# AI in Finance: Controlling, Reporting, Forecasts

Author: Marin Niehues, Niehues Consulting · Published: 2025-07-21 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/ai-in-finance/

In finance, AI shortens the work between raw data and decision, most strongly in variance analysis and report preparation, and sharpens forecasts once the data foundation is right. For senior management, this creates a steering task: which workflows of the finance function change first, in what order are they rebuilt, and who is accountable for the result? This article organizes the use cases in controlling and reporting as well as in forecasting, sets out the requirements for audit trail and traceability, and describes an approach for introduction. This article provides an overview and does not constitute legal advice.

## What is AI best suited to in controlling?

In controlling, AI is suited first to variance analysis and report commentary, then to forecasting. A finance function works in a chain. At the start are postings and plan figures from several systems; at the end is a statement on which senior management can rely when making a decision. In between lies a great deal of manual work: consolidating data, explaining variances, drafting commentary. Language models and statistical methods process numbers and text in large volumes. With fixed prompts and approved data sources plus sample checks, quality remains controllable across every close.

According to the German Federal Statistical Office, 20 percent of companies with ten or more employees used AI in 2024, and among companies with 250 or more employees the figure was 48 percent. Three use cases have emerged in the finance function:

- Variance analysis: the system flags conspicuous items and suggests causes, which a controller reviews and assesses.
- Report preparation: commentary and management summaries as well as standard reports are drafted from approved figures.
- Forecasting: statistical and machine learning methods generate predictions and scenarios, which controlling reconciles with the plan.

The close itself comes on top. A faster month-end close, known as fast close, arises when AI prepares account reconciliations and matches open items. How documents are captured automatically is described in the article on [document processing with AI](/blog/ai-document-processing/).

**The benefit of AI in finance arises where recurring analysis and writing work runs on an approved data foundation. The finance function gains time for judgment and for advising management.**

## Controlling and reporting: from variance to cause

In the traditional workflow, a controller exports figures and compares actuals with plan. They then follow up with the business units and finally write the commentary. With AI, this sequence is reversed: the system first delivers a hypothesis about the cause, supports it with the underlying postings, and the controller decides whether it holds. Time shifts from searching to judging. Every statement traces back to a source in the system; otherwise additional review effort arises.

The same logic applies in reporting. A draft of the monthly report is only as reliable as the definitions behind it. As long as two functions understand something different by "contribution margin", even the best model simply produces contradictory reports faster. Every initiative therefore starts with a uniform catalog of metrics with named owners.

The data foundation deserves its own attention, because in the finance function it is spread across several systems. The general ledger sits in the ERP system, plan figures in a planning application, and much additional information in individual functions' spreadsheets. An AI method that draws on these sources inherits every discrepancy between them. The solution is a small number of approved data sources that are authoritative for the close, with a named owner per source. Changes to the chart of accounts or the cost center structure are announced before they take effect, so that models and reports can be adjusted in time. The effort pays off twice: the close becomes more reliable even without AI, and every later application starts on verified data.

**Review questions for each use case**
- Are the source data approved and uniformly defined?
- Who reviews the result, and against what standard?
- Which decision depends on the result, and how quickly must it be taken?
- How much time does each closing cycle take today, and how will it be measured afterwards?

## How does AI improve financial forecasts?

AI methods beat the existing forecast when sufficient history and driver data are available. Two effects tip the balance. Machine learning methods recognize patterns in time series that are lost in a spreadsheet, such as seasonal shifts or the effect of individual drivers. In addition, scenarios can be calculated in a short time, so that senior management can compare assumptions instead of defending a single plan figure. This approach is called predictive forecasting: the forecast is derived from data and drivers, and controlling adds knowledge that is missing from the data.

The international forecasting competitions run by researchers show where the boundary lies. In the M4 competition, with 100,000 time series and 61 methods, combinations of statistical methods and machine learning came out ahead, while pure machine learning methods performed worse. In the M5 competition, with 42,840 retail time series that came with many drivers and long histories, machine learning methods won clearly. For the finance function this yields a clear criterion: with short histories and no driver data, the statistical method remains the benchmark; with long histories and many drivers, the machine learning method pays off.

**Forecasting competitions in research**
100,000 | time series in the M4 competition, compared across 61 forecasting methods
42,840 | retail time series in the M5 competition, in which machine learning methods came out ahead

Whether a forecast is better is decided by back-testing. Every prediction is compared with actuals once the period has ended, measured against the benchmark of the previous method. Only when the new method is more accurate over several periods does it replace the old one. This criterion belongs in the business case before an initiative starts, as described in the article [AI business case: calculating benefits and costs robustly](/blog/ai-business-case-benefits-and-costs/).

## Are AI-generated financial reports audit-proof?

AI-generated financial reports are audit-proof when every figure traces back to a posting and a named person approves every draft, within a documented workflow. In Germany, the principles for the proper keeping of books in electronic form (GoBD) require traceability and immutability as well as process documentation. Statutory auditors examine IT-supported workflows according to the auditing standards of their profession and ask about the internal control system.

**Criteria for the audit trail**
- Traceability: every figure in the report refers to the posting or source from which it comes.
- Log: inputs and outputs are stored for every run together with the model used.
- Approval: a named person reviews and approves before a result takes effect in a report or posting.
- Process documentation: the use of AI is described as part of the closing process.
- Control system: sample checks and reconciliations are defined as controls and carried out.

This sets a clear boundary. AI produces drafts, suggestions and analyses, while posting and approving the report remain with people. Involving the auditor early avoids rework at year-end.

