What are you actually using AI for in PE?
My managing partner is convinced you can basically run a deal in Claude/ChatGPT at this point. I'm way more skeptical. It can't get a decent comp table without hardcoding half the numbers, hallucinates figures, etc.
Where it's actually saved me time:
- Deep research on a company/sector/person before a call
- Summarizing CIMs or data room docs
- First draft paragraphs for memos or IC write-ups
- Turning messy notes into something readable
But it's nowhere close to end-to-end, more like a summer intern you still have to check every single output from.
Curious what everyone else is seeing:
- What are you actually using it for day to day, if at all
- Which tools, ChatGPT, Claude, Copilot, or something more niche
- Anyone gotten it to do real modeling work reliably or is that still a pipe dream
- Is it mostly used in sourcing, DD, portfolio monitoring, LP reporting, etc
Routinely use the Claude excel plug-in for first passes on datacuts and CIM reviews. Used to be very averse to using it in place of what I thought to be critical thinking. At the end of the day, these 2 tasks are somewhat repetitive and robotic so find it easier to get behind now.
Don’t see any transformational use cases from the technology but can see this changing in the coming years.
Tried a bunch of tools but for the most part, back to doing 90% of workflow the old fashioned way. Main real use case now for AI is to get up to speed on an industry (and maybe even a company) I am not familiar with. For everything else like CIM summaries, excel work, I personally found the risk of hallucinations mean that I need to cross-check and tie every number. And at that point, I might as well do it myself and understand the source material better.
you can also start fresh session to audit your current models without getting biased on its previous work
preferably should also have prompts to actually show you the calculations outside the models in its answers, and if the main lines match, should be good to go
can also cross-run GPT to see if you get a different answer
Think you're a bit too old school
"preferably should also have prompts to actually show you the calculations outside the models in its answers, and if the main lines match, should be good to go"
"can also cross-run GPT to see if you get a different answer"
Wtf are you on about?
Why do this if it doesn’t make you faster or smarter. These are extra unnecessary steps trying to coddle the LLM into not being a dumbass. I will believe in ai supremacy when I can shout unintelligible garble at it with an attached 4 year old quality markup like I can for a 1st year associate and in return I get a 98% perfect output with only cosmetic tweaks. We’re no where near that yet
Ignore title. You guys need to start using claude code or codex. I run entire models (with traceable, accurate formulas, built perfectly to how we like to cut and format data), build perfectly formatted 30+ page PPTs (little to no last mile formatting needed), have a live Q&A with the full data room materials, etc.
It has massively improved my quality of life. The only thing you need to get this right is high quality precedent examples (which any established firm has tons of) and the will power to refine skills / sub-skills
This is exactly right. LLM is just an engine you have to build the tools yourself
How do we do this? Seeking advice
You need an IT admin that’s willing to let you use the coding tools with your firm files (easier said than done). Would ask the partners and tell them about the ROI potential. Once you have the all clear, a monkey could automate all the workflows
Do you mind elaborating?
- What do you feed in?
- Are your prompts in NL or code?
- What do you get out of it? Where does if stand in terms of accuracy?
Yes, and let me also give a bit of a primer on why so many people who are "AI adopters" think these tools are not good or reliable. The key to getting everything to work: 1) lots of context / skills and 2) deterministic code supporting workflows.
For point #1, this is why Claude Code / Codex and the Opus 4.6+ class of models kicked off a massive enterprise buying frenzy (predominantly for actual coding). With Opus 4.6, you can hold 1 million tokens worth of context, which can be an entire code base in many cases. And then you also had Claude Code (and Codex) launch, which are harnesses that let you make skill files to give very granular instructions about specific tasks (which relies on having a big enough context window to work). A simple finance analogy might be: you want to take raw monthly financials and show well-formatted annual ones with add-backs, etc. -> you give Claude Code 5+ full example workbooks from other deals that show different ways of doing it -> Claude Code can look at those and retain all the formulas, logic, etc. in memory (effectively) -> you will get an accurate output (albeit, you often need to go through some painful back and forth to refine the outputs the first time you make a skill). When you do this, you have Claude create a skill that can be called something like /annual-financials, which will retain all the context of the task, any feedback you gave it about the outputs, which reference files to look at, how to validate accuracy and produce checks, and anything else you tell it related to that relatively narrow task. In the future, you can "call" that skill, and it will be able to replicate that task with a very high degree of accuracy. So, in summary, you: 1) tell Claude what you want it to do in natural language, 2) give it a ton of high quality reference examples, 3) have Claude Code build a skill around the task, 4) go back and forth with Claude until you refine the skill to be exactly right. Then you have permanently mostly or entirely automated that workflow. Carrying the example forward, any time you need to roll monthly financials to annual ones (in the way your firm or even specific partners like to see it), you can go scroll Instagram instead of writing Excel formulas by hand.
