by Tom Upchurch
Marco, what can we do about AI? Marco, are we doing enough on AI?” The questions all come from senior executives, desperate to harness the potential that AI promises.
Yet Bressan is bemused by how the technology is talked about at board level and in the media. “Currently it denotes a vision of the future; an aspect of the sci-fi imagination; something that you still can’t do. But the truth is senior financial executives have been doing AI-related work, research and deployment of products for years.”
At the most rudimentary level, AI involves teaching machines to learn and to interact in order to undertake cognitive tasks that were usually performed by humans. The type of AI featured in sci-fi films in which machines possess a human-like intelligence, sometimes referred to as general artificial intelligence, remains a distant and elusive prospect. The most optimistic experts, such as Google’s director of engineering, Ray Kurzweil, predict that AI will be able to outsmart humans by 2029. Conservative predictions expect this to take at least 100 years, if at all.
Of more immediate relevance to those working in financial services is the deployment of narrow artificial intelligence. These applications undertake specific tasks using problem solving, deduction, reasoning and natural language processing. Such programmes are being applied across financial services, from the development of customer service programmes that use natural language processing to manage and field customer queries, through to programmes that can conduct financial research and make sophisticated models of financial markets to identify trading opportunities.
The potential for narrow applications has led to a boom in AI investment. Technology companies are undoubtedly leading the way. In 2015 the giants of AI – Microsoft, Google and Facebook – spent $8.5 billion on AI research, acquisitions and talent.
In comparison, financial institutions have made a cautious foray into the field. A handful are making investments by hiring high-level data scientists or acquiring AI companies. The hedge fund Bridgewater Associates hired the former chief engineer behind IBM’s Watson supercomputer. BlackRock has also been busy hiring some high-profile names and has announced a joint venture with Google to explore how to use AI to improve investment decision-making. Goldman Sachs has invested in a number of promising AI start-ups, including the financial research platform Kensho.
Yet most financial institutions have been slow to adopt AI, even though it is likely to usher in a new type of bank, with data and technology as its heart. Failure to adapt may lead to extinction for some. As Neil Dwane, global strategist at Allianz Global Investors, explains: “Technological competence is absolutely essential for at least staying in the game. You may still lose, but if you’re not in it, you have no hope of winning.”
AI has been around for over 50 years. Many commentators are quick to say that there have been previous periods of hype about AI. Is it really different this time around? The director of the Laboratory for Financial Engineering at MIT, Professor Andrew Lo, believes that there has been a step-change in the technology: “All new technologies come with a certain degree of hype, but I do think there has been a material change in the technology today against where things stood even just five years ago.”
The key driving force is the rapid expansion in the amount and availability of data. Whether it is data created from interactions on social media sites, financial transactions or even mobile phone data, there has been an unprecedented growth in the amount of information available for collection and analysis. In combination with exponential increases in computer processing power (Moore’s Law), increased storage enabled by cloud computing and the refinement of techniques such as deep learning, this data is used to train and optimize learning algorithms.
Marco Bressan, BBVA
As Bressan explains: “The data is the key enabler. Suddenly we can work with 10 years of history of every single transaction of the millions of transactions that take place every day, and use that to train some kind of learning algorithm. We could not do that before but we can do it now.”
The barriers to experimenting with AI are also coming down. Although cutting-edge research and development still require highly specialized skills, it is becoming easier for non-specialists to develop basic AI programmes. As the resident fintech expert at Oxford University, Huy Nguyen Trieu, recalls: “Ten years ago, if you wanted to develop an app for your smartphone that recognized what you said, you would need a PhD from Stanford or MIT. Today, you can be a very basic programmer, using [programming language] Python and search libraries, and you can develop this app.”
Even as the requirements for experimenting with AI are becoming less stringent, financial services remains a sector with ample specialist skills in mathematics, statistics and data science. This pool of talent is helping encourage experimentation with technology.
As Lo notes: “Financial services now has a very large population of quants in the industry, much more than ever before, who are ideally positioned to make use of these technologies.”
The promise of AI may seem more tangible than before, yet determining where this technology will have the most profound impact in financial services is less certain. In March, Euromoney Thought Leadership launched a global survey with the law firm Baker & McKenzie exploring the future of artificial intelligence in financial services. The results showed an uneasy tension between attitudes that on the one hand identify AI as a tool for enhanced risk management and efficiency and on the other see it as a source of systemic risk, volatility and industry fragmentation.
