German tax systems present a major challenge for automation due to unstructured paper documents, contradictory rules, and high legal stakes for errors.
Success is defined by the 'touchless booking rate,' which tracks the percentage of transactions processed with zero human edits, achieving 97% automation.
Avoid a single large language model prompt by using a deterministic graph for process flow and invoking specialized AI agents only for judgment-based tasks.
Every human interaction with the system is captured and transformed into a persistent, structured rule to prevent the same issue from recurring.
A layered approach efficiently matches invoices to transactions by first using cheap deterministic filters, then a machine learning reranker, and finally an AI agent for the most complex cases.
Key principles for building agents in high-stakes environments include specializing agents, making loops predictable, learning from user actions, trusting data, and pruning cheap.
Jobs that call for the skills explored in this talk.