For point #2, as you are making a skill, Claude Code will automatically start writing deterministic code on your behalf that supports a given workflow. An obvious point that's worth making is that code runs reliably with traceable logic (i.e., same thing every time). If you take PowerPoint as an example, you can say "here are 5 examples of the slide I want you to make, please copy and paste the literal shapes from the slide, update the text, swap out this chart, etc.". If you ask for this, Claude will write actual code that says: "look at this example deck, copy this shape, paste it at this X and Y position on the page". It runs perfectly every single time. So you now have both legs of the stool: you are using inference (LLMs) where there is real ambiguity or thinking/research needed (e.g., do web research and tell me who this company competes with, then write a summary about each) and deterministic code firing when there is not (e.g. you write this excel formula to do X, Y, Z and need to make the text bold)
Over time, you can keep making more and more skills, automating a substantial share of manual work -- all with natural language prompting and 0 actual coding ability. It is perfectly accurate in most cases (and you can check, for example, Excel formulas to verify this). There is basically no limit to what you can automate (within the confines of finance workflows), outside of the time investment required to stand up more skills.
If you juxtapose this against the person who just uploads a few PDFs into the Claude website interface and asks it a question, it is an ocean of difference. The web instance has basically no context (on a relative basis) -- even if you upload the max 10 files (or whatever the max is), it is very difficult to have learnings or work translate across different sessions / tasks (so you keep "losing" work and have to re-teach Claude or ChatGPT, which is super painful and annoying), and you cannot write real deterministic code to support tasks. I recon that this is the vast majority of people who think they are "using AI", and it's a big reason behind why you see so much negativity around capabilities / this sentiment that AI is overhyped. It is not, and the best firms (and companies) of the future will standardize these skills enterprise-wide, and see material efficiency gains + QoL improvements
It did first draft of funds flow, impressive stuff. Does a lot of BS I hate doing
For a first draft, LLMs rock! Watch out for hallucinations on all of the LLMs. Claude/ChatGPT are great for getting a framework, but you really need to check all of the facts, even if they source it. Like Harvey is for legal, and OctumAI is for investment/finance, the SLMs do a bit of a better job. Co-Pilot sucks, but is pushed by MSFT. I think the large LLMs will encroach on the SLMs and niche workflow players since they have the capital and cash and data to recreate their products.
I think more people in PE need to realize that LLM's work better if you feed them in some sources of truth and analysis. If you ask an LLM to cite from a specific PDF of a CIM or industry report and past examples of outputs that you are looking for, it has a significantly lower chance of hallucination and you can usually do a quick click to verify. Just need to be explict in prompting that it is to not use outside sources. It is also generally helpful for things like writing emails or rewording points from thoughts into more professional phrasing. Also is helpful to give it a role of what it should act like and generally as much context as possible if you are asking for a more complex task. For templated modelling, I have also found it's great if you do the step of feeding in the relevant precedents.
I sure wish my new associates would use AI significantly less and actually learn how to do the analysis I'm asking them to do instead of continuously sending AI slop. It's going to cost at least one of them their job
It is useful for research, summarizing documents like CIMs or data rooms, and drafting memos, but it cannot handle reliable complex modeling and always requires manual verification.
Associate 1 in Consulting has the right idea about how to use Claude Code. I've been using it myself in the same manner, and it works beautifully.
This does involve having to be a bit of a product manager though, so you can create your own slash-commands. You need to sit your ass down, and think through exactly what you want to achieve, and explain to the AI step-by-step (as you would your grandma) how it gets there. For example, I created a sourcing workflow that is multiple slash-commands which essentially starts from giving it a 4 digit NAICS code down to identifying individual companies. For now, it generates ppts, but it can also generate xls and basically anything else you want.
The main pro of this approach is you can really tinker with it and make it yours. The limits are your creativity. The robot adapts to what you want, and you can update your Claude.md file with a style guide that makes it sound more like you vs. AI.
The one big con is that I do believe you need a fair amount of judgment in terms of prompting it the right way. If you're someone just getting started in investing, you likely need to cut your teeth doing things manually before you can create the right workflow. It shouldn't take you forever, but it's worth really, really understanding what you're doing vs. just chucking CIMs into Claude or one-sentence requests. This is also why many folks higher up in the food chain get annoyed af with AI-written junior output. The same way you learned to do multiplications by hand but now pop open an Excel for big calcs, you will benefit from being a sponge and crafting your judgment the "hard" way before you can really use the tool.
