Generative AI in Financial Modelling
Executive Summary
Key Takeaways
- ✓ Generative AI's role in financial modelling is bounded to drafting, structure suggestion, and narrative tasks; it should not be the source of a model's verified calculated output.
- ✓ Hallucination, a language model producing a plausible-sounding but fabricated figure, citation, or fact, is the single most important failure mode to guard against when using generative AI in a modelling workflow.
- ✓ A generative AI-drafted formula can be syntactically valid and read as reasonable while still applying incorrect logic, which is why formula review remains necessary regardless of drafting speed.
- ✓ Any factual claim, citation, or reference a generative AI produces in modelling-adjacent commentary should be independently verified before inclusion in material used to support a decision.
- ✓ Containing generative AI's failure modes requires specific review practices, not general caution; this guide sets out what those practices concretely are.
Objective¶
This guide sets the bounded, specific role generative AI plays in financial modelling, within AI Financial Modelling & Artificial Intelligence in Finance.
The Bounded Role¶
Generative AI, a large language model producing new text or content from a prompt, is well suited to drafting, structure suggestion, formatting, and narrative commentary within a modelling workflow. It is not well suited, and should not be relied on, to produce a model's verified calculated output directly, for the reasons set out on AI in Financial Modelling.
Specific Failure Modes¶
Hallucination. A language model can produce a plausible-sounding but entirely fabricated figure, citation, or fact, stated with the same fluent confidence as accurate output. This is the single most consequential failure mode in modelling-adjacent generative AI use, because fluency provides no signal of accuracy.
Plausible-but-incorrect formula logic. A drafted formula can be syntactically valid, well-formatted, and internally consistent while still applying the wrong calculation logic to the underlying problem, an error that is not visible from the formula's appearance alone.
Unverifiable citations. A language model can produce a citation, reference, or source attribution that does not correspond to any real source, or that misstates the actual content of a real source, in either case appearing legitimate until checked.
Context drift across a long drafting session. Over a long interactive drafting session, a language model's output can gradually drift from the original stated assumptions or constraints without an explicit signal that drift has occurred.
Review Practices That Contain Each Failure Mode¶
For hallucination: treat any specific figure, fact, or claim generated by AI as unverified until checked against an independent source, exactly as an unverified assumption from any other source would be treated.
For plausible-but-incorrect formulas: apply the same independent formula review a reviewer would apply to a colleague's work, regardless of how confidently or fluently the formula was presented.
For unverifiable citations: verify every citation against its original source before it appears in material supporting a material decision; do not accept a citation's existence on the language model's assertion alone.
For context drift: periodically restate the governing assumptions and constraints explicitly within a long drafting session, rather than assuming earlier context remains intact throughout.
Common Construction Pitfalls¶
Mistaking fluency for verification. A well-written, confident-sounding draft carries no more evidential weight than an unverified draft from any other source.
Skipping review because the task felt mechanical. Even mechanical-seeming drafting tasks, a formula pattern, a citation list, can contain the specific failure modes described above, and warrant the same review discipline as any other drafted work.
Using generative AI for tasks better suited to machine learning. Numerical prediction and classification tasks are better served by machine learning approaches with measurable empirical accuracy, addressed in Machine Learning vs. Financial Modelling.
Recommended Practices¶
- Confine generative AI's role in a modelling workflow to drafting, structure, formatting, and narrative tasks.
- Verify every AI-generated figure, fact, and citation independently before it supports a material decision.
- Apply the same formula review discipline to AI-drafted formulas as to formulas from any other source.
- Restate governing assumptions periodically during long interactive drafting sessions to guard against context drift.
Continue Reading¶
Related Pillars¶
Related Technical Guides¶
Related Comparisons¶
Related Glossary¶
How OXXON tests thisRun a free structural check with FMAE
Frequently Asked Questions
What is generative AI's role in financial modelling?
A bounded role focused on drafting, structure suggestion, formatting, and narrative commentary, accelerating the mechanical construction and communication around a model, not producing the model's verified calculated output directly.
What is hallucination in the context of financial modelling?
A language model producing a plausible-sounding but fabricated figure, citation, or fact, stated with the same fluency and confidence as accurate output, making it specifically difficult to detect without independent verification.
Can a generative AI-drafted formula look correct but actually be wrong?
Yes. A drafted formula can be syntactically valid, well-formatted, and read as reasonable while still applying incorrect logic to the underlying calculation, which is why independent formula review remains necessary regardless of how polished the draft appears.
How should citations or facts produced by generative AI be handled?
Any factual claim, citation, or reference generative AI produces should be independently verified against its original source before being included in material used to support a material decision, since such claims can be fabricated convincingly.
Is caution alone sufficient to manage generative AI's risks in modelling?
No. General caution is not a control. Specific review practices, formula review, citation verification, and number tie-outs, are what actually contain generative AI's known failure modes in a modelling workflow.
Related Articles
AI Financial Modelling & Artificial Intelligence in Finance
AI financial modelling is the application of machine learning and generative AI techniques within the financial modelling process itself, driver identification, construction assistance, scenario generation, and narrative drafting, while artificial intelligence in finance is the broader application of those same technique categories across the finance function generally. This page is the hub for the Knowledge Centre's AI financial modelling content: the foundational distinction between machine learning, natural language processing, and generative AI; how AI accelerates modelling construction without replacing the auditable calculation layer beneath it; a staged framework for adopting AI reliably; enterprise applications across FP&A, forecasting, valuation, and investment analysis; governance and risk practice; and the institutional best practice synthesis this domain builds toward.
AI in Financial Modelling
AI in financial modelling refers to the application of machine learning and generative AI techniques within the modelling process itself, rather than across the finance function broadly: identifying candidate drivers from historical data, assisting with formula and structure construction, generating scenario variations, and drafting narrative commentary around a model's output. This guide sets out where these applications add genuine value and, just as importantly, where the calculated number itself must remain the output of a structured, auditable model rather than of the AI directly.
Large Language Model
A large language model, or LLM, is a machine learning model trained on very large volumes of text to predict and generate coherent, contextually relevant language. LLMs form the basis of most generative AI tools used in finance, drafting, summarisation, and conversational assistants, and their fluency is not itself evidence of factual accuracy, a distinction central to using them reliably in a finance context.
AI Copilot
An AI copilot is a generative AI assistant, typically built on a large language model, embedded directly within a finance workflow tool, a spreadsheet, an FP&A platform, a reporting system, to support tasks such as drafting, formula assistance, and summarisation through an interactive, conversational interface. A copilot accelerates specific tasks within existing workflow; it does not itself constitute a verification or governance control over the output it produces.
AI-Assisted Financial Analysis
AI-assisted financial analysis works best when structured as a defined workflow rather than an ad hoc use of a chat tool: decomposing an analysis into discrete tasks, assigning each task to the approach best suited to it (AI-assisted or human-led), and placing a human verification checkpoint at each point where AI output feeds into a conclusion. This guide sets out that workflow structure and the checkpoint discipline that keeps it reliable.
Machine Learning vs. Financial Modelling
Machine learning and traditional financial modelling both produce quantitative output used to support decisions, but differ fundamentally in method (statistical pattern learning versus explicit, auditable formula logic), output character (a probabilistic estimate versus a traceable calculated number), and reliability characteristics. This comparison sets out those differences and why the two are best understood as complementary techniques, machine learning informing assumptions, financial modelling calculating auditable output, rather than substitutes for one another.