Governance sets the pace for AI in financial modelling

A new report from the IPFA Special Interest Group (SIG) (opens in a new tab) on AI in financial modelling finds that the technology is moving faster than the governance frameworks needed to support it. The SIG is chaired by Forvis Mazars, and the report was written by Jerome Brice (opens in a new tab), Ryan Humphrey (opens in a new tab) and Rickard Wärnelid (opens in a new tab) of Forvis Mazars. In regulated institutions that gap, and not the quality of the AI itself, is what holds adoption back.

The SIG ran six roundtables with 57 participants across EMEA and APAC, from banks, rating agencies, engineering firms, legal advisors and government agencies. Its aim was to understand how AI is really being used in project finance and infrastructure modelling today.

How far AI adoption in financial modelling has got

Not far, and with caution. Across the roundtables, adoption was still mostly at the concept and experimentation stage, and only a few members were testing AI tools in live workflows. Most participants were taking a measured approach rather than rushing to deploy. In the report’s words, “only a limited number of organisations have meaningfully embedded AI into live financial modelling workflows.” For firms worried about falling behind, that finding should reassure: the industry as a whole is still working this out.

Why governance matters more than model capability

Traditional model risk frameworks were built for deterministic spreadsheet logic and traceable formulas. AI brings probabilistic outputs and opaque reasoning, and its outputs can vary from one run to the next. That breaks the assumptions of reproducibility and auditability that lenders, auditors and regulators rely on. The report is direct about what follows: “a more capable AI model is not automatically a more deployable tool. In financial modelling, practical value depends as much on integration, governance and control as it does on raw AI performance.”

One issue came up again and again: informal use of consumer AI tools outside approved enterprise environments, even when users think they are being careful by entering dummy assumptions. Once that output goes back into a company model, there is no audit trail showing how it was produced or which assumptions sat behind it. Participants from regulated institutions, particularly banks, were clear that this kind of workaround is not acceptable.

The risk of knowledge overshadowing

Dr Simon Thorne (opens in a new tab), Senior Lecturer in Computer Science at Cardiff School of Technologies, joined one of the roundtables to share his benchmarking research. He described knowledge overshadowing: more recent information in a prompt displaces earlier context the model still needs. In a financial model, assumptions set early in a workflow can quietly stop being carried through, especially in larger or more complex models. The output still looks coherent, but it no longer reflects every input it depends on. That is a different risk from a standard hallucination, and harder to spot on a first read.

The AI tools modellers are using

Microsoft Copilot is still the tool most often available inside organisations, mainly because it sits within existing enterprise environments and is easier to align with internal security requirements. Availability is not the same as capability, though. Copilot in Excel is deliberately constrained: it works only on files in OneDrive or SharePoint, relies on structured workbook content and is restricted to context within that file.

Frontier large language models are moving faster. Participants noted that Claude in particular had moved forward since the SIG began, and can now produce more convincing first-cut project finance models aligned with globally accepted best practice guidelines.

What AI means for the skills junior modellers build

This was one of the harder questions the SIG raised. Junior analysts have always built modelling judgement by constructing models from scratch, debugging formula errors and reconciling data by hand. If AI takes over those tasks, a skills gap could open: analysts who can operate the tool but lack the intuition to know when its output is wrong. Participants saw that risk as particularly acute in project finance and infrastructure, where technical judgement built over years is central to understanding risk.

The report sets out a two-track answer: “the future skill set is twofold: practitioners must learn how to use AI effectively, while still developing the core financial modelling judgement needed to know when the AI is wrong.” Foundational modelling expertise does not matter less. It becomes something to maintain actively throughout a career, rather than something learned once at the start.

The SIG’s recommendations for governing AI in financial models

The SIG’s recommendations are practical:

• Define a risk-based control framework for AI usage anchored at the workflow level, covering scope, roles, testing and ongoing monitoring
• Mandate human-in-the-loop review for material judgements and external disclosures, so accountability stays clearly assigned
• Adopt robust documentation that captures prompts, datasets, model versions and decision logs to support reproducibility
• Establish sandbox environments with enterprise controls and data loss prevention to enable controlled experimentation
• Build a shared understanding across departments on how AI tools should be used, to avoid inconsistent application and governance gaps

The report closes on where the industry stands: adoption will be paced by assurance, not capability. Frontier models are improving, but that alone does not settle reproducibility, auditability and stakeholder trust. Those decide whether an AI-assisted financial model can be relied on for decisions that move money.

For more on this topic, read our observations on AI in financial modelling or watch our webinar AI and project finance models: industry update. To build the skills yourself, see the AI in Project Finance Modelling course.