Oh and you need to get over the initial aversion of using the terminal. Yeah the first couple of days suck as you get used to the interface and the controls (still drives me mildly crazy that I need to press PgUp instead of arrow up to read what it's outputted beyond the current screen), but it's well worth it.
Dead on and great adds. It takes time to do well (a very big J curve in my experience) and you cannot do any of this if you don't already know what you're doing. You see this with new joiners and interns, who are sometimes proficient with the AI tools but do not know what good looks like. It is detrimental to them.
And the terminal is exactly right. Works much better than cowork or claude code via the desktop app (from a pure usability standpoint). But the real reason is that it lets you parallelize work by having multiple tasks running at once
I am not so much in the weeds of excel and powerpoint anymore, so the biggest value add for me has been with legal docs -- using chatgpt to help translate, break down, summarize, provide ideas on reasoning and implications, etc. I don't think feeding it an entire legal doc with "summarize" is a good use of time, but it can be extremely helpful for certain sections, clauses, etc with the right prompt. Huge time saver for someone like me that despises this type of stuff.
Has anyone tried tech startups Endex or Rogo AI? They help with excel + modeling as an analyst.
- I ask this cuz they seem to be raising a lot of money from a16z and Peter Thiel.
- do you guys think they might be the future? Or just some ai wrapper slop.
They are going to get wiped out
Nothing they do cant be done by claude code with some set up
So true so true, they are all cooked. Why would anyone use a third party source when there is open ai or claude already better + more secure.
My firm has a lot of IT governance blocking it off. But I use it heavily for drafting / proofing emails and for reviewing memos/presentations. For the latter, I ask it to proof the deck, find discrepancies in the numbers, highlight points that may be weak, and identify obvious questions a new reader may have.
Fable 5 has basically changed the game for me. I can spend more time doing more useful work like expert calls, talking to management teams / bankers, and portco management. Anyone not using AI is going to fall behind severely. I double check all my work before sending out but the error rate has basically gone to 0 with the way I’ve set things up (combo of skills, multi-model work flows, and dashboards).
I use AI constantly for finance—everything from analyzing earnings calls and building financial models to researching industries, competitors, and market trends. The biggest challenge isn’t getting answers anymore; it’s keeping all the context organized across dozens of reports, filings, spreadsheets, and conversations.
That’s why I’ve been using Trustgraph. Instead of treating every prompt like a fresh start, it connects information into a context graph, so relationships between companies, metrics, assumptions, and previous analyses persist over time. It feels less like searching through notes and more like building a knowledge base that AI can actually reason over.
For finance work, that means I spend less time re-explaining context and more time analyzing the business. It’s become one of the more practical open-source AI tools I’ve used because it improves the quality of long-running research instead of just making individual prompts better.
Dolores sed omnis sed minima vero harum. Expedita illum laboriosam mollitia sit. Et dicta veniam quia iste. Aut nostrum non dolor error autem architecto.
Sit laborum error totam et et veritatis et. Repellendus nihil ullam soluta repudiandae voluptas. Laborum sed et eos tempora eaque placeat. Incidunt et corrupti exercitationem illo nemo ipsum architecto dolorem.
Perspiciatis dolor corrupti est. Esse vel itaque commodi labore consequatur.
Quis ad iusto expedita ut. Accusamus facere a praesentium accusamus. Est reprehenderit aut iure non. Consequatur ut aliquid voluptatem id maxime eaque. Quam non praesentium magnam qui asperiores velit.
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...
Ut beatae incidunt dicta et at recusandae. Quibusdam quidem aut non totam dignissimos. Id ea quasi magni dolores non nesciunt assumenda. Et eaque animi quibusdam repellat facere in. Est id eum consequatur qui rem nisi sed.
Assumenda dolor quos non enim ut. Ut distinctio qui ipsum quia. Ratione et illum labore ut dolores nihil nihil. Eius omnis fuga porro et minima. Aliquid exercitationem voluptate velit ex. Placeat aut cum rerum beatae natus dignissimos.
Est nostrum quia non voluptas tenetur sed soluta. Necessitatibus fugit temporibus alias vero nulla pariatur. Sed aut ullam incidunt omnis. Occaecati suscipit tempora et deserunt nihil.
Et ipsam placeat illo nihil est. Odit quos asperiores et. Occaecati consequatur cumque dolorum voluptatem dolores.