## Containing the risks of AI in controlling

The risks of AI in controlling lie in false statements and in the handling of confidential data, and blind trust in the result comes on top. Language models phrase things plausibly even when the cause of a variance is not in the data. Financial data are confidential and belong only in environments whose data processing is contractually regulated. Then there is habit: whoever accepts suggestions every day gradually checks them less carefully.

Each risk has a countermeasure. Source references and sample checks contain false statements. Approved environments protect confidential data. A fixed review quota that controlling evaluates regularly keeps attention high. A model change by the vendor is treated like a change to the closing process and tested before use.

## AI tools for controlling and finance

AI tools for controlling and finance fall into three categories: assistants in ERP and office software, planning and consolidation software with forecasting functions, and bespoke solutions built on language models. Examples of the first category are Joule from SAP and the finance capabilities of Microsoft Copilot. Planning software increasingly brings machine learning forecasting methods with it. Bespoke solutions come into consideration when metric logic and data sources are company-specific.

The choice follows the business analysis. Whoever wants to automate variance commentary first needs access to the general ledger and plan figures in one environment. Whoever wants to improve forecasts needs history and driver data. A tool that cannot reach the company's own data foundation remains ineffective regardless of its features.

**Forecasting approaches at a glance**
Criterion | Manual forecast | Standard tool without support | With Niehues Consulting
Coordination with business units | High effort per cycle | Mostly outside the tool | Anchored in the planning rhythm
Measuring forecast quality | Rarely systematic | As a metric in the tool | Back-testing against a fixed benchmark per period
Adoption by controlling | Present, but slow | Open | Developed jointly with the team
Transition to operation | Not applicable | Depends on the vendor | Responsibilities and workflows defined

## How do you introduce AI in the finance function?

AI in finance is introduced in three phases: map the workflows, build a prototype in the real close, then define responsibilities and review rules. In the first phase, you map the workflows of the finance function: which steps does the close require, which systems do the data come from, where do waiting times arise? This produces a list of possible applications, ranked by value and feasibility as well as risk. What this ranking looks like is shown in the article [Prioritizing AI use cases](/blog/ai-use-case-prioritisation/).

In the second phase, a prototype is built in the real workflow, for example for the commentary on the variance analysis of one closing month. It runs in parallel with the manual version, and the difference in time and accuracy is measured. A prototype that convinces only in a demonstration proves nothing.

In the third phase, responsibilities and review rules are defined, and the people involved are trained. Who approves the draft? Who maintains the metrics catalog? How does operation continue when a model changes or a data source is altered? Only with these answers does the prototype become part of the close. The full path is described in the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

## Niehues Consulting as a partner for AI in finance

[Niehues Consulting](/) supports initiatives of this kind from a delivery perspective. In AI management, the business case and the sequence of applications are developed, together with portfolio steering by metrics. Prototypes follow that prove the benefit in the real close, and then the rollout through to adoption by controlling and reporting, including the transfer into operation with internal or external teams. In delivery management, Marin Niehues takes on the steering himself on request, so that deadlines and decision paths remain clear through to regular operation.

The approach follows the Niehues Delivery Model with the phases understand, build and implement. Three review questions matter for the finance function: what is the baseline per closing cycle, and how do we measure success? Does the prototype prove the benefit in the real close with real data? Who is accountable for operation from day one? The projects range from VC-funded scale-ups to DAX 40 groups, with assignments in 20+ countries across six continents. The entry point is the free 30-minute initial consultation via the [contact page](/#kontakt).

**Sequence for senior management**
First clarify the definitions and the data foundation, then prove one use case with measurable time savings in the real workflow, and after that define responsibilities and operation. Whoever keeps to this sequence scales to further applications without multiplying the review work.

**Sources**
- Federal Statistical Office of Germany (Destatis): Use of artificial intelligence in enterprises, 2024
- Makridakis, Spiliotis, Assimakopoulos: The M4 Competition: 100,000 time series and 61 forecasting methods, International Journal of Forecasting, 2020
- Makridakis, Spiliotis, Assimakopoulos: M5 accuracy competition: Results, findings, and conclusions, International Journal of Forecasting, 2022

## Frequently asked questions

### What is AI best used for first in controlling?

Variance analysis in the month-end close is the best fit, because the source data are available and the workflow recurs. The time saved can be measured directly. Report commentary follows next.

### Does AI improve forecast accuracy in finance?

Yes, once the data foundation and drivers are right. The proof comes from back-testing over several periods against the previous method. Without this comparison, the improvement remains an assumption.

### What is predictive forecasting?

Predictive forecasting is a forecast calculated by statistical or machine learning methods from history and driver data. Controlling adds knowledge that is missing from the data, such as planned price changes. The advantage lies in speed and scenarios, and back-testing remains the benchmark.

### What AI tools are there for controlling and finance?

There are assistants in ERP and office software such as Joule from SAP or the finance capabilities of Microsoft Copilot, planning software with forecasting functions, and bespoke solutions built on language models. Which tool fits is decided by its access to the company's own data foundation.

### Are AI-generated financial reports audit-proof?

Yes, when every figure traces back to a posting and a named person approves. The workflow belongs in the process documentation. In Germany, the GoBD require traceability and immutability, and the finance function involves the auditor early.

### What are the risks of AI in controlling?

The main risks are plausible but false statements and the leakage of confidential data. Declining scrutiny through habit comes on top. Source references and approved environments contain these risks, and a fixed review quota keeps attention high.

### Who is accountable for results that AI produces in reporting?

Accountability remains with the finance function. A draft is reviewed and approved by a named person, and the system traces every statement back to its source.

### Will AI replace controllers?

AI changes the controller's task. Searching and drafting pass largely to the system. Judgment and advising management remain with controllers and gain in weight.
