This is not a how-to and teaching post. I am just sharing my journey with AI and financial modeling up to date (July 2026). When I first started using AI assistants integrated in Excel, it was pure joy. It reminded me of the old animated assistants in Excel back in the early 2000s. I loved them, and since then I was longing for a similar thing. Especially for a solo financial modeler who is spending most of the time in Excel, it’s just such a value-add to have an assistant to brainstorm things when you need it. So I still think the best use case for AI in financial modeling is to use it for audit and brainstorming things that you are already familiar with.

So let me tell you how I use it:

AI for Financial Model Audit and Review:

When I review a financial model, I already have a checklist in my mind of common things I need to check in the model. I have produced that checklist that I share in my Project Finance Review courses. So I converted that checklist into a skill in Claude. Now, when I want to review a model for the first time, I use that skill to generate a sheet within the model, go through the checklist, and produce a list of issues on the model.

I have already tested the skill with dozens of project finance models, and the end list is a good starting point for me to finalize my review. Once the AI produces the list, I will then need to check all the issues raised one by one. In the process, I also catch other issues that were not in the list and complete the issue list.

Can AI decide on the quality of the model? Meaning, can we rely on the AI to make the final decision to maintain or rebuild the model?

My answer is no. Why? This is a subtle decision and it depends. If it is a bad model from the start (meaning it doesn’t have any structure and is not following the best practices that make a model flexible and transparent), then the decision is easy. Any experienced financial modeler can open an Excel file and tell you straight away that an unorganized model is a waste of time and better rebuilt from scratch. However, when a model is well structured, most of the time AI will tell you to maintain the model without considering some subtleties, like how long it takes for a model to run. You might have a well-structured model, but if in order to run a debt sizing or tariff setting the model takes 10 minutes to run, then that’s a problem. So you might end up with a black-box model that looks good on the surface but can be difficult to maintain and work with.

The final verdict on using AI for model review: Definitely useful for a start, to check for common issues, but needs human eyes to go over the checklist and complete it and make the final verdict.

AI for Building the Project Finance Model:

I have done a number of trials on letting the AI build the model from scratch, by feeding in different skills for the model mechanics and providing the relevant project documents. Before I explain the process, my final verdict was that the result is not much different than starting off with a template. I already have multiple templates for different types of projects. So when I am asked to do a back-of-the-envelope analysis on a combined-cycle gas project, I just pull up that template and adjust the inputs to match the project and get a ballpark view of the main metrics. Then as things get more specific, I need to adapt the template to the project. So, personally, I do not see any added value in using AI to build it for me from scratch. Also, there are a number of rules that I have that AI fails to comply with.
I want my models to be done in a certain way. There are things that are a must for me and I cannot compromise on. I created a Grand Rules Skill for project finance modeling that lists all the rules and standards that I want my model to have. I also have separate skills for creating the Input sheet, with instructions on how I want the input sheet to look like. How I want the Calculation sheet to look like. The Check sheet. The Summary sheet. The Financial Statements sheet. A skill for the model timeline, flags, and Structure. So I have already created all these personalized skills to give exact instructions on how I want the final model to look like, at least in terms of structure.

I will make a video and maybe a seperate blog post about building the first draft of your project finance model with Claude and my process (Skills, Prompts…).

Things AI Failed to Apply in My Models:

