How AI Can Build a Reliable Excel Expense Analysis

Written byAnna Katharina Bollé

AI & IntelligenceAugust 6, 202610 minutes

AI can help finance teams move from raw transaction exports to structured insights much faster. But the safest way to use generative AI is not to upload an expense export, ask for an analysis including a final total, and assume every single line was processed. For deeper financial analysis, the ideal division of labour is simple: let the AI design the workbook, let Excel perform the calculations, and let finance validate the results. This hybrid Excel AI approach gives you speed without sacrificing accuracy. We tested this approach on SaaS and AI spend, a fast-growing cost category that is often difficult to analyse from messy ledger exports. We also turned the workflow into a reusable skill that transforms a raw expense export into a structured Excel model with standardised vendors, formulas, summaries, and charts. You can try the skill yourself as a starting point for your own SaaS and AI spend analysis, while exploring how AI and Excel can work together as a practical tool for first-pass financial analysis.

The risk of using AI as a calculation engine

A raw large language model is not fundamentally a calculation engine. Research shows that raw LLMs can hallucinate in up to 41% of finance-related queries, and their numerical calculation accuracy when operating in-head is roughly 52% (Ernst & Young). They generate responses by predicting the next likely token based on statistical patterns, rather than systematically processing transactions according to fixed accounting rules.

This probabilistic approach makes them highly unpredictable for core financial tasks. While they excel at processing language and summarising notes, they are fundamentally built for likelihoods rather than rigid rule enforcement.

When you upload a transaction log to chatbot interfaces, the tool might generate a perfectly formatted table with a clear final total, but a confident total does not prove that every single transaction was included in the math. Even when a dataset fits inside an AI's context window, spatial reasoning and attention limitations can cause the model to skip rows, cross-contaminate columns, or stop early without showing any error.

In the UK, the regulatory environment is increasingly strict about these risks. Professional regulatory guidance established in 2026 confirms that professional accountability remains unchanged when using generative or agentic AI. The human-in-the-loop principle is absolute, requiring financial officers to actively verify AI-generated financial outputs.

How deterministic Excel formulas solve the trust problem

To build a truly reliable workflow, you need to combine the creative structuring of an AI engine with the deterministic modelling of traditional spreadsheets. While probabilistic models operate on statistical patterns, deterministic Excel formulas operate on fixed, rule-based mathematical functions where identical inputs always yield identical, auditable outputs.

An effective way to achieve this is by asking your AI tool to generate a Python script that classifies your data and builds the Excel file - with all the number-crunching done by native spreadsheet formulas, not inside the chat. For example, using your chatbot as an excel AI formula generator that writes native spreadsheet functions keeps your workflows transparent. Rather than writing 'Your total SaaS spend is £15,000,' the script inserts an actual =SUMIF() formula into your sheet, so Excel does the calculation itself. This keeps the calculation logic completely visible and easy to audit when you open the file in Excel.

This might even improve output as human-built spreadsheets are famously prone to mistakes. Meta-analyses compiled by Dr. Raymond R. Panko found that 88% of audited corporate spreadsheets contain errors, while a study from Dartmouth College discovered errors in 94% of operational spreadsheets (Dartmouth College). Using an automated tool to design the initial sheet reduces manual typing mistakes, while keeping the math in Excel allows you to easily find and fix any formula errors.

Testing the workflow on SaaS and AI spend

We chose SaaS & AI spend management as our test case because it is one of the fastest-growing cost categories for modern businesses. The scale of this spending was highlighted in the Moss 2026 European SaaS & AI Spending Report, which analysed €302 million in company card transactions across Europe from January 2024 to May 2026 (Moss). Corporate AI spend in Europe alone grew to €7.1 million in the first five months of 2026, already exceeding the total spent in all of 2025 (€5.7 million) and 2024 (€1.5 million).

According to the same report, corporate AI spend in Europe is on track to hit €17 million by the end of 2026, representing a three-fold year-over-year jump. It is rapidly transitioning to the third-largest software spending category, capturing 15.4% of all corporate software spend. Keeping track of these exploding costs is crucial, yet most finance teams are still trying to manage them through messy ledger exports.

Why software spend is difficult to analyse from the ledger

Software spend is difficult to understand directly from the general ledger. This is because monthly subscriptions, annual prepayments, and usage-based charges create highly erratic transaction patterns across different periods. Traditional bookkeeping processes require teams to scroll through thousands of transaction lines, clean inconsistent vendor descriptions, and group bookings manually. It is a slow, tedious task that pulls your finance team away from strategic planning.

Furthermore, manual processes are highly prone to administrative waste. According to a study by Concur, 19% of manual expense reports contain errors, taking an average of 18 minutes and costing $52 to resolve (Concur). Using AI for pre accounting to structure this data reduces these error rates while giving budget owners immediate visibility.