When asked where they expected AI technologies to be deployed within their organization over the next three years, 49% of the survey participants chose risk assessment as the most popular application.
Bressan agrees. For him AI is, “all about decision-making, internally within the bank and, externally, on the side of the client.” Data can be fed into any manual decision-making processes and can be used to help solve internal management questions, such as whether or not to open a new branch in a new location or whether or not to give a client a new loan.
The technology’s ability to crunch huge amounts of data quickly will also have an impact on how rapidly institutions can respond to risks. As Lo argues: “Now, when investors are faced with headline risk, they can actually take action from an algorithmic perspective rather than waiting for a trader to get in, read the news and then make a decision based on his or her assessment.”
Financial research was identified as the second most-promising application, with 45% identifying it as a key area for development. Lo attributes this to recent advances in natural language processing. The ability to process contextual information was previously regarded as a capability unique to humans.
But, as Lo explains, “the progress made by machine learning with respect to natural language processing has really changed that perspective and we now have algorithms that can read annual reports, 10k filings and other kinds of text and process them in quantitative ways.”
The investment in Kensho – a specialist AI research outfit – by Goldman Sachs, is testament to the potential of this application.
Portfolio management and trading were also identified as promising areas for AI development, with 37% and 33% of the vote respectively. Oliver Bussmann, the former group information officer of UBS, sees great potential with AI in refining algorithmic trading.
He is confident that these applications will receive investment because “trying to figure out how the market is moving is something that, constrained by normal human capacities, is very hard to do.” Bussmann also believes that the direct and positive impact on revenue these applications could bring will make them attractive to senior management.
Robo-advisers have proved to be somewhat contentious. The technology is still relatively primitive. During the last months of market volatility, there have been reports of investors calling their robo-adviser wealth managers, demanding to speak to a human. Lo concedes that this is an example of the shortcomings of technology, yet he still believes that, over time, the robo-advisory offering will be refined and improved. “We’ve got version 1.0 of robo-advisers,” he says. “Version 1.5 is going to be considerably more robust and user-friendly.”
Applications aimed at sales and customer service were deemed less important by the surveyed executives. Only 14% believed that AI programmes will be used to drive sales. However Nguyen Trieu thinks otherwise and looks at how Amazon employs its recommendation engine to help drive sales and deliver, “an intensely personalized service”. If banks were able to harness their customers’ data in a similar fashion, Nguyen Trieu believes that they could begin offering truly personalized products, shaped around the customer’s needs.
In November 2015, Antony Jenkins, former chief executive of Barclays, warned of an Uber-style disruption to the structure of the banking sector, mostly as a result of new technologies. His gloomy forecast suggested that as many as 50% of jobs at big banks could be cut over the next 10 years. Alarmingly similar sentiments were shared by our surveyed executives, who identified “changes to the structure of the human workforce” as the most negative implication of AI.
Banking has already seen big reductions in headcount over the last 15 years. In this sense, AI’s introduction is just the continuation of a process of automation that is already well underway. As Nguyen Trieu argues: “AI is all about digitalization and automation. In all types of banks you still have millions of processes in which a similar task is repeatedly executed. The major trend will be going back and seeing which of these processes can be digitalized and automated.”
Indeed investors are pushing for banks to demonstrate cost reductions. Dwane of Allianz Global Investors does not invest in banks and currently finds them “un-investable”. However he does believe that any chance of banks receiving recognition “will only truly come if there are real cost benefits from such a transformation”.
In Dwane’s view, cost reductions are an absolute necessity if banks want to stay competitive and attract investment. In the age of AI, the ineluctable logic of automation will very likely mean mass redundancies.
The pace of restructuring could be quicker than some institutions anticipate. Lo believes that Jenkins’ predictions are “realistic” if investments in AI research and development continue undisrupted.
Sean Park, Anthemis
Sean Park, the founder of the fintech venture capital firm, Anthemis, believes that change will be rapid and disruptive: “The scary thing about AI and machine learning is that you don’t necessarily see a slow linear shift towards technological automation. It’s more that, in some banking businesses where there is an obvious and immediate application, the 100 people that once ran that business can be reduced to just 10 or five, or even no one in some cases.”
