Why do complex multi-step accounting skills (like an engagement letter or financial statement review workflow) keep breaking down with more exceptions and iterations?

One view expressed was that relying entirely on an LLM to handle a full process is not necessary and is likely less accurate — described as 'the worst way to go.' The suggested alternative is to do as much as possible via scripts and deterministic steps, and have the LLM (the skill file) only handle escalations — cases the script itself can't resolve. The example discussed was a financial statement review workflow (budget, engagement letter, trial balance import, inquiries/analytics, financial statement prep, issuance) that the firm tried to encode entirely as chained/iterated skills, running into an increasing pile of exceptions that made the system harder to manage rather than easier, especially after around 8 iterations.