10 Golden Rules of Financial Modeling – Revisited in the Age of AI and Cloud Collaboration

Although the insights gained from financial models are often used as the basis for major business decisions, there are no universally applicable financial modeling guidelines that can guarantee a uniform quality standard and comparability. At RÖDL, we have long addressed and maintained the resulting heterogeneity through our 10 Golden Rules of Financial Modeling.

Since the first edition of these rules, two core concepts have changed so fundamentally that we see a need for an update: the notion of “a single modeler” who builds a self-contained file locally and sequentially.

1. AI-Assisted Modeling
   AI assistants generate formulas, structures, and entire calculation logics in seconds, suggest scenarios, and provide feedback on existing models.

2. Cloud-based, simultaneous collaboration.
   Models are stored in SharePoint or OneDrive, are edited by multiple people simultaneously (co-authoring), and increasingly draw their data from centrally managed data lakes rather than manually frozen copies.

The obvious conclusion would be that traditional modeling rules are thus losing their importance. Our experience shows the opposite – for three reasons, each of which relates to those recent developments:

1. Structure enables machine readability and collision avoidance.
   A consistently structured, clearly labeled model can be reliably understood and updated by an AI assistant – and multiple people can edit it simultaneously in co-authoring mode without overwriting each other’s changes. The same discipline serves both purposes.

2. The error profile is shifting.
   AI generates results that look plausible but are nonetheless incorrect; in co-authoring, changes that do not originate from oneself are also constantly being made. Control mechanisms and an independent review thus become a safety net that must withstand increasingly complex requirements.

3. Responsibility remains with humans.
   Neither a language model nor a large number of simultaneous contributors relieves the person responsible for the model of the duty to understand its results and take professional responsibility for them.

Against this backdrop, the following ten rules, with corresponding additions, continue to apply. Adherence to these rules leads to a reduction in error susceptibility, efficiency gains in model creation and testing, increased user convenience, and greater acceptance among the model’s intended users – and it creates the conditions necessary for both AI support and collaborative work in the cloud to be used safely and with added value.

1) Stringent workbook design

To maintain clarity and ensure a consistent data flow, a model’s worksheets should be grouped into the sections “Raw Data,” “Assumptions,” “Analysis,” and “Output” and color-coded to distinguish them from one another.

• The “Raw Data” section contains the imported source data, which is linked to within the model. This ensures that the data source remains traceable at all times.
• The “Assumptions” section collects the editable premises and serves as a “model cockpit” for scenario and sensitivity analyses.
• The “Analysis” section is where the actual calculations take place; these calculations draw data exclusively from the “Raw Data” and “Assumptions” sections. No direct data entry takes place here.
• The “Output” section presents the key results in a format tailored to the intended audience and – where helpful – in graphical form.

This four-section separation is also the most effective way to make a model readable for an AI assistant. Clear sheet names, labeled rows, and unambiguous units serve as documentation that the AI can immediately use for analysis and expansion tasks.

This same section structure forms the basis for conflict-free collaboration. Clearly delineated sections allow multiple people to work simultaneously on different parts of the model. The previous recommendation to map the entire model in *a single* file and strictly avoid external links must now be reviewed: What matters is not “the single file,” but rather the traceable and centrally managed data source. Data connections via Power Query to a centrally stored SharePoint source often fulfill this goal better in a world of co-authoring and shared data sources than a manually frozen copy – provided that the source, update logic, and transformation steps are documented.

2) Comprehensive assumption sheets

The premise sheet should be organized in a clear and logical manner. It is recommended to define top-level sections (e.g., “I. General Assumptions,” “II. Income Statement Forecast,” “III. Balance Sheet Forecast”) with corresponding subsections that can be shown or hidden as needed using the grouping function.

Variable assumptions that are modified for scenario and sensitivity analyses should be highlighted in color and self-explanatory (if necessary, with instructions such as “Enter expenses as negative values”). Assumptions from the “Raw Data” section are directly referenced so that the data source remains traceable at all times. Assumptions whose sources are not available in Excel (such as PDF files) are entered directly; the source and underlying assumptions are documented in a separate comment column.

A well-maintained, labeled premises sheet is the ideal starting point for AI-supported sensitivity and scenario analyses, because the AI can immediately identify the model’s control variables. The comment column also gains additional importance: It records *where* a premise comes from and whether it was researched with AI support – and is therefore subject to validation (see Rule 12).

The premises sheet is also the “single source of truth” for all co-authors – there should therefore be exactly one authoritative location per premise. Anyone who sets or changes a premise can use the thread-based comments in modern Excel versions – including @mentions – to link directly to the person responsible; this replaces the static comment column previously used for coordination purposes and makes discussions about individual assumptions traceable.

3) KISS principle (Keep it smart and simple)

A financial model should be as detailed as necessary and as simple as possible. The data flow should correspond to natural reading habits: time series from left (history) to right (forecast), calculations from top (assumptions) to bottom (results). Excessively long formulas and deeply nested “IF functions” should be avoided and broken down into comprehensible intermediate steps. Intermediate and final results should be calculated only once and then linked to avoid redundancies and errors.

