Credit scores have a well-known problem: they measure your history of managing debt, not your actual ability to repay right now. Someone with thin credit history but a stable income and healthy bank balance gets rejected, while someone with a long credit history and maxed out cards sails through.
Abound, a London fintech, built its entire lending model around fixing that gap using Open Banking data instead of a bureau score, and the technical approach behind it is worth understanding even if you never touch consumer lending.
Why bureau scores are a weak proxy for repayment ability
A traditional credit score is essentially a lagging indicator, it summarizes how you've handled debt historically, compressed into a single number. It doesn't know your current income, your upcoming expenses, or whether you've just taken a pay cut. Abound's underwriting model instead pulls real transaction level data directly from a borrower's bank account via Open Banking APIs, income, spending patterns, remaining balance after fixed costs, to assess affordability in something closer to real time.
That's a fundamentally different kind of signal. It's higher resolution, it's current rather than historical, and it's much harder to game than a credit utilization ratio.
The engineering challenge: transaction data isn't naturally structured for this
Raw bank transaction data is messy. Categorizing thousands of transactions per user into meaningful buckets, essential spending, discretionary spending, existing debt repayments, income (and distinguishing regular income from one-off transfers), is a non trivial classification problem at scale.
Get the categorization wrong and your entire affordability model is built on bad inputs. This is presumably why Abound built its underlying decisioning engine, called Render, as a standalone product rather than a one-off internal tool, the categorization and risk scoring logic is complex enough to be a product in its own right, not a side script.
The two-sided business model this enables
Once you've built a working AI underwriting engine, you have two ways to monetize it: lend directly using it, or license it to other lenders who don't want to build the same infrastructure themselves. Abound does both, direct consumer lending on one side, and licensing the Render platform to other banks and lenders on the other.
That's a pattern worth noticing architecturally: the hard, reusable infrastructure (real time affordability modeling) becomes valuable independent of the product it was originally built for. It's the same logic that turns an internal tool into a platform play once it's proven to work reliably at scale.
Where this kind of model actually breaks
Cash-flow underwriting isn't strictly better than credit scoring, it's differently vulnerable. A bureau score is stable and slow changing. Real time transaction data is volatile, someone's spending pattern this month may not represent their situation reliably, seasonal income, one off medical expenses, or a temporary job gap can distort a model that's overly reactive to recent data. Any system built this way needs to smooth out noise without smoothing away genuine signal, and getting that balance wrong either produces false rejections or approves loans that shouldn't have been approved.
There's also a data availability constraint baked into the model: it only works for borrowers willing and able to connect a bank account via Open Banking, which assumes a certain level of banking access and digital literacy that isn't universal.
The takeaway
The interesting technical lesson here isn't "AI is better than credit scores." It's that swapping a stable, low-resolution proxy signal (credit history) for a noisy, high-resolution real signal (actual transaction data) creates a genuinely different engineering problem, one about data classification, model stability, and noise handling rather than simple pattern matching against a known feature set.
It's the kind of tradeoff that shows up in more domains than lending: whenever you're deciding between a clean-but-lagging signal and a messy but current one, the real work isn't picking the "better" data source, it's building the pipeline that makes the messier signal trustworthy. It's also a pattern that keeps surfacing across UK startups building alternative data models in adjacent spaces, insurance, employment verification, rental screening, all wrestling with the same underlying tradeoff.
If you want the full business model breakdown, funding history, and company details behind Abound, you can read the original piece on Entrepreneur Plus UK.
Top comments (0)