Introduction

My first attempt with Claude on Excel? Instructive, but frankly frustrating.

I had created a consulting mission tracking model, a file I use regularly to manage workload, milestones, and billing. I thought: why not test Claude to optimize it? My initial brief was three words: “Optimize this file.”

Result: visually stunning. Impeccable formatting, well-aligned columns, consistent colors. Then I looked closer. Hardcoded numbers everywhere. No dynamic formulas. Broken business rules. Lost dependencies between sheets. It took four to five iterations to get a truly functional file. And most importantly, one question: would I have been faster alone?

According to McKinsey (2024), 58% of generative AI users report a longer learning curve than expected. But 87% report net gains after a few months. In hindsight, I understood: Claude works very well, provided you guide it properly. Here are my five mistakes, and most importantly, how to avoid them to transform AI into a genuine performance lever for finance functions.

1. Why My Vague Brief Generated Unusable Results

“Optimize this file.” Three words. No context. No rules. No constraints.

Claude interpreted. He restructured, reformatted, added calculated columns. But he also replaced formulas with fixed values, ignored business rules (for example, calculating daily rates based on consultant status), and broke links between sheets. The result was clean, but wrong. This is a key point: a visually perfect Excel file can be financially incorrect.

According to Gartner (2024), the majority of generative AI failures stem from poorly formulated instructions, not technical limitations. Claude supports files up to 30 MB each and up to 20 files uploaded per session (DataStudios.org, August 2025), enabling analysis of multi-sheet models and datasets containing hundreds of thousands of rows. The problem is therefore not technical capacity, but brief quality.

What I should have done: explain my actual usage (weekly workload tracking by consultant), specify my business rules (variable daily rates by profile, alert thresholds at 80% and 100% capacity), and provide concrete examples of edge cases (what happens if a consultant is between assignments?). A good Excel model starts with proper scoping. With AI, it’s even more critical. At Penon Partners, we apply this logic to all our finance transformation projects: the quality of the deliverable depends first on the clarity of the requirement.

2. My Value Proposition Has Changed: Architecture, Not Execution

While Claude worked, I felt useless. Then I understood: my role is changing. I’m no longer just building files. I’m defining their logic.

Claude executes. But he doesn’t understand the business. He doesn’t know that a daily rate must be consistent with consultant status, that a missed milestone automatically delays billing, or that a between-assignment period should be isolated in a dedicated sheet. My value proposition becomes model architecture: defining inputs (what data? what format? what frequency?), outputs (what tables? what KPIs?), business rules (what formulas? what thresholds?), and exceptions (what edge cases? what errors are possible?).

This is exactly the logic of a “finance factory” applied to Excel. My expertise doesn’t disappear. It moves up in abstraction. I spend less time writing formulas and more time structuring business logic. This is precisely what we observe at Penon Partners on our digital transformation missions: AI doesn’t replace finance expertise, it redeploys it toward higher-value tasks.

3. The Risks I Underestimated (and How to Anticipate Them)

My test also revealed several important risks. First risk: loss of control. A file can appear correct without being so. Errors are less visible, especially if formulas are replaced by fixed values. Second risk: tool dependency. Without understanding the model, it’s difficult to correct or evolve the file. Third risk: auditability issues. Hard to explain the logic if it wasn’t designed clearly. Fourth risk: illusion of competence. The tool creates an impression of mastery without guaranteeing quality.

Also worth noting: Claude cannot write, read, or execute VBA macros (NineTwoThree.co, 2025), which represents a significant limitation for organizations using legacy banking systems or VBA-based automated processes. If your model relies on complex macros, Claude cannot reproduce or optimize them.

Concretely, how to protect yourself? Systematically verify formulas (never trust a first version), document key rules in a dedicated sheet, test edge cases (missing data, inconsistent values, duplicates), and maintain simple, traceable logic. At Penon Partners, we apply these principles to all our finance process redesign projects: traceability and documentation are not options, they are prerequisites.

4. The Real Issue: Measuring AI ROI on Excel

At first, I lost time. Clearly. Creating the right brief, iterating, correcting—it’s longer than doing it myself on a small model. But the gain doesn’t come from the first version. It comes from subsequent iterations.

AI becomes profitable when the model is complex (multiple sheets, dozens of nested formulas, multiple business rules), when it’s reused regularly (weekly or monthly), and when it requires frequent evolution (adding KPIs, scope changes, new rules). It’s not worth it when the need is simple (a pivot table suffices), the timeline is short (immediate urgency), or control must be total (regulatory reporting, external audit).

Claude pricing ranges from $20/month (Pro) to $100-160/month (Max) (NineTwoThree.co, 2025). The Max tier is necessary for 200,000 token context windows and priority Opus routing. For a finance team of 5 to 10 people, the investment remains modest compared to time savings on complex models. But you must measure actual ROI: time saved, errors avoided, ability to evolve models quickly. This is exactly the approach we recommend in our AI project ROI framework: measure, iterate, adjust.

5. The 5 Tips That Transformed How I Work with Claude

Tip 1: Create a clear “Excel framework.” Formalize your logic before launching Claude: authorized formula types, business rules, sheet structure, data format. This drastically reduces errors and accelerates iterations. Concretely, I now create a “Rules” sheet in each model, with key assumptions, thresholds, and edge cases.

Tip 2: Structure your brief. A good brief contains four elements: objective (what should the file do?), constraints (what rules to follow?), sources (what input data?), expected result (what outputs?). For example: “Create a weekly workload tracking table by consultant, with automatic occupancy rate calculation, alert thresholds at 80% and 100%, and monthly consolidation by practice.”

Tip 3: Iterate in a targeted manner. One correction at a time. Systematic validation. Don’t ask “fix everything,” but “fix the formula in the Workload sheet, column E, so it accounts for consultant status.” This prevents regressions and keeps you in control.

Tip 4: Test exceptions. Missing data, inconsistent values, duplicates, edge cases. Claude generates and debugs Power Query M code smoothly (NineTwoThree.co, 2025), but it doesn’t guess your specific business cases. Systematically test atypical scenarios.

Tip 5: Provide examples. An existing model beats a long explanation. If you already have a working file, upload it and ask Claude to reproduce it with new rules. It’s much more efficient than an abstract description. At Penon Partners, we apply this logic to all our projects: start with what exists, identify what works, and iterate.

Conclusion

My first test with Claude on Excel was frustrating. But above all, revealing. The problem wasn’t the tool. It was how I worked with it.

Today, I structure before executing, test before validating, think architecture before formulas. My role is evolving. Less execution. More design. And above all, one conviction: AI doesn’t replace Excel. It changes how you use it.

For finance teams, this is an opportunity. Provided you don’t lose control. FP&A forum users report that traditionally week-long tasks (building an indirect cash flow statement) are compressed into an afternoon with Claude (NineTwoThree.co, 2025). But this performance is only accessible with a structured brief, clear rules, and systematic validation.

Generative AI is transforming the finance function. But it doesn’t dispense with business expertise. It redeploys it. This is exactly what we support at Penon Partners: helping finance teams build competency on these new tools while maintaining control of their processes and quality.