AI + Project Finance Models: Where the Industry Stands and Where It’s Heading

This article summarises insights from the Forvis Mazars webinar: AI + Project Finance Models; Industry Update (February 2026). Presented by Rickard Wärnelid (Partner, Head of APAC Financial Advisory and Energy), Joshua Grimm (Director, APAC Energy, Singapore) and Rael Levinsohn (Associate Director, APAC Energy).

Artificial intelligence is no longer a fringe topic in project finance. Over recent months the market has seen a rapid acceleration in the development of AI-driven tools for Excel and financial modelling, and the conversation has shifted from “should we explore this?” to “how do we adopt it responsibly?”

When Forvis Mazars announced a webinar on AI and project finance models in February 2026, the response was striking – close to a thousand registrations, far exceeding the typical audience for a financial modelling webinar. As Rael Levinsohn noted during the session, there has scarcely been a training course in the past eighteen months where someone hasn’t asked about AI: What’s going on? What do you think about it? Are we going to lose our jobs?

The headlines tell part of the story. Financial news outlets are asking whether AI will kill Excel and entry-level investment banking roles. AI firms are paying investment bankers to help train their engines. Major consulting firms have faced scrutiny for using AI in government reports without adequate disclosure. LinkedIn feeds are filled with claims of building full financial models in minutes. The pace of change is real – but so is the noise. And for the project finance community, separating the signal from the noise matters enormously.

Why AI adoption in project finance modelling is different?

Much of the current AI discussion centres on corporate finance, M&A and FP&A workflows. Project finance modelling, however, operates under a distinct set of constraints. Models must be auditable, transparent and traceable. Lenders, sponsors and government stakeholders all need to trust the numbers, and that trust has historically been built on the ability to follow every calculation from input to output. Black-box solutions, no matter how impressive, do not meet this standard. For an example of how this transparency works in practice, see our tutorial on cash flow available for debt service (CFADS).

“The outputs are somewhat irrelevant if you don’t have a spreadsheet with logic that explains every step. That’s completely different to almost every other field. If you’re building a house and someone shows you the architect’s blueprints, no one says show me the working papers. But in project finance, if you don’t have the model, you say – how do I trust that this is right?”

Rickard Wärnelid

This distinction is what makes the project finance sector’s response to AI fundamentally different from what we see in consulting, legal or even statutory audit. As Joshua Grimm observed, the demands and requirements have always been different for project finance – and the same is proving true for AI adoption.

AI governance, data leakage and the reality of adoption across project finance

The webinar mapped out a wide range of concerns the presenting team hears from clients, lenders and advisors. At the governance level, organisations are grappling with questions around responsible use, data leakage risks and the distinction between enterprise-grade platforms like Microsoft Copilot and the fast-moving landscape of AI startups. The Deloitte incident – where AI-generated hallucinations in a government report attracted significant scrutiny, followed by internal guidance discouraging staff from emailing files to use on personal AI tools – underscored how seriously stakeholders take transparency around AI usage.

On the modelling side, the concerns are more specific: AI hallucinations and verification risk, the auditability of AI-generated outputs, undocumented assumptions or logic, and the question of review liability – if AI contributes to an error, who is accountable? And then there is the broader anxiety around controlling templates and modelling methodology in an environment where anyone with access to an AI tool could potentially generate model structures outside established standards.

“For most organisations, the governance and the responsible use of AI is that there is no AI. The tools are banned. You can’t get access to them. That is the vast majority of people’s experience. And even those who do have access have zero guidance, zero encouragement and no framework for how to use it.”

Rael Levinsohn

One of the most common anxieties, however, is not about AI itself but about being left behind. Many organisations report a feeling that everyone else is really advanced on this topic. The reality is that adoption across project finance remains very uneven. In training courses, when participants are asked who has access to AI tools at work, the answer is overwhelmingly “zero access” about ninety-five percent of the time. This gap between perception and practice is important – it means there is still time for organisations to develop thoughtful, governance-first approaches without feeling they have already fallen irreversibly behind.

AI platforms for financial modelling: ChatGPT, Copilot and Claude

The webinar identified seven platforms and categories generating the most discussion: OpenAI’s ChatGPT, Google Gemini, Microsoft Copilot, Anthropic’s Claude, the Shortcut platform, and a constantly shifting roster of AI startups. However, there is a notable disconnect between what people talk about and what they actually use at work.

ChatGPT dominates private usage – it is the platform people know and feel comfortable prompting. But in the professional project finance environment, if people have access to anything at all, it is Copilot. The webinar stressed the importance of starting with the tools you actually have access to, rather than waiting for a theoretically superior platform your organisation may never approve.

“A very common thing we hear is people saying, I’m waiting for Copilot, but what’s the point? Claude is so much better. But they haven’t actually tested Copilot yet. We’re seeing a lot of people extremely informed about what they think, and having done very little.”

Rickard Wärnelid

Excel Agent Mode was highlighted as a significant development – the entry point for understanding AI-assisted modelling. In a recent internal workshop, the team described an exercise where Agent Mode built dashboards with charts in moments: half the output was completely useless, the other half was amazing. For spreadsheet-level tasks the results can be remarkable, but for structured project finance models the tools lean into approaches like the LET function and named ranges everywhere – technically functional but not human-readable. It enables people with low skill to do impressive things, which also makes it extremely powerful and extremely dangerous. For a practical example of readable Excel techniques, our tutorial on using LOOKUP to replace VLOOKUP and HLOOKUP functions illustrates the kind of clarity project finance models require. It enables people with low skill to do impressive things, which also makes it extremely powerful and extremely dangerous.

