Faster judgement ≠ new judgement

As I approach my fourth Milken Institute event of the year (don’t cry for me, the professional conference circuit attendee), and having continuous conversations around what I do as an AI-driven private credit manager - I found myself thinking about the man himself. Michael Milken was really the original credit innovator. Milken built a pipe between the capital that Garn-St Germain had just freed up and issuers everyone else considered radioactive, and he stood at the end of that pipe insisting the risk was mispriced, not bad. The junk bond market was a plumbing story, not a productivity story. Structural innovation in finance has never just been about doing the old thing faster. It's about deciding who gets to hold what risk, and why.

I mention this because everyone I meet right now wants to talk about AI in asset management like it's the second coming of the spreadsheet (which ok, it might be), but a faster spreadsheet is not what changed finance in 1983. So before I get overexcited about agentic underwriting or AI-native private credit, I want to apply a Milken test to it. Not ‘does this make the process faster’. The question is whether it expands who can hold risk and what can get financed in the first place, or whether it just lets the same people price the same risk in less time. Most of what's out there right now fails that test, and I’m not just talking my own book. Covenant extraction, memo drafting, document review, deal sourcing copilots are all genuinely useful. That’s throughput, not architecture. I read a sharp substack essay by Joshua Easterly of Sixth Street recently making exactly this point through the lens of Baumol's cost disease: financial services has always been protected from the deflation that hit manufactured goods because labour is 50-70% of cost, and labour doesn't get more productive on its own, the way a string quartet still takes four musicians an hour to play Beethoven. The interesting claim isn't that AI lowers the cost of a unit of judgement. It might let a fixed stock of senior judgement produce many more units in the same hour, which is a throughput story, not a cost story, and throughput is where the real operating leverage lives. That's a real and probably underpriced shift. It is not, on its own, a Milken shift. It doesn't change who can access capital or what kind of borrower becomes financeable. It changes how many memos an MD can sign off on before lunch.

Where I think the Milken-shaped opportunity actually sits is in underwriting that reaches borrowers the old system structurally couldn't see. There's a real push right now to underwrite emerging-market SMEs and thin-file borrowers using alternative data, mobile payments, satellite imagery, transaction flows nobody was digitising five years ago. That's the interesting case, because it isn't pricing existing risk faster, it's manufacturing a category of financeable borrower that didn't exist for institutional capital before. It’s the thesis behind a lot of what we do at Liquidity: underwriting asset-light companies a traditional balance sheet lender would balk at, because we see revenue quality and burn dynamics a covenant checklist was never designed to understand. Milken needed a captive pool of newly liberated S&L money to make junk bonds work. The open question for this decade is what plays that role for AI-originated credit: insurance balance sheets moving into algorithmically underwritten private debt, sovereign capital in the Gulf allocating into AI-scored EM credit, tokenised structures giving smaller pools of capital access to origination they were previously locked out of. Whoever solves that matching problem, not the underwriting-speed problem, is the one doing something structural.

I'd also gently resist the temptation to treat this as uncomplicated progress, because Milken's own market is the cautionary tale sitting right there in the record. Junk bonds worked fabulously until they didn't: spreads compressed, quality slipped, the market got flooded, and by 1990 Drexel was gone. Easterly's essay makes almost the identical point about active management, where indexing turned out to be the automated version of a portfolio manager and beta got manufactured at near-zero cost, and fees simply migrated toward the marginal cost of the automatable piece of the job. Cheap, scalable capacity meeting a transparent market doesn't reliably reward the person who built the capacity. It usually just gets arbitraged into price. The same multiplier that could make agentic AI transformative for one firm is the mechanism that turns it into an industry-wide deflation engine for everyone else, and the difference between those two outcomes comes down to one rather unglamorous fork: is the compute and the data behind it a commodity everyone rents at the same rate, or is it proprietary and scale-sensitive in a way that compounds with size? If it's the former, you get a faster Bloomberg terminal. If it's the latter, you might get something closer to Drexel's client network, a durable structural edge this industry has genuinely never had, because for two centuries everyone sat at roughly the same point on the same labour-dominated cost curve.

I’ve been repeating the tired bit that perhaps only the front-office personality hires in finance are truly safe. The reason isn’t charm, it’s that the thesis still has to come from somewhere, and relationships are how you hear about opportunities before they’re in a data set.

This is roughly the argument even the most aggressive AI forecasting ends up unintentionally supporting. The case for 2027-era systems doing the work of a research engineer or analyst is rather unsettling forecasting. Even that version of the argument concedes that superhuman reasoning processes information faster, but it doesn't manufacture new facts about the world on its own. Milken's actual insight, the one the pipe was built to exploit, was that default rates on below-investment-grade debt were persistently overstated relative to the yield being demanded. That's a mispricing thesis, and someone had to have it before any amount of processing power was useful. My honest guess is that the next real structural break in this industry still starts with a human noticing that some category of risk is being systematically misjudged, and agentic AI is the thing that then scales that thesis into a market at a speed and precision no trading floor of analysts could ever match. The multiplier is extraordinary, it just needs something to multiply, and that part is still, stubbornly (or perhaps happily), a human problem.