A few months ago, I was experimenting with an AI model on a relatively straightforward financial modelling task. I gave it a clean dataset, a clear objective, and a detailed prompt. The output was good enough that it could have served as the starting point for a real engagement.
Then I changed the prompt.
Not dramatically. I rephrased a few instructions, added some context, and removed a sentence that I thought was unnecessary. The model still produced a coherent answer, but the methodology had changed. Some assumptions had shifted, a few adjustments disappeared, and the final output was no longer identical.
That is not a criticism of AI. It is simply how these systems work.
It also illustrates why professional financial analysis cannot stop at the point where an answer is generated.
The conversation around AI often focuses on capability. Can it build a discounted cash flow model? Can it summarize a 10-K? Can it write Python code to reconcile two datasets? Increasingly, the answer is yes.
The more interesting question is different.
Can another analyst independently reproduce the work six months later, explain every assumption that was made, and defend every number if it is challenged?
Those are not the same problem.
Reproducibility Is a Professional Requirement
Financial analysis should be repeatable.
If two analysts begin with the same information and follow the same methodology, their conclusions should not depend on the wording of a prompt or the version of an AI model available that day. They should depend on the assumptions that were documented and the analytical process that was followed.
AI can absolutely become part of that process. It can automate repetitive work, identify patterns, generate code, and reduce the time spent on mechanical tasks. None of those capabilities diminish the need for a methodology that another professional can review and reproduce independently.
A model is only as reliable as the process behind it.
Context Is More Than a Token Limit
Much has been written about context windows and token limits. Those technical constraints matter, but they are not the real issue.
The real challenge is that financial engagements accumulate context over time.
A valuation may begin with historical financial statements, but the analysis gradually incorporates management interviews, customer contracts, industry reports, revised forecasts, board presentations, accounting policies, and countless discussions with the client. The analyst develops an understanding that is built over weeks or months, not from a single collection of documents.
AI is exceptionally good at processing information that it is given.
Professional analysts are responsible for recognizing what information has not yet been considered.
That distinction becomes increasingly important as engagements become more complex.
Every Number Should Be Explainable
One question has followed me through almost every analytical engagement.
“Where did this number come from?”
It sounds simple, but it captures the essence of good financial work.
The answer should never be, “The model generated it.”
It should be possible to identify the source data, explain every adjustment, justify each assumption, and reproduce the calculation without relying on memory or undocumented steps. If another analyst cannot follow that path, confidence in the result begins to erode, regardless of whether the number itself happens to be correct.
In practice, this means every output should have a clear lineage. Source data leads to transformations. Transformations lead to calculations. Calculations lead to conclusions. Every step should be visible.
Traceability is not administrative overhead. It is part of the analysis.
Speed Does Not Reduce Responsibility
One of AI’s greatest contributions is that it removes a great deal of repetitive work.
That is a welcome development.
Analysts should spend less time cleaning spreadsheets and more time thinking critically about the business problem in front of them. Better tools have always changed how professionals work. Spreadsheets replaced manual ledgers. Statistical software replaced hand calculations. Programming languages automated repetitive analysis.
AI belongs in the same category.
What it does not change is responsibility.
Someone still needs to decide whether an assumption is reasonable. Someone still needs to notice when two datasets contradict each other. Someone still needs to explain why a valuation changed after new information became available. And someone ultimately signs their name to the work.
Those responsibilities have not disappeared simply because the tools have improved.
Confidentiality Deserves the Same Attention as Capability
As AI becomes part of everyday workflows, another question deserves more attention than it currently receives.
What happens to the information being analysed?
Financial engagements often involve internal forecasts, pricing strategies, contracts, transaction documents, customer information, and other commercially sensitive material. Before incorporating AI into those workflows, firms should understand where data is processed, how it is stored, who has access to it, and whether the process aligns with client obligations.
The discussion should not begin with model performance.
It should begin with governance.
The Value Has Never Been the Spreadsheet
There is a temptation to think that if software can produce a financial model more quickly, then financial analysis itself has become less valuable.
I think the opposite is true.
The mechanics of analysis have become cheaper. Judgment has become more valuable.
Anyone can generate a spreadsheet. The difficult part is deciding whether it reflects economic reality, whether it answers the question that was actually asked, and whether another professional could review it and arrive at the same conclusion.
That is where trust is built.
AI will continue to improve, and so should the way we use it. Firms that ignore it will fall behind. Firms that rely on it without disciplined processes will face a different problem.
The opportunity is not to choose between AI and human expertise.
It is to combine the efficiency of one with the accountability of the other.