We all agree that inputs should be separated from calculations and outputs. So I have a specific sheet where I store the inputs. But when I am using the inputs to calculate different calculation blocks in my calculation sheet, I want the assumptions to be next to the calculation block and then used in the formula. So that when I want to check the formula, I just press F2 and I see all the elements of the calculation within that same sheet and don’t have to do a zigzag to understand how any item is calculated. That is something that AI failed to respect. You might say my instructions were not good enough, but I tried a number of ways and it just cannot consistently do it.
AI failed to use a consistent formula across the row. Even though this is in my grand rules and I also specify in a prompt to respect this rule (that every row in the model should use only one formula so that I can copy it across), still, there are a number of cases where, in the middle of the timeline, the AI-produced formula is not consistent. So if you are using AI to build a formula in a row, copy it across and see if the sum changes.
AI fails in keeping formulas simple. I also experienced that in VBA codes. I ask it to create a simple copy-and-paste code with a loop, and it just makes the code too long. With my limited knowledge of VBA, I know that I can do the same with 3 lines of code. When it comes to formulas, it is the same thing. Although I indicated in my grand rules that I forbid the use of OFFSET, HLOOKUP, VLOOKUP, and nested IFs, I still notice that, so many times, the formulas suggested by AI are more complex, and there are more elegant and simple ways to do the same. So make sure you challenge the output and iterate for simplicity.

AI as a Companion While Building a Model:

The best use case of AI I have found was to help me in the process of building a model.

• AI can do a good job in building the databook from different project documents. AI can do this step for you. You provide the project documents and give instructions on creating a list of parameters from the project documents, separated by type, and also provide the reference in each project document. Then you can go through the list and cross-check if the source is correct. I hope this will not make financial modelers skip the step of reading project documents and understanding the project. To me, that’s such an essential step in building a model. Unless you know the project, you can’t build even the model structure for it, let alone define the base case and make sure the model is accurate.
• Coding: One skill that financial modelers lack is VBA coding. That’s why we are so proud of our copy-and-paste codes and our optimization buttons in our project finance models. Then we end up with models that take 10 to 15 minutes to run for one iteration of debt sizing or tariff setting. Early in my career as a financial modeler, I realized that this circular issue in project finance models is a real problem, and the only way to get rid of circular references is to reprogram it. I was lucky to get the chance to work with Professor Edward Bodmer, who created this brilliant Parallel model technique, which opened a whole new portal. I have been talking about it for years, and most people understand the value of it but are scared of maintaining the code on their own. Many experts told Professor Bodmer that he is the only one who can do this Parallel model technique. But now with AI, I think that fear can be reduced. Because you can use AI to help you build that complexity in your model. Not to do it for you from scratch, same concept as building the model. It will not be able to build the Parallel model, but it can help you debug it when you face an issue or modify the code, once you understand it. We are planning to make webinars and master classes on these topics, so if you are interested, contact either me or Professor Bodmer and we’ll keep you in the loop.
• AI to generate the annual financial statements from your periodic financial statements. I find that AI is best to use when you already have a structure in place and you just ask it to replicate it. So things like building annual statements based on periodic statements can be done, but again you need to either give detailed explanations (use SUMIF, XLOOKUP…) and cross-check the result to make sure that no complex formula is used and the formulas are consistent across every line in each statement.
• Compiling all checks in the model in a single sheet. When you are building a model, you always create a number of checks, and then finally you compile all the checks within one sheet and create a master check. You can ask your AI assistant to do it for you, and of course cross-check to make sure all checks are included and the master check is a simple formula (SUM or COUNTIF).

I asked Claude to read this blog post and defend itself againts the claims I made:

Here’s the argument and my views:

Claude: Most of the post is accurate as a description of one expert’s workflow. Where it slips is in generalizing from her own setup to claims about what AI can and cannot do. You asked me to defend myself, so here is where the arguments are weakest, strongest counterargument first.

Claude: The “AI adds no value building from scratch” conclusion is rigged by your baseline, not by AI’s ceiling. You are comparing AI against a decade of your own templates, grand-rules skills, and per-sheet skills. Of course starting from a mature combined-cycle template beats starting from a prompt. But that says your templates are good, not that AI building is weak. For a modeler who does not already own a calibrated template library, AI-from-scratch is a meaningfully different value proposition. The honest claim is “AI does not beat my templates,” which is narrower than “AI from scratch adds no value.” Confidence: high.

Hedieh: These template are not specific to me. You can go to any Financial model market place and download one. Better to start with a template that has already been proven.

