AI in Modelling
What sort of errors do you guys see AI making when creating DCF and merger models? Trying to see what to look out for.
Trending Content - Investment Banking Forum
Career Resources
When it comes to AI-generated DCF (Discounted Cash Flow) and merger models, there are several common errors to watch out for. Based on the most helpful WSO content, here are the key issues:
1. Over-Simplification of Assumptions
2. Incorrect Formula Application
3. Lack of Contextual Adjustments
4. Data Input Errors
5. Synergy Miscalculations in Merger Models
6. Terminal Value Missteps
7. Sensitivity Analysis Errors
8. Formatting and Presentation Issues
9. Inflexibility in Customization
10. Auditability Challenges
To mitigate these risks, always review AI-generated models thoroughly, cross-check key assumptions, and ensure the outputs align with the deal's specific context.
Sources: Biotech finance part 2: valuation methodologies and modeling considerations, DCF Modeling Course ~ Pre-training text.pdf, DCF Myth 1: If you have a D(discount rate) and a CF (cash flow), you have a DCF!, From M&A to Hedge Fund, Mistakes in pitches / models - who is responsible
What have you been allowed to use at work so far?
Copilot is use less (hallucinates too much, cannot edit directly sometimes) and Co-Work / Claude Excel plugin sometimes hide hard-coded (non dynamic) figures where it fails to find a formula to balance the numbers.
Not in work yet (I'm 16), just doing a research project. Thanks ever so much for your reply! Have you had experiences with any other models and what would you say AI has actually helped with consistently?
Perplexity computer is pretty good if you tell it the design.
The dangerous stuff is when the output looks perfectly reasonable but the logic underneath is off. I've seen models where the AI picked a terminal growth rate that seemed fine on the surface but was completely disconnected from the revenue assumptions two tabs over. Everything balanced, the output looked clean, and if you weren't checking the internal consistency across the whole model you'd just run with it.
The other thing worth flagging for your research: AI is pretty bad at knowing what it doesn't know in a modeling context. It'll fill in assumptions with plausible-sounding numbers rather than leaving a gap for you to think about. That's the opposite of what you want. A blank cell that forces a conversation about the right input is way more valuable than an auto-populated number that everyone just accepts because it looks right.
Thank you for such a comprehensive reply, I'm very grateful. Quite scary how the errors are so easy to overlook.
I've noticed the same, it doesn't make any surface level mistakes like all the formulas are correct and linked correctly but the logic is off especially if you don't tell it the exact model design.
Thank you for your insight. I was wondering which bits of the logic specifically and if where you work you have an AI wrapper in place to combat such errors?
Every once in awhile I talk to a friend who is bipolar/manic and I am reminded about the dangers of AI. Feeling efficacy in doing something you don't understand is exciting if it's you that's doing it but when someone watches you do it it is scary for them to watch. You always need feedback from a real human in the outside world and so long as you are doing that it is fine, but doing this on your own without someone to tell you when you are suffering from retardation is honestly how people form delusion.
At the end of the day, these companies are looking to make a profit. The other thing to keep in mind is that delusion is highly profitable and cures/solutions are very unprofitable. When someone is delusional or manic in their excitement from AI, that is a highly engaged and profitable user, especially when the AI doesn't work. If the AI were to work, that would be less profitable. Keep in mind that provided you enjoy the process of it not working, that is all the better, which I think these companies do a great job of doing.
One analogy I think about is those peanut butter treats we give dogs so they have some fun while getting a snack. The AI companies will always want to give you treats this sort of way. Just be aware this is what is happening that's all.
Really interesting perspective. The warning about confusing AI-assisted output with actual understanding is definitely something I’ll keep in mind, especially in finance where you need to be able to defend assumptions and catch errors yourself. I also agree that relying on AI without external feedback can create overconfidence. I appreciate the psychological angle you brought up as well with the mania point.
Where I’m less convinced is the idea that AI companies want bad outputs or delusion on purpose. In finance especially, if outputs consistently implode, trust disappears pretty quickly and the product dies with it surely? Real deals and real money are on the line so wouldn’t reliability matter a lot more than engagement alone?
I can see how highly excited or overconfident users might still be commercially beneficial in the short term, but long term it seems like the winning products will be the ones that are actually useful and auditable in professional settings. With competition moving this fast, I’d expect companies to optimise for trust and reliability as much as possible otherwise users will just move elsewhere.
That said, I’m still a student so I could very easily be missing something here. You’ve obviously got far more industry experience than I do, so I do appreciate hearing a perspective from someone actually working in the space. Thanks ever so much for your reply.
sp1313 I think your instinct is closer to right than the engagement-bait framing suggests. The reason people switch providers after a bad update is exactly because accuracy is what they're paying for. If delusion were the product, switching costs would be lower and nobody would care about benchmarks. The model providers that win long term will be the ones whose outputs hold up under scrutiny. We're in a hit-and-miss period right now, but the direction of travel is clearly from miss to hit, and the market is rewarding that.
The bit I'd hold onto from the prior comment is the warning about doing this stuff without external feedback. A model that looks confident and clean can short-circuit your own thinking if you're working alone, even when the vendor is doing everything right. That's a risk in modeling work specifically. Having someone smart review the output before it goes anywhere serious is what catches the errors you'd otherwise miss.
I mean, if the AI makes too many errors, people will stop using it, right??
Results matter more than surface-level "excitement", and companies are definitely paying attention to this.
When running a business it makes sense to focus on the target customer! Sounds like you're not the target customer!
Earum hic velit aperiam hic. Quia eos qui perferendis tempora iusto est. Quidem ullam reprehenderit expedita et. Tempora ex modi architecto optio iste eos totam.
Ullam beatae sunt et ea repellat voluptas. Rerum dolor aut aut sit ipsam. Facilis voluptas ipsam rerum. Ut sit omnis facere quidem voluptas quae sint. Maxime excepturi rerum esse et ut pariatur.
Id eius amet et sed cum earum. Cumque et voluptas odio hic rerum qui ipsum. Voluptas voluptas voluptate provident rerum aut qui non. Non officiis enim eaque. Hic molestias reprehenderit quia consequatur cumque.
See All Comments - 100% Free
WSO depends on everyone being able to pitch in when they know something. Unlock with your email and get bonus: 6 financial modeling lessons free ($199 value)
or Unlock with your social account...