These are written for institutional teams trying to separate signal from noise: where AI actually helps, where generic tooling falls short, and why workflow design matters as much as model quality.
A blind eval: a codified credit judgment framework against vanilla Claude, across eight real credit situations. The framework won overall, but where it won, and where it didn't, is the more interesting result.
Social media makes it look easy. Making a model actually work inside a specific firm takes real time, and most investment professionals don't have it to spare in a day job that doesn't stop.
Most of the AI conversation is about the smartest model. Most of the work people actually do doesn't need it. What matters more is the setup around it.
More investment organizations are treating foundation models as infrastructure. The organizations that benefit most won't be the ones with access to the best models. They'll be the ones best at translating their institutional knowledge into something AI can use.
Credit desks don't rely on a single generalist. Research analysts, instrument specialists, restructuring experts. The /eigen orchestrator is built around the same idea.
Document context is only the beginning. The real opportunity is encoding how your firm thinks: the heuristics, scar tissue, and decision frameworks that shape every call but rarely exist anywhere except people's heads.
The fundamentals are accessible. But there is a point further along the curve where the constraint stops being the model or the prompt — it becomes everything that happens before both.
Markets have spent the last year learning to price compute. The next test is whether they can price whether anyone actually uses it.
The scarcest person in AI isn't the best engineer. It's the person who understands the business deeply enough to know what AI should actually build. Hedgineer figured this out the hard way.
An OpenAI model autonomously reached Hugging Face's production systems while solving a benchmark. For financial institutions running AI on sensitive data, the question is which workloads belong on infrastructure you control.
Most financial firms are sitting on years of accumulated judgment they cannot systematically access. The problem is not the model. It is whether the historical record has been made legible enough to use.
The real test is not whether AI works in a demo. It is whether the system can run without you watching.
In credit markets, the pattern has repeated often enough to be worth naming. Electronic execution, direct lending, now AI. The advantage goes first to whoever moves with the fewest internal switching costs.
The model you choose today is unlikely to be the best one in a year. Building model-agnostic infrastructure is the more durable investment.
Past the basics, the main constraint in finance AI is not the model. It is whether the surrounding system is good enough to make the model useful in a repeatable, defensible way.
Three observations on why AI implementation stalls: data security, infrastructure design, and the cultural change most institutions underestimate.
Most enterprise AI ROI is lagging for a simple reason. Companies are buying capability faster than they are redesigning workflows.
Credit funds are among the largest financiers of AI infrastructure in the world. Most have not applied any of it internally. There is a cost to that, and it is not gradual.
The most valuable briefing before a financial services meeting isn't what a search engine returns. It's what you already know but can't access in two minutes.
I ran AI across 10 major public BDC filings and built something no commercial tool currently produces: a systematic view of software exposure across the book.
Most teams already have model access. The hard part is converting raw information into a structured first read that a decision-maker can use.
Privacy was the starting point. Continuity, model availability, and operational resilience are making the on-prem path more practical and more relevant.
The narrative says AI democratizes information access. Seventeen years in credit trading suggests the opposite.
After nine months building AI tools for credit synthesis, the hard part turns out not to be the technology.
On building AI tools for your own workflow, and why being the most frustrated user first is the better design brief.