The ai Corporation is having a good 2017 and is “looking to be even more active in the market throughout the rest of the year”, writes CEO Mark Goldspink in a note circulated this month.
“The early part of 2017 has continued to be as exciting as the end of 2016,” he writes. “Artificial intelligence/machine learning (ai’s specialities) are becoming more accepted and understood in the market. We’re seeing a growth of confidence in our approach.”
As a global sponsor of the Retail Risk conference series, ai will have plenty of opportunities to spread the news about its unique self-service fraud prevention solutions, designed with retailers very much in mind.
The company will be giving keynote presentations at Retail Risk conferences in Toronto on 22nd June, in New York on 14th September and in Leicester (UK) on 5th October. To sign up for any of these conferences, free to all retailers, go to www.retailrisk.com and sign up for a delegate pass.
With its acquisition of a payment gateway and launch of several new products in 2016 ai has been increasingly visible to retailers. But the company has been around since 1998, and its recent breakthroughs with machine learning stem from many years of research and the steady build-up of a global presence.
Matt Attwell, ai’s Risk and Client Service director, emphasises that its approach to machine learning is based on a wealth of human experience as well as data.
“We have two unique insights,” he says, “we have a global footprint, as we are working with institutions worldwide, and we deal with every aspect of the payment chain.”
Currently ai services 100 banks worldwide, and processes transactions for over three million merchants as well as we scrutinising more than 20 billion transactions annually.
In conversations with retailers and other clients, ai has heard three core themes repeated again and again, Attwell says.
The first is accuracy, so that clients can identify the fraudulent transactions without declining several times that number of genuine transactions. Second is the sophistication of fraudsters, who for example are now using sim swap frauds to intercept calls from a firm or institution trying to confirm an order or a change of delivery address.
“Finally, internal fraud is a huge problem. Fraudsters are increasingly professional, they look at return on investment, they look for the weak link, and quite often unfortunately it is the staff in the store.”
While there is now a variety of anti-fraud tools, the rapid multiplication of channels – not just e-commerce, but m-commerce, w-commerce, kiosk, pop-up stands and the proliferation of payments methods – ApplePay, QR codes and Bitcoin, for example – leads to tremendous difficulty in monitoring entry points, each of which might need a different acceptance threshold.
Retailers are on a spectrum in terms of monitoring the risks, ai has found. Some still have a single channel perspective, where one team is looking at in-store transactions, another team is dealing with e-commerce and they are not really liaising. Others are at a multi-channel stage, where there is still a silo mentality but at least there’s a sharing of insights. The next evolution is cross channel, where monitors are able to see all the incoming and outgoing activity relating to a customer and thereby build a better picture of anomalous behavior. And finally there is omni-channel – having a single view of the customer and a single holistic strategy.
“But even retailers that have adopted the omni-channel model said they were still struggling,” Attwell says. “They were challenged to manually review all the orders that get referred, and they couldn’t see how to reduce that burden.”
“We started asking whether the human mind is suited to dealing with all aspects of fraud. We looked at behavioural economics. This is based on the realisation that the human mind is still wired the way it was in the days of cavemen. We’re still wired to take the easiest path, to minimize risk etc.”
ai identified 10 biases of the human mind that can influence humans’ ability to manage fraud – for example, the herd instinct, where people go with the flow, and inattentional blindness – a tendency to see only what you are looking for and fail to notice other outstanding anomalies. Attwell cites a study at Harvard medical school where 24 radiologists were asked to search for cancer in chest scans. A picture of a gorilla had been inserted into the scans but all but four of the radiologists failed to report this, although the cancer they were looking for was about 40 times smaller than the gorilla.
By contrast a software programme does not have these hard wired biases, and the more data it receives the more finely tuned are its judgements.
While machine learning is already widely applied, for example, in recommendations to customers, spam filters and credit card fraud detection, ai, says Attwell, believes the technology is moving into a new realm in terms of expertise, speed, output, scope, implementation.
“There is often a perception that machine learning is about people in lab coats with PhDs. We very much base our technology on the idea of self-service within the business. You don’t need to be an expert with data. You just need to understand what your problem is and what you’re trying to achieve.
“And these days the solutions are provided much faster than in the past, if not within days, then within hours of posing the problem. People think machine learning is a black box but modern ML gives results in human readable form. ai’s SmartRule solution gives human-readable business rules.”
“The scope has also changed, in that the technology is now agile enough to mix and match to meet a business’s needs, and the hardware required is now just a business PC” Attwell says.
He cites the example of a fashion retailer that was helped by ai to improve the accuracy of its fraud detection rules engine.
“The retailer did not perceive a problem with their fraud detection rate. The problem was lack of accuracy, a huge number of referrals and the requirement for a large manual review team that was struggling to cope with a deluge of information. This was killing the business.”
ai submitted the retailer’s data to its SmartRule machine learning technology, and the outcome was five simple rules that that could be put in a rules engine to improve accuracy.
“The result was that we were able to increase their fraud detection by 25%, reduce false positives by 60%, and the most important aspect was speed of implementation. Instead of it taking a week for their team to create a new rule set, with SmartRule it took hours. And we agreed to give them a monthly automated refresh to keep on top of the latest emerging risks. This meant that analysts who had been spending their time creating these rule sets were able to go off and target other problems and make further enhancements to the business.”
ai recently cited other standout performances with SmartRule – the completion of multiple rule-sets for a global fuel card issuer in 24 hours instead of five days per set, the time this was taking previously, and the creation of an anti-fraud rule-set for a large international bank in just two hours.
Five rule sets were created for the bank in 10 hours, with an average of 12 parameters, meaning SmartRule created a total of 70 rules in just one day’s work, a record breaking achievement, says ai. It adds that besides being implemented quickly the product detected $800,000 of fraud and reduced false alerts by 14,000 over a three week test period.
For CEO Mark Goldspink, the self-service capability that comes from making machine learning outputs easily “human readable” are a strong selling point for ai’s products.
Commenting on the company’s research and development, he writes: “I am particularly excited about the work our R&D teams have completed in our new machine learning solutions, not only with improvement on fraud detection, but the unique self-service capability that is streamlining our customers’ processes by as much as 85% – a paradigm shift from the conventional approaches that have been used so far.”
ai is continuing to invest around 45% of its revenues in research and development, a ratio matched by few other innovators in recent years with the exception of Twitter.
With the acquisition of a payment gateway last year, and the launch of several new products, the company has seen its net worth and net assets grow by 165% and 172% respectively while total liabilities fell by 18%. Founded in 1998 and headquartered in Guildford, UK, ai now services 100 banks worldwide, processes transactions for over 3 million merchants and scrutinises over 20 billion transactions annually.
ai’s researchers have come to the conclusion that three of the four stages to a data-driven decision – analysis and definition, data preparation, data analysis and modelling, and analysis and implementation – can now largely be left to machines, although the first stage, analysis and definition of a problem, involving ideation and contextual analysis, should still be purely human activities.
“Let people do what they’re good at that, let computers do what they’re good at, and if we marry the two together we are on to a very powerful thing,” Attwell says.
And what is good for other companies is also good for ai, Goldspink acknowledges.
“We are utterly committed to releasing human creativity within the ai team”, he writes, “by investing further in training our team members and automating activities that are best done by machines.”