The questions AI lenders should be asked at Money20/20, and the answers that count

Lending will be one of the busiest categories on the Money20/20 Middle East floor, and almost every company in it will describe AI-powered underwriting. The phrase covers two quite different things, and the difference determines whether a company is expanding access to credit or just processing applications faster. The first version uses machine learning to speed up a decision a bank would also make. Documents are read automatically, checks run in parallel, a score is produced in minutes instead of days. That is real value and it is not a new credit model. The borrower who gets approved is broadly the borrower a bank would have approved, arriving at the answer sooner. The second version underwrites on data a bank cannot see. A lender with access to a merchant's daily transaction flow, procurement history, delivery records or platform ratings is assessing something a credit bureau does not hold. That is what allows lending to a business with no audited accounts, no collateral and no credit file, which describes most micro-businesses in most markets. Only the second changes who can borrow, and the two are described in identical language. The Saudi context makes this worth pressing on. Nayla, a lender to micro-businesses built through Sanabil Studio, closed a pre-Series A of about $18 million around LEAP 2026, combining equity led by Idrisi Ventures with a debt facility led by BLOMINVEST. The structure itself is informative: a debt facility alongside equity means a credit investor was willing to lend against the loan book, which is a harder test than persuading a venture investor to fund growth. TechScoop reported that round separately. Elsewhere in the market the pattern from embedded finance holds, with lenders that already operate their borrowers' software raising credit facilities against what they can see. Foodics Capital and SILQ's Fina platform are the clearest examples. So here are the four questions worth putting to any lender at the event, in order of how much they reveal. What is the size of the loan book today, and what was it twelve months ago. This is the only question that establishes whether lending is happening at all. A company with a product and no book is running a pilot. Growth rate matters more than absolute size at this stage of the market. What is the default rate, and over what period. This is where evasion is most common, and the evasion is usually technically honest. A loan book that has grown quickly has a low apparent default rate simply because most of the loans are young and have not yet had time to go bad. The meaningful figure is cohort performance: of the loans written twelve or eighteen months ago, what proportion defaulted by the end of their term. A lender that has been through a full repayment cycle can answer. One that cannot should say so. Who funds the book, and on what terms. A lender using its own equity is subsidising loans with venture capital, which does not scale and hides the true cost of credit. A lender with a debt facility from a credit fund or a bank has been diligenced by someone whose money depends on the underwriting being right. The identity of the funder is a third-party assessment of the credit model, and it is more reliable than any description of the model itself. What data do you use that a bank does not have. If the answer is transaction flow through a platform the lender operates, or procurement records, or verified delivery data, the underwriting is genuinely different. If the answer is a description of model architecture, the lender is doing conventional credit assessment quickly. There is a fifth question that matters in this market specifically, and it is a regulatory one. Which entity holds the licence, and what does it permit. Saudi lending activity sits under SAMA supervision, and the range of licences is wide: consumer finance, SME finance, buy now pay later and debt-based crowdfunding are different permissions with different conduct obligations. A company can market a lending product while the actual credit is written by a licensed partner, which is a legitimate structure and a materially different one for the customer. The counterpoint to all of this scepticism is that thin-file lending genuinely does work when the data is good, and the Saudi market has an unusual amount of the right kind of data: high electronic payment penetration, national digital identity, and now licensed open banking that can surface bank transaction history with consent. The conditions for a real expansion in SME credit are more present here than in most markets, and dismissing every AI lending claim would be as lazy as accepting them all. The distinction is empirical and the figures are simple. Book size, cohort default rate, funder, data source, licence. Five answers, none of them about the model.