Underwriting Is Not the Edge

A qualified account of what ML underwriting does

Eric Glyman, whose company Ramp does underwrite with machine learning, has offered an unusually restrained assessment of the technique: that ML underwriting was among the most overstated claims of the 2010s fintech wave, and that it functions as a marginal-improvement lever rather than a business-defining one. Responding on the MAD Podcast to Matt Turck's observation that the underwriting-as-a-service wave largely concluded that traditional scoring "was actually pretty good," Glyman placed the claim in time before assessing it: "This is a big part of the New York tech story probably a decade ago, when there was a lot of fintech lenders." His verdict: the capability "is real. A lot of those companies overstated how useful it is."¹

He described the gains in terms of approval rates for sales prospects. By his account, no heuristics approve roughly 60 to 70 percent, good heuristics reach about 80 percent, strong programs reach the low 90s, and great ML models take that to the mid 90s. His conclusion: "it's not going to totally change the nature of the business." He located the actual value not in loss prevention but in growth: the models mainly let the company "say yes to more people, as opposed to prevent losses." In his framing, underwriting ML is an approval lever more than a risk lever.

The structural context Glyman cites

Glyman connected this to why Ramp chose the corporate credit space. In his telling the choice was partly about the loss environment, which he says restricts how much complexity underwriting needs to carry. He contrasted annual credit losses across segments: subprime consumer cards in the 5 to 10 percent range, general prime consumer around 0.5 to 2 percent, and corporate cards, citing historical Amex disclosures, around 0.1 to 0.2 percent. He claimed Ramp had run below that corporate range, attributing the difference to its use of modern techniques.

He was careful that low loss rates do not make risk unimportant. "If margins are on the order of a percent, one loss can wipe out the earnings on 100 other companies," an asymmetry he presented as the reason underwriting discipline stays binding even when the headline rate is small. He also noted that underwriting at Ramp began as rules and human underwriters and became progressively more ML as data accumulated, whereas fraud detection was heavily ML from early on.

Where Glyman relocates the advantage

The consequence Glyman draws is that if ML underwriting buys only a few approval points, the durable edge in card fintech sits elsewhere. He points toward the data and workflow position, developed in Proprietary Data Moat, and toward an aligned business model in which the card's economics reward spend rather than debt, a view connected to Purchase Volume as the Single Variable. He presents this as the risk-side counterpart to the broader "scoring was fine" concession, and as a calibration template for evaluating AI claims generally: the technology can be real, the marginal numbers still matter, and moving from 90 to 94 percent is a different proposition than changing the nature of a business.

Attribution and open questions

The comparison to Amex is Glyman's own claim, made at a relatively early and smaller stage of Ramp's book than the historical figures he cites, and is self-reported. His argument is also specific to card underwriting on rich bank-transaction data. In thin-file or emerging-market segments, where data itself is the binding constraint, ML may move the needle further, and the source offers no counter-evidence on that case either way.