金融 AI 的最后一公里：可靠数据、行业审阅与不能被分数代替的判断

Kim says financial reasoning training feeds into the main model, so developers can build on that shared intelligence rather than assuming the first-party product has an entirely separate financial brain.

A usable product surface can unlock capabilities that already exist, as Kim's internal adoption example links a desktop release with a sharp increase in non-engineering use.

Bundled data can justify its cost when existing connectors fail often enough that the product team must control the retrieval experience end to end.

A vertical product may depend on a small but decisive layer of domain-specific data, interface choices, and integrations built on shared platform primitives.

As agents become more autonomous, their permissions and review paths should resemble the graduated access and supervision used for new colleagues rather than relying on an undefined human in the loop.

A factual error that changes a conclusion demands a different safeguard from an unattractive slide, even though both can reduce the usefulness of the work.

Professional taste can be trained with expert comparisons, but useful financial work still requires both verifiable checks and preferences about how the result is structured and behaves.

Moving from junior execution to investment judgment makes the learning target less clear because a promising idea may require years of outcomes and unusually strong expert feedback to evaluate.

A high benchmark score is only useful when the questions, answer keys, and grading rules remain valid proxies for the professional work the product is supposed to improve.

Domain experts and product managers become central when the team cannot assume that its model trainers already know what excellent work looks like in the target profession.