When we show Structrz to prospective clients, we often are asked if AI will be able to perform all of this complex financial analysis on its own. While we are power-users of AI for programming and have a deep understanding of the particulars of the technology, we are skeptical that AI will be able to provide the value it has for language without some standardization first.
Most spreadsheets have different ways of setting up inputs, linking formulas, and naming conventions. We think AI will be able to assist in providing some evaluations of spreadsheets, it will be hard to train an AI on examples and then ask it to create spreadsheets based on a new set of facts. It will have the same issue as when you ask a junior analyst to create a spreadsheet; it will probably get a lot right, but will miss the subtelties and ability to make judgements about granularity. You will be left with an analysis that you will not have confidence in and will likely be hard to scale as requirements for a given analysis evolve.
While we are not actively developing Structrz as an AI driven tool, the underlying architecture and standardization should provide the right level of generalization to make this the right tool for leveraging AI. Most AIs are going to bue consumers of APIs. Excel as an API is too general for an AI to simply build a financial modeling spreadsheet, even if trained on lots of examples. Better would be to have an API on top of Structrz that and AI could review and use an LLM to determine how a user’s input could be translated to use the Structrz API.
For instance, an AI agent could review the possible protoype schedules we have and review the user’s input to see which might fit best based on linguistic matches and how the inputs and outputs might be constructed. With this, the AI agent could start constructing incrementally more complex models by feeding the structured model context back into itself to further the user’s requirements. After using LLMs so extensively and seeing how good they are at language versus tabular formulas, that the LLMs will be better at interacting with an API to describe a spreadsheet versus parsing large sets of evergrowing formulas that do not use clearly defined architecture. Also, becuase Structrz schedules have a limited set of isolated input variables and inputs streams, each building block of a model can get away with a smaller context, which should improve results.
Time will tell and we are certain that programmers will try to come up with spreadsheet readers and spreadsheet training data, but we don’t expect these efforts to replace the power of AI coupled with fully thought through spreadsheet architecture.