As U.S. banks wrestle with the decision of whether to use artificial intelligence to help calculate credit scores and make loan decisions, a potential role model is MyBucks, a company that’s been doing this for more than a year —and has even begun offering 15-minute, AI-based loans through WhatsApp and Facebook Messenger.

MyBucks is a Luxembourg-based fintech that owns several banks and provides loans and basic banking products in seven African countries, Poland and Spain; it’s expanding rapidly into other countries.

U.S. regulators have signaled a willingness to accept banks’ use of AI in lending. And the evidence so far, at least in MyBucks’ case, shows that AI can improve credit quality and reduce defaults.

How the Haraka app works

MyBucks’ Haraka app, which is now offered in Zimbabwe, Uganda, Swaziland and Kenya, and in early 2018 is expected to be introduced in the Philippines and India, can score a customer within two minutes. The customer downloads the app from the Android store; MyBucks pays for that download through reverse billing, so even customers who are broke can still access the app. The app then asks for permission to scrape the phone for data, including text messages, call-history patterns and geolocation information.

The information can be very useful. For example, the text messages of heavy users of mobile payment programs like M-Pesa contain all their mobile money transaction verifications.

“That gives us insight into customers’ income and expenses,” said Tim Nuy, deputy CEO at MyBucks. “We see any payments in and out of that person’s account.”

Tim Nuy, deputy CEO of MyBucks
Future of lending?
“I think it’s where everybody is moving: to create a user experience that’s as flawless as possible and that makes it easy and natural for a customer to get their credit,” says Tim Nuy, deputy CEO at MyBucks.

Customers also log into their social media accounts from the app to help MyBucks verify their identities: The company compares the applicant’s social media feed against the information in their mobile wallet.

“We look at behavioral traits,” Nuy said. “Very active social media accounts are likely to be real people, and we make sure the information on the cellphone and the social media account tie together.”

Can MyBucks obtain a complete picture of a person’s financial situation from that phone-scraping?

“The reality in an emerging market is that it’s never perfect, but at the end of the day you need to make a judgment call,” Nuy said. “People have the option of simply deleting their messages and expenses, in which case we wouldn’t get a full picture. But we get a full enough picture to grant a loan with a default rate that’s acceptable to us.”

The bank starts with very small loans. A first loan to new customers might be a mere ten euros. If such borrowers successfully repay, they will qualify for a larger loan the second time. The returns made on those repeater loans make up for the slightly higher default rate — around 20% — on the first loans.

MyBucks charges a low interest rate for Haraka loans, but its processing and mobile money disbursement fees add up to a 20% average cost per loan. It makes 10,000 to 20,000 Haraka loans a month, for a total of about 200,000 in the past year.

“Haraka has been one of our bigger successes, particularly because it allows us to do loan applications at a smaller size more profitably, and it allows us to give people who are otherwise unable to secure financial services access to that initial loan,” Nuy said.

During the year, Haraka’s AI engine has improved significantly by analyzing defaults.

“It’s mainly refining how it reads mobile money wallets, what transactions to include and exclude —do high volumes of transactions indicate creditworthiness?” Nuy said.

The default rate for Haraka loans has dropped from more than 30% to less than 12%. Across all its loan products, MyBucks’ default rate is 7%.

The company plans to bring Haraka to Mozambique and Botswana, Australia, Indonesia, Myanmar, Vietnam, and Cambodia.

“Our view is that this has global applicability in emerging markets, particularly markets with a high penetration of smartphones,” Nuy said.

The fact that some countries, including Zimbabwe and Myanmar, are in turmoil has not affected MyBucks.

“We’ve done extremely well in Zimbabwe, because we’re one of the lenders that’s been there from the beginning and have been serving people through difficult times,” Nuy said. “People who live in these countries want a politically stable situation and access to financial services.”

In addition to the very small Haraka loans in emerging markets, MyBucks also makes more traditional loans in mature markets like South Africa. It pulls credit reports and verifies income and employment, yet it also uses AI and lends over mobile devices. It offers checking and savings accounts as well as loans.

“Having a bundled offering is an attractive part of the company’s business model,” noted Craig Focardi, senior analyst at Celent. “They can acquire low-cost deposits from some customers and lend them out to others.”

This is a more sustainable fintech model than some of the challenger banks in U.S. and Europe that offer mobile checking and savings accounts but not loans, he said.
MyBucks began offering loans through a chatbot in WhatsApp and Facebook Messenger in October — just to South Africans so far. The entire loan process and customer service queries are handled through the bot, so customers do not have to download an app.

Behind the scenes, MyBucks pulls the applicants’ credit report and obtains their three latest monthly bank statements (which it feeds to the AI engine). It sends the funds to customers’ bank accounts or e-wallets within 15 minutes.

“I think it’s where everybody is moving: to create a user experience that’s as flawless as possible and that makes it easy and natural for a customer to get their credit,” Nuy said.

To encourage use of the app (and gather more data about customers), MyBucks offers better loan terms to those who download it.

In Africa, bank regulators value the ability to offer loans to people who have no credit history.

“They are aware that for some people there’s virtually no information available, so the fact that we’re willing to use alternative sources and give them a loan is satisfactory,” Nuy said.

Could this happen here?

Many of the markets MyBucks operates in would look foreign to U.S. bankers: developing countries where screen-scrapable mobile payments are widely used and regulators are open-minded.

Yet some of MyBucks’ methods could work for U.S. banks, such as the use of AI in credit scoring, underwriting and fraud decisions and the use of mobile apps and chatbots for lending, especially to serve the underserved.

The appetite for using AI in lending, especially among large traditional U.S. banks, has so far been small. BankMobile announced its implementation of Upstart’s AI-based lending platform last week; it is one of very few.

“AI in credit decisioning in general is still a new and developing technology,” Focardi said. “Few U.S. companies would implement technology without having a better idea of what their expected loss rates are likely to be.”

U.S. banks are testing AI on existing loan portfolios and new loans to compare its predictive ability to their traditional underwriting models.

“That’s the most prudent way of implementing new scoring analytics in the U.S. lending markets,” Focardi said.

U.S. bankers worry about compliance: providing a reason code for approving or denying credit, making sure an AI engine wouldn’t violate fair lending and disparate-impact rules. Providers of AI lending software, including Upstart,, James, Zest, Kabbage and Enova, say their programs can provide a clear reason code and test for fair lending and disparate impact.

The first U.S. banks to use AI in lending could have an edge over those who hesitate. Yet many see no need to change, as regression-based credit models still perform OK.

Marc Stein, CEO of, said that for a large U.S. bank to adopt AI-based lending, it would have to decide to expand credit access into underserved markets without significantly increasing risk.

“They aren't able to do that with their current methodologies,” he said. “The only way to do that is to effectively underwrite thin-file millennials. The only viable model for doing that is using machine learning.”

AI-based underwriting, Stein pointed out, is a lot cheaper than human underwriting.

“We were asked by a bank in a former Soviet state to replace their underwriting staff off 150 with one of our algorithmic models,” Stein said. “This is an extreme example, but what I expect to see here in the U.S. is increased AI-based automation aimed at increasing the productivity of human underwriters, and underwriting more volume with fewer people.

“At some point, the benefit becomes too obvious even for a major bank,” he said.

Editor at Large Penny Crosman welcomes feedback at