Maximilian Eber

The through-line

Eber's recurring preoccupation is the gap between what AI can already do in financial services and what it is still barred from doing, and the practical question of what would have to change to close it. He frames his own work as navigating AI adoption in financial services [1], and his interest sits at the seams of the org chart rather than in the model itself: the handoffs between marketing and risk, between risk engineering and product, between model builders and compliance. On the technology side he is specific about the two blockers he expects to fall, hallucination and inference economics, and about the class of use cases each one unlocks [2].

On what has to happen before AI can sit in the decision path

Two developments interest him most. The first is eradicating hallucinations, a problem he notes researchers are actively publishing on, including work on "why models hallucinate in the first place" [2]. Solving it matters because it "unlocks a whole class of use cases which today are blocked because people really want very very high reliability not just 98% but 99.9" [2]. The second is the falling cost and latency of inference, which he treats as effectively certain to happen. The constraint is concrete: "if you are in the off flow of a credit card swipe, you don't have much time to actually call a model", and as a result "most of these use cases are out of reach today for modern AI" [2]. Faster, cheaper models make real-time decisioning of that kind "attackable and really attractive" [2].

On the friction between marketing, risk and product

He describes a pattern he says he hears often: marketing pulls in the cheapest leads, and risk then rejects exactly those, because they are "cheap to acquire" but "won't work just economically for the company" [1]. His interest is in how a lender actually syncs the marketing, risk and product sides of the house so that the two functions are not optimising against each other [1]. The same instinct shapes what he considers a dream tool, which he locates in the work between risk engineering and product [1].

On credit systems you cannot A/B test

Eber distinguishes one-off lending from what he calls the permanent models of credit: repeat lending, or a credit line, "which is obviously the the ultimate version of repeat lending" [1]. His point is that these systems resist ordinary iteration. A scoring model can be swapped, measured on a statistical metric and A/B tested for a quarter, but a graduation ladder that runs from a small overdraft to a large loan takes years to play out: "You can't just AB test a ladder that runs for two years and know it's a great ladder" [1]. That leaves the hard question of how to roll out changes safely and still be convinced the system works, which is where he pushes hardest in conversation with practitioners [1].

On where AI is landing in financial services now

He reads one of the most immediate wins as the collapse of the traditional BI loop, summarising it plainly as a system where AI "generates SQL queries on your warehouse" so more people can reach the same level of information without going through a specialist function [1]. He also reports seeing a lot of traction for document-related use cases, while noting that these skew toward SMB and come up more often outside the US, where the underlying data infrastructure is different [1]. When it comes to what goes from zero automation today to full automation in the near term, his framing is deliberately short-horizon, asking about six to nine months rather than three years [1]. On the governance burden around model building, he draws a comparison to medicine, where transcribing the patient-doctor conversation and auto-filling the admin gives clinicians their time back, and applies the same logic to credit: have model builders "spend more time building models" and less time documenting what they have built [1].

Takeaways

• The reliability bar for high-stakes financial use cases is not 98% but 99.9%, and eradicating hallucinations is what would get AI there [2].
• Real-time decisioning inside the authorisation flow of a card swipe is out of reach for current models on latency and cost grounds, and becomes attractive as inference gets faster and cheaper [2].
• Marketing and risk routinely pull apart because the cheapest leads to acquire are the ones risk rejects on economics [1].
• Credit lines and repeat lending cannot be validated with a single A/B test, because a graduation ladder can take two years to play out [1].
• AI writing SQL against the warehouse effectively replaces the BI loop and widens access to the same data across functions [1].
• Document intelligence use cases show real traction but skew toward SMB and toward markets outside the US [1].
• The right target for automation in model governance is the documentation burden, so builders spend their time building rather than explaining, much as clinical transcription tools return time to doctors [1].