When Artificial Intelligence Automates Reporting: The Management Controller at the Heart of Trade-Offs

The role of management controller was built around one essential mission: helping an organization understand its results and use its resources in line with its objectives. Preparing budgets, analyzing costs, monitoring margins, forecasting and working with operational managers form the foundations of this activity. Reporting, meaning the regular production of information used to steer the business, is only one component of it. Interpreting results and proposing corrective measures are already part of the job, as the Apec job profile points out¹.

Artificial intelligence therefore does not create the analytical dimension of management control. It introduces new possibilities in data preparation, the detection of variances and the drafting of the analyses that feed this function. The question then becomes less about how many dashboards are produced than about how reliable and useful they are for decision-makers.

Difficulties in accessing actionable information remain significant. An international survey published in May 2025 by Deloitte and the Institute of Management Accountants, covering more than 440 finance and accounting leaders, indicates that 54% of respondents lack the reporting they need or believe they must improve the transparency of their cost and profitability information. This finding describes a need reported by professionals, not proof that AI alone will solve it².

This is where the transformation of the profession lies: automating what can be automated, while strengthening the ability to understand what the figures actually make it possible to assert. A generated commentary is not necessarily a demonstrated explanation, and a recommendation is not a decision.

A preliminary distinction is essential. Adding up entries, updating a dashboard or calculating a budget variance does not necessarily require artificial intelligence. These operations can rely on conventional calculation rules.

The specific contribution of AI lies notably in recognizing unusual patterns, assisting with certain analyses and generating text from the information provided. Generative AI thus produces content, but its convincing wording does not guarantee its accuracy, a limitation highlighted by the CNIL³.

Several documented uses illustrate this shift.

These features show that automation can go beyond simply formatting reports. On their own, however, they do not constitute independent proof of productivity gains or better decision-making in every company.

For any deployment, the relevant question therefore remains a practical one: how much time is actually saved once data preparation, verification of results and correction of commentaries are taken into account?

MSc AI & Data for Finance & Accounting

The Bac+6 in AI & Data dedicated to careers in Finance, Accounting, Audit and Management Control. No programming prerequisites.

12 months — Bac+6 Entry with a Bac+5 Work-study, initial or continuing education Paris-Villejuif campus

Discover the program → RNCP Level 7 certification

The transformation does not mean turning the management controller into a passive supervisor of algorithms. On the contrary, it can strengthen a responsibility that is already central: making sense of results and making assumptions open to discussion. Working with operational teams and formulating improvement proposals are already within the scope of the role¹.

Financial skills remain the foundation of the profession. Understanding how costs are built up, how profitability works and what happens when an assumption changes remains essential for assessing an automatically generated result.

Nevertheless, professionals recognize the need for training. A CIMA survey published in December 2025 of 1,446 finance and accounting executives and managers indicates that 56% of respondents see generative AI as their main skills gap. This survey covers finance and accounting functions broadly, not only French management controllers⁷.

It can help, but reliability must be assessed at several levels: the accuracy of the figures, the relevance of their interpretation and the robustness of the forecasts. An improvement at one level does not guarantee the other two.

What the evidence allows us to say

A fictional teaching example helps clarify this distinction.

A company plans to sell 1,000 units at €100, for revenue of €100,000. It ends up selling 1,200 units at €90, or €108,000. Revenue therefore exceeds the budget by 8%.

With a constant variable unit cost of €60, however, the contribution margin falls from €40,000 budgeted to €36,000 actual, a decrease of 10%. Here, this margin corresponds to revenue minus variable costs, without taking fixed costs into account.

A commentary limited to “volume growth is driving strong performance” would be incomplete. The calculations, which can be done without AI, show that the rise in revenue is accompanied by a decline in this margin. However, they do not support the conclusion that the commercial policy is necessarily bad: the company may have accepted lower short-term profitability to achieve an explicit objective.

The controller must then verify this objective, its expected effects and the conditions under which the decision will be reassessed. The figure describes a consequence; it does not, on its own, settle whether the trade-off is sound.

The same caution applies to forecasts. A model that reproduces past data very well may perform less well on new periods. The evaluation must use data that were not used to fit the model and compare its results with a benchmark method, over a relevant forecast horizon. The methodological work of Hyndman and Athanasopoulos highlights precisely this distinction between goodness of fit and actual forecast accuracy⁹.

Limitations to watch for

Several risks also require particular attention. A generative model can produce a plausible but unsupported justification. An analysis may overlook information missing from the file provided. Excessive confidence in the answer can reduce the necessary checks, as the CNIL points out³.

To assess a system, it is therefore preferable to track concrete criteria: rate of incorrect reconciliations, number of corrected commentaries, traceability of figures, forecast accuracy and total review time. The goal is not to produce more analyses, but to obtain usable analyses without introducing a disproportionate verification burden.

The most plausible evolution is not one of uniformly autonomous management control. It will depend on data quality, how work is organized and which decisions the company is willing to delegate to tools.

• ✓ More assisted file preparation: systems will be able to aggregate information, flag inconsistencies and prepare summaries. The time saved will only translate into better advice if the organization actually reallocates that time to analysis and to engaging with the client.
• ✓ More continuous risk monitoring: fraud, transaction anomalies and changes in certain indicators can be tracked more frequently. This monitoring also increases the volume of alerts and therefore requires prioritization and control mechanisms.
• ✓ Generative assistants playing a bigger role in client relations: they will be able to facilitate document searches or prepare responses. Deploying them will require well-managed knowledge bases, confidentiality controls and a precise definition of the situations in which a human must take back control.
• ✓ Agentic systems within controlled scopes: some assistants will be able to chain several tasks together. In sensitive processes, the key question will be the permissions granted to the system: reading, proposing, transmitting, modifying or executing are not equivalent levels of autonomy.
• ✓ More structured AI governance: the ECB has included strategy, governance and AI-related risk management among its supervisory topics. In particular, it plans to continue targeted monitoring of banks’ generative AI applications.⁸
• ✓ A greater need for regulatory understanding: bankers will have to work in an environment where banking regulation, data protection, digital resilience and the AI Act overlap. Not all AI systems will be subject to the same requirements; the use case, purpose and level of impact will remain decisive.

Freed-up time will therefore only become analysis time if the company organizes this reallocation. The transformation of the profession remains as much a matter of management and training as of technology.

Automating reporting does not change the purpose of management control: making performance understandable and helping the organization act with full knowledge of the facts. It does, however, shift part of the focus toward validating data, discussing assumptions and justifying recommendations.

This responsibility goes beyond the mere pursuit of immediate cost reductions. A trade-off can improve a financial indicator while undermining the company’s quality, lead times, skills or future capabilities. The controller’s role also involves making these effects visible and spelling out what the dashboard does not measure.

The value of a tool must therefore be assessed based on the problem it solves, its total cost and the resources it requires. Conventional automation may be sufficient when a calculation rule meets the need. A more complex model is only relevant if its added value has been demonstrated.