AI reduces the cost of complexity – it effortlessly generates nested formulas and elegant code constructs. That is precisely why KISS is even more important today. However, the focus is shifting: it’s less about foregoing technical sophistication and more about auditability – every calculation must be verifiable by a human, and in case of doubt, by an independent auditor as well as a co-author, without disproportionate effort. Anything that no one checks is potentially dangerous, no matter how elegant it may seem.

The following applies: While AI increases the complexity that any individual can incorporate into models – dynamic arrays and INDIRECT aggregations can now be implemented without extensive background knowledge – it also enhances the auditability of that very complexity. Modern financial modeling should make use of both aspects.

Macros (VBA) should also continue to be used with great caution, as they are difficult for third parties without the appropriate knowledge to understand. Today, more transparent alternatives are available: `LET` and `LAMBDA` for reusable, documented logic within the worksheet, dynamic arrays for rule-based calculations, as well as Office Scripts and “Python in Excel” for more complex tasks.

4) Column consistency (uniform timeline)

When modeling across multiple time periods, strict attention must be paid to maintaining a consistent timeline across worksheets. The timeline is defined once in the assumptions sheet and linked to in all other sheets; by “pinning” it to the top of the sheet, it remains visible even in lower sections. A few columns are reserved for comments between the row labels and the start of calculations.

Column consistency significantly reduces the likelihood of errors and greatly simplifies model review, as linking errors between periods are more easily detected when consistency is broken.

A consistently uniform timeline is one of the most effective structures for enabling AI-assisted reviews: An assistant can detect period breaks, misaligned links, and inconsistencies across sheets much more reliably when each period is located in the same position – the same applies to a human reviewer who is consolidating contributions from multiple authors.

5) Row Consistency (one formula per row)

By properly fixing rows and columns (using dollar signs), the same formula should be used for all periods within a single row. Traditionally, the formula was written for the first period and then filled in for subsequent periods via copy and paste; the reviewer would then “simply” check the first period and could quickly rule out formula breaks.

Dynamic arrays or overflow ranges take this principle to a new level—a single formula fills an entire range, so that a formula break within a row cannot technically occur in the first place. Where this is not possible, the tried-and-true principle still applies. However, strict adherence to row consistency can lead to increased complexity if too many contingencies are modeled in a single formula; in such cases, the KISS principle (Rule 3) should be considered, and, if necessary, intermediate calculations or “flags” (Rule 6) should be used. Where formula breaks are unavoidable, they should at least be highlighted in color.

6) Use flags to improve transparency and efficiency

“Flags” refer to binary auxiliary values (1 or 0). By multiplying by the flag row, individual factors can be “activated” (multiplied by 1) or “deactivated” (multiplied by 0) along the time axis—for example, in business valuations to distinguish between the detailed planning period (“0”) and perpetual annuity (“1”).

Through the deliberate use of flags, complex “if” functions can be partially avoided, and row consistency (Rule 5) can be maintained while adhering to the KISS principle (Rule 3). Flags also increase transparency regarding when specific events occur and allow for variability across periods.

7) Implement robust control checks

Model creators and model users are often different people. To prevent erroneous changes, all calculation formulas should be set to be uneditable by the user; only the premise fields in the assumptions sheet may be modified.

To quickly identify errors, the model should include “check rows.” A distinction is made between technical checks (e.g., verifying that total assets equal total liabilities) and source checks (e.g., verifying that the net income in the model calculation matches the net income in the source file). The check rows are inserted at the bottom of the worksheets and linked to a shared worksheet (“Master Check”), which provides an overview of all checks and immediately displays any model errors along with their source.

This rule is now becoming increasingly central to our daily work. AI generates plausible-looking errors that elude visual plausibility checks. A balanced network of meaningful check rows and a comprehensive Master Check are the most effective means of catching AI-induced errors: Whenever an AI-supported change is made to the model, the Master Check must be reviewed first – if a check row is triggered, the change is incorrect or incomplete. In cases of co-authorship, changes that do not originate from oneself are constantly being made. The master check thus becomes a shared early-warning system for the entire team – it immediately indicates when another author’s intervention has made the model inconsistent.

There is also a growing focus on “big-picture” plausibility checks, which are typically qualitative in nature: Outputs must be structured in such a way that trends can be explained clearly and comprehensibly by their creators.

8) Ensure user friendliness

Layout and user-friendliness are particularly important for “standard models” used by multiple users for comparable scenarios (e.g., a standardized valuation model). Users often do not know the details of the model; they simply need to be able to operate it and understand the results.

A “model menu” has proven effective: a separate worksheet with clear links to the relevant sheets, supplemented by a backlink to the menu in each worksheet (e.g., in cell A1). Similar to a website or app, the entire model is controlled via the menu. A clear menu and consistent operational logic help not only human users but also agent-based AI tools to navigate the model. However, as models are increasingly operated via natural language, this does not replace good labeling and structure – it requires them.

If a standard model is used by many users via SharePoint, it’s important to note that not all Excel features are available identically in Excel on the web (e.g., certain macros). Robust, platform-independent user interface designs are therefore preferable; access to individual areas can be controlled via permissions.

