Give AI context to build better financial models

Current LLMs can struggle to create waterfalls even with fairly simple structures, likely because they simply do not have sufficient knowledge of the structures to accurately calculate waterfalls without help. If you know the structures, you can definitely use AI to build waterfall models. But be aware that the LLMs often make wrong decisions on interpreting structures and will require your corrections to guide them in the right direction.

Earlier this year I tested a few LLMs to build a cap table model and a fund forecast model and the results were not great. But over this year, I've learned that asking an LLM to build a model and critiquing the result is less interesting than thinking of how they can improve our models and how we can improve our use of the tools.

Here's a few thoughts on emerging best practices in using AI in financial models, specifically with cap tables and exit waterfalls, but likely applying a bit wider.

1. Use AI to audit a waterfall. Start with an existing cap table and ask it questions like "What looks unusual?", "What could create problems in an exit?", Are there ownership, option pool, SAFE, note, or preference issues I should understand?". LLMs are good at looking across a set of information, finding relationships, and helping you understand what deserves attention. Instead of staring at a waterfall and trying to figure out where to start, you can ask the model to help identify the areas that deserve investigation.
2. Use AI to model a waterfall. Take an existing cap table and waterfall and ask the model to create a liquidity distribution. For a simple set of common and preferred shares with vanilla rights, it should be fairly straightfoward for the AI to get it right, but as the structures get more complicated iterative prompting and corrections may be necessary. Simply exporting a cap table from Carta and dropping the data into an LLM to build a waterfall has not worked particularly well for me. The problem was that the model did not always know which information mattered, where to find it, or how the different pieces of the export related to one another. Well structured existing cap tables and data on new rounds provide better context for better results.
3. Give the model a structure to follow. If you already have a model showing how a waterfall should work, give the model that structure and ask it to apply the same logic to a new round. The more structure you give them, the less they need to improvise.
4. Use review loops, not one-shot prompts. You can also improve reliability by having the model review its own work. Instead of just asking for the LLM to perform a calculation or task, prompt the LLM to do the task, review, recalculate independently, compare, identify discrepancies, and resolve. Ask the LLM to use subagents to do the calculation 5 times, then compare. Using multiple agents do not magically make the answer correct, but it adds checks to a process that otherwise depends on a single probabilistic response.
5. Context is part of the calculation. Instead of relying on the LLM to know how everything works, provide is extra information or context so that it knows the specific treatment or best practices for this situation. Skills, MCP servers, connectors, these can provide specific detail for the LLM to use, instead of canvassing everything it knows.
6. Build a persistent cap table and waterfall environment. Instead of starting with a blank chat every time, create a persistent project for liquidation analyses. Inside the project, give it instructions explaining how you want calculations performed. Give it examples of correctly modeled transactions. Connect it to the relevant data and calculation tools, which could include your cap table management provider's MCP server, an internal database, an external best practices resource my MCP server is an example, any source of information of data that you use to make the environment better for you to use. Track what happens, and when you discover that the model misunderstands something, update the instructions. This creates a layer that can it across the cap table, legal documents, financing history, and investment data and help you understand what is happening, model what happens next, and check whether the numbers actually make sense.

Combined, these can help us use LLMs better to audit and build our models.