Where redundancies occur is dependent upon a number of different factors, including the repetitive nature of the job and the availability of data, relevant to the specific function. It will also be dependent on the particular strategy of the financial institution, as Bressan explains: “If I’m worried about margins, the main impact will be on the back end. If I’m worried about increasing value to my customers, in developed saturated markets, the main impact will be in the customer experience. If I’m worried about growth in the market, the main impact will be in knowledge about my non-customers.”
Disruption is expected to go beyond the internal structure of financial institutions and is likely to impact the structure of the industry as a whole. Of our surveyed executives 56% believe that AI will drive market diversity, with more small and medium-sized participants entering the marketplace. Only 8% believe that there will be no change to the structure of the financial services sector.
Nguyen Trieu agrees that AI will accelerate the fragmentation of bigger financial institutions, as seen in the private banking sector where, “you see many more people setting up their own small family office or boutique firm. This is much more easily achieved, as from an infrastructure perspective it is much simpler to outsource infrastructure to smaller companies.”
Over the past few years a number of high-profile data scientists have moved from technology companies to financial institutions. They have been hired to help banks leverage the wealth of customer and market data they hold.
These scientists bring a rare blend of skills in computer science, mathematics and business. Their rarity means they command enormous salaries. In line with this demand, our surveyed executives identified the shortage of specialist skills in technology as the second-most significant obstacle with introducing AI into their organization. “If I was sitting on the board of a big bank, the biggest challenge they have is acquiring the best talent,” says Park.
Financial institutions will struggle to compete with big technology companies in acquiring the very best talent in mathematics, data science and computer science. Companies like Google and Facebook can offer the compensation, intellectual stimulation and brand appeal that brilliant young people crave.
As Park puts it: “Today, if you’re, say, Goldman Sachs, forget about trying to poach the best technologist talent away from an Amazon, Google or Alibaba, it’s very difficult.” This will make it more challenging for banks to engage in experimental AI research and development.
However banks do not necessarily need the very best technologists to develop successful AI systems, products and programmes.
Huy Nguyen Trieu,
Nguyen Trieu believes the future talent challenge for the banking sector lies in finding individuals with hybrid expertise – some technology expertise combined with knowledge of their industry.
As the level of AI being applied to financial services is moving away from the realm of pure R&D and towards technically simpler applications, Nguyen Trieu believes that tomorrow it will be about “finding the people who understand how to apply AI to different business environments and different processes.”
This challenge could be harder to overcome than finding pure scientists. In many financial organizations there is a great disconnect between business and technology. Nguyen Trieu suggests that institutions should be encouraging the cross-fertilization of skills by arranging tailored graduate programmes that encourage those in business to train in technology and vice versa.
Financial institutions can also benefit from the increasing importance of open-source. Don Duet, head of technology at Goldman Sachs, thinks this has an important impact on the type of individuals he needs to hire to run a successful team. As technology companies share most of their latest advances in an open-source format, this means that Goldman Sachs can focus on hiring people with domain expertise: “Being able to bring together people with deep domain expertise and bring them together with people with deep technological expertise helps drive to success… that intersection point is very important.”
Bressan is less convinced that the emphasis should be placed on hybrid-talent. He believes that business knowledge is more easily acquired than the technical knowledge of artificial intelligence, deep statistics and mathematics. The real key lies in “building multidisciplinary teams that complement each other… rather than to have mono-talent.” Despite the competition that technology companies pose to financial institutions, Bressan is confident that FIs can attract top talent, chiefly because “we have very rich data and we have non-solved, deep problems.”
As financial institutions become increasingly dependent upon AI and data, they will be exposed to new risks and vulnerabilities. As Lo explains: “The big risk is Moore’s Law meets Murphy’s Law.”
As computer power increases exponentially (Moore’s law), the chance of it all going wrong also increases (Murphy’s Law.) We have already seen this combination in high-frequency trading where new technology has enabled organizations to trade and make mistakes at far greater speeds.