The data is there. The first usable analysis is not. That is why we built a skill that could turn a raw expense-account export into a predefined Excel model rather than asking finance to build the workbook, formulas, summaries and charts manually each time.

Four lessons from building an Excel AI expense tracker skill

When we built our own skill to analyze our SaaS & AI expense account using AI that builds an excel, we learned that a successful workflow requires a clear division of work. We found two highly practical lessons for planning the calculation logic and two for designing the visual output.

Designing calculation logic before prompting

First, you must agree on the complete calculation plan before building. Start in plan mode and define the workbook on paper, specifying which sheets, tables, and columns you need and how they should connect. Once the structure is agreed, tools like Claude Code can implement the design in one focused iteration. This avoids running the skill only to discover that a critical column or formula is missing, saving you time, tokens, and repeated rebuilds.

Second, ask the AI to review the working first version. A functioning first draft gives the model something concrete to analyse and improve. At this stage, it is much easier to identify missing checks, unclear sections, or useful additional formulas. This step-by-step process helps you build a solid financial tool rather than a cluttered, overbuilt spreadsheet.

Using visual mock-ups and screenshots for design

Third, if design matters, always design the workbook layout before prompting the final interface. Trying to describe a complete visual design through text alone often produces layout results that are noticeably off. We found that creating a mock-up in Claude Design first, and then feeding a screenshot to the code builder, allowed the AI to implement the intended design perfectly in a single run.

Fourth, use screenshots to show visual problems during development. When spacing, formatting, alignment, or hierarchy looks wrong, providing a screenshot is far more effective than spending several messages describing the issue. The model can instantly analyse the visual reference and produce a faster, more accurate correction.

Learning: For calculation logic, define the rules; for visual output, show the target.

How Our Skill Builds a SaaS and AI Spend Workbook

For finance teams looking for a practical way to use AI for Excel, a repeatable custom skill provides a straightforward starting point. We developed a chat-based skill that turns a raw expense export into a working Excel model in minutes. The tool automatically identifies SaaS, AI, and licensing vendors while filtering out unrelated costs. It then consolidates chaotic vendor descriptions under clean, standardised names and organises everything by month, category and spend.

This removes two of the most time-consuming parts of the traditional finance workflow. Your team no longer needs to scroll through thousands of transaction rows to build a supplier list, nor do they have to spend hours writing formulas and formatting charts from scratch.

The AI handles the repetitive groundwork, letting you focus on analysing the results.

What the automated SaaS spend analysis includes

The resulting workbook provides a comprehensive overview of your software costs. This structured analysis includes the following key components:

• Visual insight graphs: High-level charts showing your top vendors, spending by category, and monthly trends.
• Detailed analysis sheets: Granular, deep-dive views broken down by vendor, month, and department category.
• Smart anomaly checks: Automated flags for sudden spending spikes, irregular billing patterns, and inactive vendors.
• Budget comparison: Optional budget-versus-actual templates to track variances automatically.
• Reconciliation backup: A clean copy of the original transaction list to serve as your source of truth.
• Transparent Excel formulas: Visible, native formulas connecting the raw transaction data to your summary dashboards.

Because every summary remains linked to the raw data through visible formulas, you are never locked into the AI's first guess. You can easily adjust classifications, run an internal audit of specific rows, or extend the model for forecasting and board reporting.

The tool does not replace financial judgment; it simply handles the file setup so you can start analysing immediately.

How to run the AI software spend skill in ChatGPT or Claude

Go ahead and try the skill yourself to find the licences draining your budget. For the best results, we recommend uploading around 12 months of software, SaaS, and licensing transactions so the analysis captures annual prepayments as well as monthly subscriptions, giving you a more complete view of recurring software commitments.

Here is how you can run the skill in your preferred chatbot:

• Using ChatGPT: Upload the skill ZIP file and your raw transaction workbook directly into the chat interface. Copy and paste the provided prompt into the text box, then let the system run the analysis.
• Using Claude: Add the ZIP file to your skill library first. Open a fresh chat, upload your transaction workbook, and run the /moss-software-spend-intelligence command in the text box.
• Validating the output: Once the tool generates the file, open it in Excel, verify that the total row count and total value match your source export, and review the main formulas.

Using this skill helps finance teams bypass manual setup and move straight to investigating which licenses drove the largest cost increases, where tools may overlap, which teams are the main cost drivers, and which plans could be reviewed or renegotiated.

The wider benefit comes from building or using repeatable skills like this one. Instead of asking AI to recreate the workbook from a new prompt every time, finance can save the analytical logic, structure, formulas, and quality checks as a reusable process. Excel continues to handle the calculations, while the skill prepares the model consistently and makes the workflow significantly faster.

That is the practical shift: less time building the same analysis again, and more time using the workbook to make better decisions.

Common questions about using AI to build an Excel