THE TEARDOWN
Fifteen million euros, or 3% of global revenue. That is now the price of a credit model that cannot show its work.
The Structural Signal
The EU AI Act's high-risk rules took effect on August 2, 2026. Credit scoring for individuals sits on that list.
So does pricing for life and health insurance. Trading algorithms do not.
Read that again. Brussels put the loan officer under a rulebook and left the trading desk alone.
The penalties are written into the law. Up to 15 million euros or 3% of global turnover for a high-risk breach.
National market surveillance bodies enforce it. The European Commission's AI Office handles general-purpose models.
The Mechanical Breakdown
Here is what an AI credit decision actually looks like inside a lender today.
Step one is data. The model pulls bureau files, bank transaction feeds, device signals, and the lender's own repayment history.
Step two is the score. A gradient-boosted model ranks default risk. It uses thousands of features, where a scorecard uses a dozen.
Step three is the cutoff. The lender sets a threshold, and applications above it get approved with a price attached.
Step four is the reason code. If the answer is no, the applicant must be told why, in specific terms.
That fourth step is where models break. A tree ensemble has no native explanation.
Lenders bolt on a second system to explain the first one. It is a guess about a guess.
Step five is new as of August 2. Every decision needs an automatic log and a named human who can stop it.
Monitoring does not end at launch either. Lenders must watch live behavior and report serious failures.
Step six is the paper. Technical files, a risk record, and a sign-off that the model is safe before it goes live.
Nothing in that chain requires a person to read an application. The person is now the auditor, not the underwriter.
The end of the dollar as you know it
The downward slide has begun.
According to new research from Bloomberg, the U.S. dollar's share of global reserves has just fallen to the lowest level this century.
While everyone is distracted by hyped-up IPOs and the AI bubble, the world is walking away from the dollar – the foundation on which all of our lives are built is crumbling.
And I believe the consequences for the country – and your financial security – are extremely serious.
President Trump knows it. That's why he has taken emergency action by signing executive order 14241 to initiate the first full reset of the American dollar in half a century.
That means every dollar you have saved and invested… every good, every service, every asset… all of it could be about to be repriced against a new monetary anchor.
It’s not gold, or crypto – but something far more unexpected. An asset so fiercely contested and so critical that Vladimir Putin once claimed whoever controls it “will become the leader of the world”
Nobody can tell you exactly how this reset will play out.
But I do know that the last time America changed its money like this – half a century ago – it split the country in two. Between the folks who understood what was happening and responded accordingly – and those who got brutally left behind.
That line is being drawn again. And what you do with your money in the months ahead could decide which side you end up on.
I’d like to show you which investments could thrive – and which could be the most dangerous – inside Trump’s new monetary order.
Legacy vs. Autonomous
The legacy setup is a scorecard plus a human. FICO-style models use a small, stable set of variables that anyone can explain.
Humans handle the edge cases. Thin files, self-employed income, and recent job changes go to a queue.
The trade is speed for defensibility. A scorecard is slow to update but easy to defend in an audit.
The machine setup wins on accuracy. Klarna told the SEC its model beat the VantageScore benchmark in the US.
It claimed more than twice the predictive power on default. That is Klarna's own estimate, filed in its 2025 listing papers, using December 2024 data.
Where machines fail is narrower than skeptics claim and worse than vendors admit.
Drift is the first failure. A model trained on 2023 borrowers misprices 2026 ones, and nobody sees it until charge-offs move.
Explanation is the second. Reason codes made after the fact are approximations, and the AI Act now demands they hold up.
Judgment on strange cases is the third. AIG's Peter Zaffino spoke in May 2026 about claims work.
He put Claude at 88% as good as an expert. Remarkable, and still not a number you want deciding the last 12% alone.
Capital Flow Implications
Three pools move.
The first is manual underwriting headcount. Credit analysts who read files by hand are the direct substitute for step two.
The second is the bureau score franchise. FICO and VantageScore sell a standard number that a good in-house model now beats.
The third is model governance. Validation, documentation, and sign-off work is a line item that did not exist in 2024.
Follow the winners. Large banks already run model risk teams for stress tests.
For them the AI Act is an extension, not a build. Small lenders have to start from zero.
That cost is fixed, and fixed costs favor scale. This is what the "AI democratizes lending" pitch keeps missing.
The rulebook made the model cheaper to run and more expensive to prove.
The same clamp is forming elsewhere. The Reserve Bank of India issued draft model risk guidance in 2026 covering AI and third-party models.
On the scoreboard this lands on legacy fee pool compression. Underwriting labor shrinks while a new compliance line grows beside it.
Verdict
The underwriter's judgment lost. The underwriter's paperwork got a promotion.
Consumer credit decisions are now a machine function with a human signature on the file. What compresses is manual underwriting labor and the bureau score's pricing power.
What grows is model governance. It goes to firms big enough to already have the team.
Territory: +machines on the credit decision, +incumbents on the moat.
Manual underwriting and bureau-score pricing both shrink, while compliance costs build a wall only large banks can clear.