9) Tailor outputs to the audience

One of the greatest challenges is summarizing the model’s complexity for the target audience by focusing on key metrics and insights. Often, individual output sheets must be created for different audiences (e.g., a management summary for executive leadership, supporting information for the controlling department).

Meaningful outputs include tabular and graphical analyses; they are self-explanatory and include notes (currencies, dates) as well as interpretations of the results. As far as possible, they should be formatted for presentation and printing in accordance with the corporate design. If selected assumptions are stored variably on the output sheet, the impact on results can be tracked in real time.

Instead of sending file copies via email, outputs can be made available as a single, up-to-date result shared via SharePoint (“Single Source of Truth”). This avoids the classic source of error – outdated versions circulating in parallel—and ensures that all recipients have access to the same version.

AI can significantly accelerate the creation of outputs tailored to the audience – from consolidating key metrics to drafting an interpretation of the results. However, the professional responsibility for the selection, weighting, and interpretation of the presented metrics remains with the creator; an AI draft is a starting point, not the final result.

10) Apply dual review and model validation

In practice, errors often creep into the creation of financial models, which can lead to incorrect results and significant financial consequences (e.g., excessively high purchase prices in corporate transactions). A structured model review—ideally conducted by someone who did not participate in the model’s creation (the “four-eyes principle”) – can remedy this. The review is divided into three levels:

• Technical review: Are all links set correctly and are the formulas error-free? This is supported by specialized review software (e.g., Operis Analysis Kit) as well as, increasingly, by AI-powered audit tools.
• Structural review: Is the model architecture rigorous and the data flow consistent?
• Logical review: Are the calculation logic and underlying assumptions appropriate?

The technical review can be partially automated and accelerated by AI. The logical review – do the assumptions hold up, and is the business logic sound? – remains a human task and, with AI, an even more challenging one, because an AI model can convincingly formulate logic that is factually incorrect. Two principles are essential here: First, the “second pair of eyes” performing the review must not be the same AI that created the model. Second, AI-assisted review does not replace independent human review – it complements it.

Three New Rules for the Modern Modeling Environment

The ten rules above pertain to the model itself. AI and collaborative cloud-based work add a second dimension that concerns the process—and is indispensable, especially in a professional setting.

11) Governance and Confidentiality

Before a model is shared in the cloud or made accessible to an AI tool, it must be clarified which data goes where and who is authorized to access what. Client, deal, and personal data are subject to professional confidentiality obligations and data protection (GDPR). This rule ties both threads together, as governance involves two aspects:

• Data input to AI: Distinguishing between approved enterprise solutions with contractually guaranteed data processing (no use of inputs for training purposes, defined retention periods, EU data storage where applicable) and unapproved consumer tools; anonymization or pseudonymization of sensitive content wherever possible.
• Access and sharing in the cloud: deliberate management of permissions in the SharePoint/OneDrive storage location, restrictive use of external sharing links, and the use of sensitivity labels (Microsoft Purview) to technically protect sensitive models.

A confidentiality breach—whether caused by the wrong tool or an overly broad sharing link—cannot be remedied after the fact.

12) Traceability, Versioning, and Validation

The common thread of traditional financial modeling – the complete traceability of every number – applies all the more to AI-generated content and collaboratively edited models. This gives rise to three obligations:

• Document which parts of the model were created with AI support (e.g., in the comments column or a change log) – and thus also what requires special review.
• Cloud storage automatically maintains a version history; this is an advantage, but it does not replace deliberate change control. Key versions should be saved as named milestone versions and accompanied by a brief change note – precisely because AI and co-authorship greatly accelerate changes.
• No AI output and no external change is incorporated into the authoritative version without review. AI is a co-pilot, not the pilot. Validation is directly linked to the check mechanisms (Rule 7) and the dual-review process (Rule 10).

13) Responsibility Remains with Humans

AI supports the creation of content, and the cloud distributes it among many contributors – but responsibility must remain clearly attributable. Whoever is responsible for a model – especially as a professional in the context of a valuation, a fairness opinion, or due diligence – must be able to fully understand and professionally defend its results, regardless of how many contributors and tools were involved. Professional and legal liability cannot be delegated; it cannot be transferred to either a language model or a team. Efficiency gains through AI and collaborative work are welcome – but they must never come at the expense of comprehensibility and one’s own professional understanding of the model.

Conclusion

AI-powered tools and cloud-based, simultaneous collaboration are changing how financial models are created, reviewed, and used – but they do not change the definition of what constitutes a good model. On the contrary: A rigorous architecture, consistent structures, and robust control mechanisms are now essential prerequisites for AI to reliably understand a model and for multiple people to collaborate on it without conflict. Both developments raise the bar for quality—but only where professional discipline lays the groundwork for it. It is precisely this discipline that the rules presented here describe.

*Note: This version is a substantive revision of the original “10 Golden Rules of Financial Modeling” (as of 2019). The ten basic rules have been retained and supplemented with guidance on AI and collaborative cloud environments, as well as three new rules.*