Claude: The factual error is the circularity claim. “The only way to get rid of circular references is to reprogram it” is wrong. Circularity in project finance (IDC depends on debt, debt depends on funding need, funding need depends on IDC; or DSCR sculpting feeding tax feeding interest feeding debt) can be broken several ways: copy-paste/macro value-pasting, Excel iterative calc, goal-seek/Solver, and crucially algebraic resolution where you solve the loop in closed form inside ordinary formulas with no code at all. Bodmer’s copy-paste and UDF approaches are good engineering, but they are not the only exit from circularity. Reprogramming reduces iteration time and fragility; it is not the sole method. Confidence: high.

Hedieh: Copy and paste is what making the models slow to run and there’s no algebraic solution for the main circularities in any project finance model. So i insist that the only way to resolve circular reference in at least project finance models is to reprogram it.

Claude: Your strongest concrete criticism is the inconsistent-formula-across-the-row failure, and it is correct. Models break formula consistency most often at period boundaries (first period, COD, end of tenor), and AI does exactly this: it special-cases the edge column and breaks the copy-across property. Your test (copy the formula across, check whether the sum changes) is the right detector. This is a real, reproducible weakness, not a prompting artifact. Confidence: high.

Hedieh: OK

Claude: The over-complex-formula and verbose-VBA criticisms are accurate but overstated as inherent. AI defaults to defensive, nested constructions and long loops, yes. But verbosity is steerable in a way formula-inconsistency is not. “Three lines of copy-paste VBA” versus a long loop is usually a single corrective iteration, not a ceiling. So this belongs in “needs challenging,” which is where you put it, rather than “cannot do.” Confidence: moderate-high.

Hedieh: Here I want to challenge the idea of building a financial model from scratch with one iteration. Offcourse most issues can be resolved with several iterations that might take more time than building it yourself from scratch.

Claude: The “F2 and see everything” failure is partly self-inflicted by an instruction that contradicts itself. You assert inputs must be separated from calculations (own sheet), then require the assumptions to sit next to the calculation block and be used directly in the formula. Those pull in opposite directions. What you actually want is input cells mirrored or referenced adjacent to each calc block, which is a specific and legitimate layout but not the same as “separate inputs.” Institutional standards like FAST deliberately do the opposite of what you want here: they discourage scattering input-looking cells across calc sheets precisely to keep one source of truth. So AI’s inconsistency is at least partly it ricocheting between your two rules. The model is not failing a clear instruction; it is failing an ambiguous one. Calling it purely an AI failure is not quite fair. Confidence: moderate.

Hedieh: I honestly refined the instruction and provided examples but still AI could not do this for me.

Claude: The model-review verdict is where you are most right and you undersell your own example. You say AI cannot make the maintain-versus-rebuild call because of subtleties like a 10-minute run time producing a good-looking black box. But run time is observable and testable, which means the failure is that the AI was not given or did not gather that signal, not that the judgment is intrinsically beyond it. The genuinely hard part you gestured at and then dropped is the one worth naming: a model can be structurally clean and still be wrong on the things that matter, base-case logic, tax and accounting treatment, whether the structure actually matches the deal. That is the real ceiling on AI review, more than run time. Confidence: high.

Hedieh: Run time is a big issue! If you have ever been in a meeting with running the model live infront of people, it’s not nice to make people wait 10 mins to run a scenario!

Claude: Two points are simply correct and need no defense. Databook extraction with source references, and the warning that it must not let modelers skip reading the documents, is right and is the most important sentence in the post. Annual-from-periodic and master-check compilation are good fits precisely because the structure already exists and you are asking for replication plus verification.

Hedieh: Ok.

Claude:  Net: your practical conclusions are sound and your detectors (copy-across test, cross-check sources, challenge for simplicity) are exactly right. The weak spots are three overgeneralizations (no value from scratch, the circularity absolutism, and one self-contradictory layout rule blamed entirely on the model) where the post treats properties of your particular setup, or one factual overstatement, as properties of AI.

https://youtu.be/FUPu2UcIYKk