How well can AI handle financial modelling?

We test Claude, GPT, Gemini and other AI models on a standardised 13-cell financial modelling workbook and score each one on accuracy, formula logic, structure and presentation. The best score so far is 61 out of 80 (76%), from a skilled-prompt run. This tracker is a living scorecard, updated as models improve.

I started testing AI on financial models because I kept hearing the same claim: “AI can build your model for you now.” After 20 years of financial modelling across PwC, KPMG, Investec and NAB, that sounded like someone saying a calculator can do an audit.

So I built a framework to test it properly. Thirteen evaluation cells, each scored on the same four criteria I would use to review work from a junior analyst: accuracy, formula logic, structure and presentation.

The results surprised me. Not because AI was brilliant, and not because it failed completely. The surprise was how specific the failures were, and how consistent they became across models.

No hype, no vendor pitches. Just structured evaluations against a real workbook covering revenue builds, operating expenses, P&L consolidation, balance sheets, cash flow, debt schedules and more.

Key terms

AI financial modelling Using AI tools such as Claude, GPT or Gemini to build, review or extend financial models, including 3-way integrated models, revenue builds, debt schedules and valuations.

Evaluation cell One of 13 defined financial modelling tasks in our scoring framework. Each cell is assessed independently for accuracy, formula logic, structure and presentation, scored out of 80.

Cold prompt An AI model given a task brief with no prior context, examples or coaching. Tests the model’s baseline capability on a financial modelling task.

Skilled prompt An AI model given a structured brief with context, examples, formatting rules and step-by-step instructions. Tests the model’s ceiling when guided by someone who knows what to ask for.

Human reference score The benchmark score set by a qualified financial modeller completing the same task under the same rubric. Every AI score is compared against this baseline.

3-way financial model A fully integrated financial model linking the profit and loss statement, balance sheet and cash flow statement, where a change in any assumption flows correctly through all three.

Can AI actually build a financial model?

Yes, but with significant limitations. Across four completed evaluation runs, the best AI score reached 61 out of 80 (76%). That is roughly equivalent to a junior analyst who gets the structure right but misses reconciliation checks and makes formula errors in edge cases.

The gap between a cold prompt and a skilled prompt is large. When I gave Claude a bare task brief with no context, the output missed fundamental modelling conventions: hardcoded values where formulas should be, no error checks, inconsistent sign conventions. When I gave the same model a structured brief with examples and formatting rules, the output improved by 15 to 20 points.

That tells you something important about AI financial modelling today. The tool is not the bottleneck; the prompt is. If you do not know what good modelling looks like, you cannot coach an AI to produce it.

How do Claude, GPT and Gemini compare on financial modelling tasks?

No single AI model dominates across all 13 evaluation cells. Each model has strengths in specific areas and consistent weaknesses in others.

Claude tends to produce cleaner formula structures and follows formatting instructions more reliably. GPT often generates longer, more detailed outputs but introduces more hardcoded values. Gemini handles narrative explanations well but struggles with multi-step formula chains.

These are patterns, not absolutes. Every model version changes the picture, which is why this tracker exists as a living scorecard rather than a one-off comparison.

The evaluation runs so far include:

Scores are indicative and updated as new runs are completed. Click any run card in the tracker below to see the cell-by-cell breakdown.

What does each evaluation cell test in AI financial modelling?

Each of the 13 evaluation cells maps to a real financial modelling task that an accountant or analyst would perform inside a 3-way integrated model. The cells are not abstract exercises; they are the actual build steps in a monthly driver-based model.