Claude was singled out as the most hyped solution in project finance modelling, with many arguing it is the most powerful tool available. But the webinar noted a striking gap: almost no one in the project finance context has access to it at work. It remains a weekend activity for enthusiasts. The startup landscape presents a similar dynamic – new tools launch weekly, but before governance teams approve one, another will have taken over.

Where AI adds value in the project finance modelling lifecycle

A key theme was the importance of moving away from the binary expectation that AI can either do everything or nothing.

“This notion that AI can do absolutely everything, or AI can do absolutely nothing, is false. The key thing is: what iterative piece in the process could you make slightly better? It has to be a step-by-step process.”

Rael Levinsohn

Model documentation is one of the clearest wins. AI excels at reading complex formula structures and producing plain-language explanations. When someone puts in a seven-line formula that very few people understand, you can use AI to explain what it actually does and convert it into a multi-step calculation – powerful for bringing transparency where it was previously lacking. Our tutorial on debt sculpting to target DSCR without VBA is a good example of breaking complex logic into auditable steps, and this is precisely the kind of transparency AI can help bring where it was previously lacking.

Data extraction from contracts into structured input sheets is another practical use case. The team recommends a stepwise approach: extract into a table first, confirm it has worked, then restructure – much like instructing a well-trained colleague one step at a time, rather than expecting a single prompt to produce a finished model. When updated contracts come through, often running to two hundred pages, AI can quickly identify what has changed. Bringing those updates into an existing model remains harder – as Joshua Grimm noted, spreadsheets it gets, but when it comes to a project finance model, the structure is actually quite complex. Tutorials such as features of a cash flow waterfall in project finance and the loan life coverage ratio (LLCR) illustrate the layered logic these models require.

Credit and investment paper preparation is an area where AI performs well once the task shifts from numbers to narrative. As soon as numbers are being converted to text, it does exceptionally well. The underlying metrics AI needs to interpret, such as average DSCR and cash sweep mechanics, are covered in our tutorial library. Model review is similarly promising. Automating integrity checks and flagging errors suit AI’s pattern recognition capabilities, though a full model audit with its strategic and commercial alignment questions remains beyond current tools.

Full model development – generating a complete, bankable model from scope documents – remains at the frontier. The tools produce formulas that are technically functional but not human-readable, undermining the trust that project finance models exist to establish. As Joshua Grimm observed: will AI eventually build a full model? Absolutely. But it will be much more of an analyst than a full team.

How AI adoption in project finance compares to consulting, legal and audit

The webinar placed project finance in context against other professional services disciplines. In consulting, AI is already generating auto-decks and storyline drafts. In legal, dedicated platforms for clause detection and precedent retrieval mean firms are hiring fewer graduates and paralegals. In statutory audit, AI is being used for volume review of invoices, contract tagging and benchmarking – largely within internal platforms to protect data. In financial due diligence, tools for GL anomaly detection and tie-out automation are maturing rapidly.

The common thread, however, is consistent: AI is not doing everything. There is still a human point of analysis, review and decision-making in every case. Where content and language processing are the core task – consulting reports, legal documents – adoption has been fastest. Where numerical precision, auditability and stakeholder trust are paramount – project finance modelling – the uptake is slower but no less inevitable.

The future of AI in project finance: organisational learning, tailored agents and building trust

The webinar outlined several developments likely to shape the sector’s trajectory. Organisational learning – the ability to capture modelling methodologies, cross-project experience and institutional knowledge into AI-accessible formats – represents one of the most promising near-term opportunities. Imagine a project finance bank uploading five years of credit papers alongside the models and source documents that supported them. That creates a powerful training ground for AI to learn what good looks like in a specific institutional context. The challenge is scaling this without compromising confidentiality across projects.

Tailored agents – purpose-built for model development, credit paper drafting or sector benchmarking – are a logical next step beyond general-purpose AI tools. And building what the webinar called “project finance trust” requires full lineage, audit logs, change tracking and sign-offs. The parallel with Python is instructive: nearly a decade ago the question was whether Python-coded models would replace Excel. The answer was no, because despite being more efficient for certain calculations, Python lacked the transparency that stakeholders required. AI faces the same test.

“In project finance, everyone is looking at that model to check and understand exactly what’s going on. It’s why we continue to use Excel. It’s a similar thing when it comes to the adoption of AI – yes, we could use it in certain elements, but we still need to keep that trust.”

Joshua Grimm

The ultimate vision – from a Teams call to a model specification to an automated bankable model – remains aspirational. Some elements are already achievable, but the output, as the presenters observed, is not yet very useful for a genuine project finance context. The gap between a technically functional model and a bankable one is not a gap that technology alone can close.

The organisations best positioned will be those building the foundations now: developing governance frameworks, upskilling their teams and integrating AI into their workflows in a controlled, step-by-step way. As Rael Levinsohn concluded: the prize is huge – projects going to market faster, deals getting done with less pressure and fewer late nights. All of our jobs are here to stay, but they will be done quicker, faster, better and smarter.

—

To explore these themes in a hands-on setting, Forvis Mazars has launched a new two-day course – AI in Project Finance Modelling – available in Sydney and Singapore. The course equips participants with practical workflows, a prompt library and a governance playbook designed specifically for project finance teams.