variance commentary in the monthly report	high: 1–2 days per month	yes: report structure fixed	controller signs off	holds up
Flag anomalies in postings and budgets	high: errors caught before the close	yes: account logic stable	finance professional classifies	holds up
Statistical forecast as reference next to the plan	medium	only in a stable business	controller compares with plan	holds up with caveats
Structure interview notes and KPI glossary	medium: days in scoping	yes	author checks every entry	holds up
Copilot on a documented semantic model	medium to high	yes, if model is maintained	model delivers figure, human reads	holds up with precondition
AI calculates KPIs from accounts	negative	–	nobody in the meeting	does not hold up
Agent without data foundation	unclear	no	not possible	does not hold up
Data agent for end users on raw tables	low	no	end user cannot	does not hold up
"AI replaces the controller"	–	–	–	does not hold up

What holds up

The first draft of the variance commentary. The monthly report is built, variances against budget and prior year are calculated, the ±5 % threshold has flagged what needs a comment. A language model that knows the variance table, the last three months of comments and the posting rules drafts every flagged position: what, how much, against which reference, with which presumed cause. The controller cuts, corrects, adds the cause only he knows, and signs off. With thirty commented positions that saves one to two days a month – and it holds up because the structure is constant and a human validates.

Anomalies in postings and budgets. A supplier that suddenly lands on a different account; a budget line ten times its history; a provision account with no movement in the closing month. A machine finds such patterns in large posting tables more reliably than a person on closing day. But whether the outlier is plausible or a problem, only the finance professional can judge (Meisser 2024, p. 33). The use case holds up because it secures the close, not because it replaces it.

The statistical forecast as a reference. A time-series model projecting three years of revenue history is not a plan but a second opinion: if the head of sales plans plus 12 % for the fourth quarter and the projection says plus 3 %, the difference is the question for the meeting. That holds up in a stable business with seasonality; in a project business with five large orders a year the rules are not constant. And the forecast has to be transparent, or users will not accept it (Meisser 2024, p. 33).

Structuring interview notes and the glossary. In scoping we talk to every executive: objective, problem, early warning, analysis, report, effect. Turning twelve sets of notes into a decision inventory and a first KPI catalogue used to take days; today the model provides the structure and we check every entry. The same goes for the KPI glossary – the definition the chatbot in the opening scene was missing.

Copilot on a clean semantic model. Copilot in Power BI does not calculate on its own; it translates the question into a query against the semantic model. If "EBIT" is defined there exactly once, as a measure with description, synonyms ("operating result") and documented dimensions, the meeting gets a figure – the same one as in the report. If the model is a collection of tables with technical column names, Copilot guesses, and it guesses fluently. Dimensions and measures cleanly named, relationships unambiguous, notation fixed: that is the precondition, and it is controlling work.

What does not hold up

AI that calculates KPIs. A language model handed the chart of accounts and a trial balance that "calculates" EBIT delivers an approximation nobody in the meeting can check. KPIs are deterministic and company-defined; the right tool for them is BI, not AI – the two are symbiotic, not competing (Meisser 2024, p. 34, fn. 3). Our rule: the figure comes from the model; AI may comment on it, explain it, put it in context. It may not invent it.

Agents without a data foundation. 2026 is the year of agents: tools that plan steps on their own, fetch data and assemble results. On a documented model that can work. On eight source systems with three different customer numbers and an Excel file as the consolidation layer, the agent executes the same confusion the human had before – only faster and without the nagging feeling.

Data agents for end users. The idea that a division head asks the agent directly and takes the answer straight into the meeting fails the third test question: the end user cannot validate whether the agent took revenue gross or net, with or without intercompany, by invoice or delivery date. Self-service on a verified model with predefined measures: yes. Free questions on raw tables: no.

"AI replaces the controller." Christian Langmann put it soberly back in 2019: controlling will keep the authority to interpret, but only if it understands the data analytics the machines apply (Langmann 2019, p. 53). What gets replaced is not the controller, but the controller who cannot ask the question and cannot check the result.

The sequence stays: decision, KPI, model, then the machine

Our framework "From decision to KPI" does not change with AI. Decision → question → KPI → data → report → effect → review – the machine only enters at "data", once decision, question and KPI are defined. Before that stands the purpose question: what are these numbers supposed to say, and what are they for? Whoever cannot answer it for the chatbot cannot answer it for the report.

Sometimes the best "AI idea" needs no AI at all. At an industrial company the question was which procurement risks should be visible in the division cockpit. The first reflex was a model predicting risk from lead times and stock levels. The rule that held up was simpler: a cheap part that sits in many bills of material and belongs to the fast movers is a risk – because when it fails it stops the most expensive products, and nobody looks at it because it is cheap. Three attributes, one sort order, one KPI with a weekly cadence. The rule is constant, purchasing understands it, and it reconciles. An AI might have found the same thing, but nobody would have known why.

That is the yardstick for every AI use case: would a constant rule in the model be simpler, more traceable and easier to reconcile? Then it belongs in the model. Reconcilability before beauty applies here too – the report hangs on the close, and the AI answer has to match it.

Meisser sees the greatest benefit not in forecasting or dashboards, but with controllers who understand their data and build automations themselves with Copilot or ChatGPT – two to five times faster than before. His line: the winners of this development are not today's fastest programmers, but those who understand the business problem while being less versed in coding (Meisser 2024, p. 33). That is the role we expect of controlling: not the recipient of AI output, but the person who builds the model the machine works on.

What the project review checks

With an AI initiative in controlling we look first not at the initiative but at the model underneath. Three checkpoints: first, model maturity – is every steering KPI defined exactly once, with owner, reference and cadence, and does the report reconcile with the close? Second, the question – which decision should come out differently because of the AI result, and who validates it beforehand? Third, notation – is there a concept in line with IBCS® that fixes how variances, references and thresholds look, so that a generated comment speaks the same language as the report? Where the three are in place, AI is a building block of a few weeks. Where they are missing, AI is the most expensive symptom of a model problem – and that is what we write in the second opinion of the project review, with a cut list of what comes before the AI.

What you can do tomorrow

Hold the AI ideas circulating in your company against the three questions: value in francs, constant rules, who validates. Whatever gets "unclear" twice is parked. Then ask your dashboard chatbot the three questions from the opening scene and compare the answers with the monthly report. If they do not match, your next project is not an AI project but a model project.

If you would like an outside view: our project review takes a few days, comes at a fixed price and ends with a written second opinion – which AI use cases hold up, what the model needs first, and what can be cut.

Sources

• Gartner (2026): Gartner Survey Shows 45% of CFOs Say Their AI Investments Lean Toward Productivity, While 20% Say These Investments Lean Toward Decision Quality. Press release, 20 July 2026, survey March 2026, n = 204. gartner.com
• BARC (2026): The Planning Survey 26 – AI use in corporate planning doubles within a year. Press release, 9 June 2026, n = 804. barc.com
• Dillenberger, K. L., Spang, S. (2025): Digitalisierung im Controlling. Benchmarking study, results report (excerpt). Centre for Performance Management & Controlling (CPMC), Frankfurt School, commissioned by the International Association of Controllers (ICV). icv-controlling.com
• Meisser, F. (2024): KI im Rechnungswesen und Controlling: Was sind sinnvolle Anwendungen? In: SwissAccounting «Standard» 3/2024, pp. 32–35.
• Langmann, C. (2019): Digitalisierung im Controlling. Springer Gabler essentials.