The cells cover:

1. Revenue build Revenue drivers, pricing assumptions, volume growth, seasonal patterns
2. Cost of goods sold Variable cost structures tied to revenue drivers
3. Operating expenses Fixed and semi-variable cost schedules with escalation logic
4. Employee costs Headcount planning, salary escalation, on-costs and superannuation
5. Depreciation Asset register, useful life calculations, depreciation methods
6. Working capital Debtor days, creditor days, inventory days, all formula-driven from revenue and COGS
7. Debt schedule Drawdown, repayment, interest calculation, covenant testing
8. Tax Corporate tax calculation with carry-forward losses
9. P&L consolidation Bringing revenue, COGS, opex, depreciation, interest and tax together
10. Balance sheet Full balance sheet with opening and closing balances that reconcile
11. Cash flow statement Indirect method cash flow derived from P&L movements and balance sheet changes
12. 3-way integration check Does a change in one assumption flow correctly through all three statements?
13. Presentation and structure Formatting, sign conventions, labelling, print layout, error checks

Each cell is scored on four criteria worth 5 points each (total 20 per cell, 80 across four scored cells per run). The four criteria are accuracy, formula logic, structure and presentation.

Where does AI fail in financial modelling?

AI financial modelling breaks down in predictable ways. After four evaluation runs, the failure patterns are consistent across Claude, GPT and Gemini.

Hardcoded values instead of formulas. Every AI model tested inserted at least some hardcoded numbers where a formula-driven reference should be. In financial modelling, a hardcoded value is a maintenance risk: it does not update when assumptions change, and it breaks the model’s integrity.

Missing reconciliation checks. None of the cold-prompt runs produced a balance sheet that included a reconciliation check (assets minus liabilities minus equity equals zero). Skilled prompts improved this, but only when the brief explicitly called for it.

Inconsistent sign conventions. AI models frequently mix positive and negative signs within the same schedule. Expenses appear as positive numbers in one section and negative in another. This makes P&L consolidation unreliable without manual correction.

Circular reference avoidance that breaks logic. When a model encounters a potential circular reference (common in interest calculations that affect tax, which affects cash, which affects debt), AI tends to break the chain by removing the dependency entirely rather than implementing an iterative solution or a workaround.

Weak presentation standards. Formatting, column widths, number formatting and print-area settings are consistently poor across all models. AI treats presentation as cosmetic; in professional financial modelling, presentation is a quality signal.

Is AI ready to replace financial modellers?

No. Based on four structured evaluations, AI is a useful assistant for specific financial modelling tasks but cannot independently produce a model that meets professional standards.

The best AI score of 76% came from a skilled-prompt run. That means a qualified modeller spent time writing a detailed brief, providing examples and specifying formatting rules. The time saved by AI generating the output was partially offset by the time spent crafting the prompt and reviewing the result.

Where AI adds genuine value today:

• First-draft generation. AI can produce a starting structure faster than building from scratch, provided you review every formula.
• Formula explanation. AI is strong at explaining what a formula does and suggesting alternatives.
• Error checking support. AI can scan a model and flag potential issues, though it also flags false positives.
• Documentation. AI writes assumption registers and model notes faster than most analysts.

Where human oversight remains essential:

• Integration logic across all three financial statements
• Sign convention enforcement
• Circular reference resolution
• Professional presentation standards
• Commercial judgement on assumption reasonableness

How we score each AI model on financial modelling?

Every evaluation follows the same protocol to keep results comparable across models and prompt types.

Step 1: Task brief. The AI model receives a written brief describing a financial modelling task. For cold prompts, the brief contains only the task description and source data. For skilled prompts, the brief adds context, examples, formatting standards and step-by-step instructions.

Step 2: Output generation. The AI model produces its output, typically an Excel workbook or structured data. No manual corrections are made before scoring.

Step 3: Cell-by-cell scoring. Each of the 13 evaluation cells is scored independently against four criteria: accuracy (are the numbers correct?), formula logic (are formulas used instead of hardcoded values?), structure (is the layout logical and consistent?) and presentation (does it meet professional formatting standards?).

Step 4: Human reference comparison. The AI score is compared against the human reference score for the same task, completed by a qualified financial modeller under the same rubric.

Step 5: Documentation. Every run is documented with the full prompt, raw output, scored rubric and commentary. This documentation is what you see when you expand a run card in the tracker above.

What is a skilled prompt vs a cold prompt in AI financial modelling?