Beyond the realm of trading, what could go wrong with employing AI in other parts of your business? Bressan believes that the main, all-encompassing risk is something he refers to as “algorithmic quality.” Although machine learning techniques have advanced over the years, the successful training of algorithms is highly dependent upon the correct use of data. If not enough data is used, or if the data is biased or unbalanced, then this will be reflected in the algorithm.
As Bressan explains: “If you use biased data, the algorithm will also learn your biases. If you use data that is biased against a particular minority, which is not adequately represented in your data set, the output algorithm will completely ignore that minority because it was not able to learn about the behaviour of that minority.” As a consequence you will have an algorithm that is intrinsically discriminative.
Algorithms also need constant updating, to match the changing nature of reality. Any failure to do this may make the algorithm irrelevant and liable to error. Bressan mentions an interesting example from the field of epidemiology. A few years ago researchers estimated that the outbreak of flu virus could be predicted more accurately by analyzing Google search terms and key words. At first, the algorithm did indeed prove more accurate than the prediction models employed by national health organizations.
However after a few years of using the same algorithm, they found it was no longer working. The reason for this was because the way people searched Google had changed. As Bressan warns: “These algorithms need to have the right data, they need to evolve over time and they need to be robust and consistent.”
The scope for algorithmic failure opens up a range of potential legal risks. These risks are yet to be clearly defined, as regulation and legislation are slow to catch-up with rapid technological change. However if flawed investment decisions are made as a result of poor data or malfunctioning algorithms this could pose corporate liability risks. Data protection and privacy risk also arise, if, for example, personal investor data fell into the wrong hands.
As Steve Holmes, TMT partner at Baker & McKenzie, points out: “Some of these risks are very specific, including increased likelihood of security breaches, data loss, fraud or intellectual property infringements. While digital represents a significant opportunity for most companies, there are also legal risks that need to be considered.”
Incidents such as the Knight Capital crash, in which the trading firm lost $440 million due to a programming error, are likely to reoccur, particularly as more systems become automated. Despite these risks, however, Nguyen Trieu believes that greater than the threat of flash crashes is the business risk of not investing in the technology and being left behind.
For Nguyen Trieu: “AI is a little bit like electricity.” Just as those companies that did not invest in electricity eventually suffered huge competitive disadvantages and were quickly left behind, so too will companies resistant to AI. He believes that this is already happening now, across a variety of businesses “where automation of previously manual processes are creating cost-savings of at least five times.”
When it comes to regulation, our survey demonstrated a lack of confidence in the regulators’ ability to stay ahead of technology trends – 69% of respondents were either ‘not very confident’ or ‘not confident at all’ in the regulators’ understanding of technology. However the challenge for regulators is extremely difficult to manage. They are under-resourced in talent and capital, compared with financial institutions. Perhaps more importantly, they have to strike a balance between preparing for new technologies while avoiding the temptation to indulge in anticipatory regulation.
Lo, himself a member of the OFR Financial Research Advisory Committee, sees a challenge ahead for regulators: “Instead of being two steps behind, regulators could be five steps behind if the pace of financial innovation continues. That’s the concern.” Regulators themselves will have to consider making similar investments in AI technology to help them oversee markets.
Is AI the future of banking? If referring to AI as a wider process of automation then the answer has to be an unequivocal yes. The more pertinent question is whether large financial institutions are in a position to adapt to this future. Park remains puzzled by the paradoxical situation where banks “have huge financial and human resources to deploy, but there is still a challenge in taking these technologies from the laboratory to the real world… culturally, it’s like turning a super tanker. Not easy.”
Neil Dwane, Allianz
The first big obstacle is adopting new technologies, while being encumbered by a dense and debilitating network of legacy IT systems. AI can actually play a part in helping to reduce complexity and improve the efficiency of old computer systems. Duet mentions that Goldman Sachs has used AI to do just that.
But Dwane remains sceptical that banks are in a position to change: “I know many CTOs at large banks currently spending 100% of their money on effectively keeping the lights on the bank’s technology.” If all efforts are focussed on just keeping the machine running, then it is unlikely that investments in AI will prove successful.
The second and perhaps bigger obstacle is the cultural shift that AI and associated technologies demand of organizations. Banking leaders need to fully understand, practice and preach the art of total data.
Park’s view is damning: “If you look at the executive suites or boards of banks, and you’re objective about it, they don’t have the right people to make these changes.”