AI Is Changing Financial Modeling. Can You Still Trust the Model?

Written by: Elizabeth Renteria

Artificial intelligence is rapidly changing the way financial models are built.

Tasks that once took hours can now be completed in minutes. AI can generate model structures, write formulas, suggest assumptions, explain calculations and even produce an entire first draft of a financial model from a simple prompt.

There is enormous potential here.

But as AI becomes more capable, I think an equally important question needs to be asked:

Do we understand how the model was built to know where it needs to be challenged?

That is a different question from whether the model works. A financial model can balance. Its formulas can calculate correctly. Its outputs can look entirely reasonable.

And it can still be wrong.

AI introduces a different type of model risk

Traditional financial modelling already requires careful review. But AI introduces some additional risks that reviewers need to consider.

For example, AI can confidently generate information that doesn't exist. It can create sources, references or market data that appear credible but cannot be verified. It can also produce calculations without providing the same transparency around the modelling logic that an experienced analyst would typically expect.

And perhaps most importantly, AI is designed to respond to a prompt and produce a useful answer. Financial modelling works differently.

A good model should not be designed to arrive at a predetermined answer. The assumptions and logic should drive the result.

If the model says an investment generates a particular return, the purpose of the model is to demonstrate why that return occurs, not to construct a series of calculations that gets you to the desired number.

That distinction matters.

Because if AI has played a significant role in constructing the model, the reviewer needs to understand not only what the model does, but how it got there.

The review process needs to start earlier

This is where I think the role of the model reviewer is going to evolve. A good review of an AI-assisted financial model shouldn't begin with opening the completed Excel file and checking whether the formulas work.

It should begin with understanding the model's provenance.

• Which parts were built by an analyst?
• Which parts were generated or modified by AI?
• What prompts were used?
• Were assumptions suggested by AI, sourced independently by the analyst, or both?
• Did AI generate individual formulas, entire sections of the model, or the underlying model architecture?
• Were any AI-generated outputs subsequently modified by an analyst?

Understanding this doesn't mean reviewing every cell differently simply because AI was involved. It means understanding where the model's risk sits and tailoring the review accordingly. For example, an experienced reviewer might apply greater scrutiny to an area where AI has:

• Generated the underlying modelling logic;
• Introduced assumptions or external data;
• Created complex formulas that an analyst has not independently validated;
• Interpreted contractual or technical information;
• Built calculations where the rationale isn't immediately transparent; or
• Made changes to an existing model without a clear audit trail.

Conversely, where an experienced analyst has independently developed and validated the methodology, and AI has simply been used to automate repetitive tasks or assist with formula writing, the review risk may be very different.

The point isn't to treat AI-generated work as automatically wrong. The point is to understand what was done by AI, what was done by people, and what level of review each component requires.

This is where governance becomes practical

I don't see governance as a separate layer sitting on top of the modelling process.

For AI-assisted modelling, governance should be embedded in the way the model is developed and reviewed. A sensible process might involve maintaining a clear record of:

1. What AI was used for

Was it used to generate formulas? Develop model architecture? Research assumptions? Write documentation? Check for errors?

2. What the analyst did

Which assumptions were independently sourced? Which calculations were developed or validated by the analyst? What judgement was applied?

3. What was independently verified

Were AI-generated sources checked? Were technical assumptions validated? Were key calculations rebuilt independently?

4. What the reviewer needs to focus on

Which parts of the model carry the greatest risk because of how they were developed?

This information gives the reviewer something incredibly valuable: context.

Instead of applying a standard checklist to every part of the model, they can focus their expertise where it is most needed.

This makes technical expertise more important, not less

And this is why I don't believe AI will reduce the need for experienced financial modelling advisors. In many respects, it increases it.

The value of a good reviewer isn't simply knowing how to check whether an Excel formula is correct.

• It's knowing whether the formula is appropriate in the first place;
• It's understanding the commercial, technical, accounting, tax and financing considerations that sit behind the calculation;
• It's knowing which assumptions matter;
• It's knowing where an apparently reasonable result should make you uncomfortable; and
• It's knowing what questions to ask when the model doesn't tell you the whole story.

That requires industry knowledge as well as modelling expertise.

An AI tool may be able to produce a technically valid calculation for a debt repayment profile, but an experienced infrastructure finance advisor can challenge whether the repayment profile actually reflects the financing documentation.

AI may be able to build a production forecast, but an industry specialist can assess whether the operating assumptions make sense in the real world.

AI may produce a tax calculation that looks perfectly logical, but a tax specialist can determine whether the underlying treatment is actually appropriate.

That's the difference between checking a model and reviewing a model.

The future of model assurance

As AI becomes increasingly embedded in financial modelling, I expect the strongest modelling processes will be those that combine three things:

• AI capability to improve speed and productivity;
• Human expertise to provide commercial and technical judgement; and
• Governance and independent review to connect the two.

The objective shouldn't be to prevent AI from being used. It should be to make sure that when AI is used, we understand what it has done, where the risks are, and how those risks have been independently addressed.

Because ultimately, the question isn't:

"Was this model built by AI?"

It's:

"Do we understand how this model was built, have we reviewed the right things, and can we trust the answer it produces?"

As AI takes on more of the mechanics of financial modelling, I believe that is where experienced advisors will add the most value. Not simply building the model. Not simply checking the spreadsheet. But understanding how the model was built, where the risks are, and whether the model can be trusted to support the decision being made.