A cold prompt tests what an AI model can do with minimal guidance. A skilled prompt tests what it can do when guided by someone with domain expertise.

The distinction matters because it separates two questions: “How capable is the AI model?” and “How capable is the AI model when used by a qualified professional?”

In our evaluations, the gap between cold and skilled runs ranges from 15 to 20 points out of 80. That gap represents the modeller’s contribution through prompt design, not the AI’s native capability.

For accountants and finance professionals evaluating AI tools for financial modelling, the skilled-prompt score is more relevant. It reflects realistic usage: a professional using AI as an assistant, not a replacement.

Why we built this AI financial modelling tracker?

Most AI benchmarks test language tasks: summarisation, coding, reasoning puzzles. None of them test what matters to accountants and finance professionals: can this tool help me build, review or extend a financial model?

I built this evaluation framework because the question kept coming up in training sessions, client workshops and conference Q&As. “Which AI should I use for modelling?” There was no structured, comparable answer.

This tracker fills that gap. It applies the same scoring rubric across every model, every prompt type and every evaluation cell. When a new model version drops or a new tool enters the market, it gets the same 13-cell test.

The framework is designed for the accounting and finance profession specifically. The evaluation cells map to the tasks in a standard 3-way integrated financial model. The scoring criteria match the quality standards a senior modeller would apply when reviewing work.

We re-run evaluations as models improve and publish the results here so you can see which tools are ready for your workflow and where human oversight is still essential.

Frequently asked questions

Can AI build a financial model?

Yes, but with significant limitations. In our testing across 13 financial modelling tasks, leading AI models (Claude, GPT, Gemini) produce workable outputs for simpler cells like revenue builds and operating expenses, but consistently struggle with balance sheet reconciliation, cash flow linkage, and circular debt schedules. No model has yet produced a fully integrated 3-way model without manual correction.

Which AI is best for financial modelling?

Based on our structured evaluations, Claude (Anthropic) currently scores highest across financial modelling tasks, particularly on formula logic and structural consistency. GPT-4o (OpenAI) performs well on revenue and cost builds but drops on integration tasks. Gemini shows improving trend lines but lags on accuracy. Results are updated regularly as models improve.

How do you evaluate AI financial modelling accuracy?

We use a structured evaluation framework that tests each AI model against 13 financial modelling cells — from revenue builds and operating expenses through to full 3-way integration. Each cell is scored on four dimensions: accuracy (correct numbers), formula logic (proper Excel formulas, not hardcoded values), structure (layout, labelling, flow), and presentation (formatting, readability). Scores range from 0 to 10 across a 8 grading rubric.

Should accountants use AI for financial modelling?

Yes, as an assistant rather than a replacement. AI adds genuine value for first-draft generation, formula explanation, error-checking support and model documentation. It is not ready to produce a model that meets professional standards without human review. The skilled-prompt results show that the value of AI in financial modelling depends heavily on the quality of the instructions you give it.

What is an AI model evaluation framework for financial modelling?

An AI model evaluation framework is a structured scoring system that tests AI tools on defined financial modelling tasks. Our framework uses 13 evaluation cells covering revenue builds, operating expenses, balance sheets, cash flow and more. Each cell is scored on accuracy, formula logic, structure and presentation, with results compared against a human reference score. The framework allows consistent comparison across different AI models and prompt approaches.

How does the scoring rubric work?

Each evaluation cell is scored on four criteria, each worth up to 5 points: accuracy (correct numbers and calculations), formula logic (formula-driven rather than hardcoded), structure (logical layout and consistency) and presentation (professional formatting, labels and error checks). The maximum score per cell is 20. Runs typically score four cells at a time, giving a maximum of 80 per run.

What financial modelling tasks does this evaluation cover?

The framework covers 13 tasks that map to a standard 3-way driver-based financial model: revenue build, cost of goods sold, operating expenses, employee costs, depreciation, working capital, debt schedule, tax calculation, P&L consolidation, balance sheet, cash flow statement, 3-way integration check and presentation standards. These are the actual build steps a financial modeller would complete.