Risk Prevention Models Have Tried an Early Version of AI Before
An interview with Laurent Clerc by Sean Murphy
In the early 1990s, when Laurent Clerc was building credit scoring models for a living, a new technology arrived promising to sweep the old methods away. Neural networks, the first commercial wave of artificial intelligence, could find patterns in a sample of borrowers that no statistician would spot. For a while the industry was enchanted but then the models met the real world and their performance fell apart. Within a few years neural networks had disappeared from risk modelling almost entirely.
He watches with the calm of someone who has seen this cycle before.
Clerc has spent almost thirty-five years in lending, the first ten inside banks and financial institutions working on credit risk, the rest as an entrepreneur. His first venture, a lending technology firm serving Central and Eastern Europe, was built on the older on-premise architecture of its era and sold to a Polish IT group after a decade. In 2014 he started again with twenty-five partners, this time designing for the cloud from the first line of code. The result was Circeo, a Luxembourg-based company whose platform, runs lending and leasing operations for banks and finance houses from application through to the final repayment. He leads the business today as president of its management board.
His caution about AI is not a lack of respect for what the tools can do. Finding a strong model on a given sample of borrowers, he says, was never the hard part of credit scoring; classical statistics could manage it, and modern AI can probably manage it better. The hard part is stability. A lender applies its model to people it has never seen, and if those new applicants differ in profile from the training data, or simply behave differently, the cleverest model in the laboratory will misfire in production. That is precisely where the neural networks he saw in his early career failed, and he is not yet persuaded that the new generation has escaped the same trap.
He also raises a second concern about AI with a bit more edge to it. An AI system fed rich personal data can learn to discriminate without being instructed to, finding proxies for the characteristics the law forbids it to consider. European authorities are alive to the danger, which is why creditworthiness assessment appears among the high-risk applications singled out by the EU's AI Act.
Where Clerc turns genuinely enthusiastic is somewhere less obvious: the data a lender receives at the moment someone applies. For as long as consumer credit has existed, lenders have relied on documents as stand-ins for the truth, and a payslip rendered as a PDF can be forged by anyone with a laptop. In a growing number of European countries, an applicant can now consent to their income being verified directly against tax authority records, a route that works today even in France, a country Clerc affectionately files under the old-fashioned. Open banking performs the same service for transaction data. In both cases the forgeable document drops out of the process, fraud becomes markedly harder and lending decisions arrive faster. Set against the application forms still common across Europe, which can stretch to fifty questions for a modest consumer loan, the difference is stark. The infrastructure is spreading quickly; a forecast cited by J.P. Morgan puts open banking users at 183 million in 2025, on the way to 645 million by the end of the decade.
Serving lenders in many countries from one codebase brings its own puzzles. Consumer credit regulation in Germany bears little resemblance to its American counterpart, and Clerc's team counted twelve distinct methods in use around the world for something as basic as how many different ways there are to calculate daily loan interest. Rather than rebuild for each market, Circeo wrote all twelve into the platform's core. A student loan in the United States requires one method, car finance in Kazakhstan another, and in each case the correct treatment is chosen from a drop-down list rather than commissioned as new code.
The philosophy underneath is simple enough: the fundamentals of lending are the same everywhere, so they live in the core, and local regulation is handled as configuration.
The company behind all this is an oddity in its market for another reason. Circeo has never taken venture capital. Among the founding partners were seasoned bankers with wealth of their own, and they funded the business themselves so that nobody outside it could force a sale, cut off financing, or veto a strategy. Ownership stays close to the work: nearly all shareholders are employed in the business, and staff can become shareholders after a few years, which he credits for the loyalty the company enjoys. "That's called entrepreneurship," he says of the decision to risk his own money. "That's your own child."
Asked what borrowing will look like in three to five years, Clerc points not to some new intelligence sitting in judgement but to the steady spread of verified data through the application process. The technology is ready, he says, and the lenders are ready; it is borrowers who still hesitate, wary of what happens once a stranger can see their tax file or their current account. He finds the fear understandable and misplaced, since few industries are as tightly supervised as this one. "It is just getting data, analysing the data, and making a proper risk and affordability assessment," he says. "That's all." He expects the shift to take years, and to arrive without fanfare when it comes. It will not be a revolution, he says, and he does not mean that as a criticism. But it will make borrowing quicker, safer and easier, for the lenders and for their customers.
