The world's biggest wealth manager UBS is building "virtual agents" that can perform investment research to near-human levels.
Speaking at the TechXLR8 event at London's Excel centre last week, Annika Schroeder, AI lead at UBS Group Innovation said that the bank is "trying to build virtual agents that can imitate the quality of an investment analyst".
"It can screen through market data, through SEC filings and can actually do a company valuation with all of the inputs that a human analyst would use and can produce text in a fairly decent quality and almost human mimicking language, so we are getting very close there," Schroeder said.
UBS isn't ready to put this into live production yet, and Schroeder wouldn't be pushed on a potential timeline. She said: "We have reached fairly high confidence levels of what we can achieve but this will take time before there has been enough data to train the machine to a high enough confidence level so that we feel comfortable actually applying this in live production".
Schroeder also spoke about efforts to eliminate innate biases from the behaviour of its portfolio managers. She said: "We ran a trial identifying typical behavioural biases, like selling too early, or selling on a Friday and buying on a Monday.
"We have been able to identify the very high probability, typical biases and put out alerts" so that these typical behaviours are avoided when they don't produce the optimal result. "Again this is not something which is ready but we are always in the experimenting phase here," she added.
Schroeder said she believes AI will lead to "better investment performance, which obviously will only hold as long as the competition takes to recreate what you have done".
She also recognises that some of these developments could have an impact on jobs. She said that UBS believes AI to be a "collaborative play" with humans. "The real benefit here lies not in replacing employees but increasing the convenience and supporting employees and not engaging in any high risk processes", she said.
UBS is also "investing heavily in digital literacy to explain the new technology to the workforce," Schroeder said. "We are organising machine learning meetups and education across the organisation to hopefully captivate that fear that is out there very much."
How UBS is utilising AI
UBS has been evaluating applications of AI since it launched a dedicated programme in early-2016. This consists of "sourcing ideas from employees through hackathons, talking to startups and strategic vendors and running focus days around specific topics", Schroeder said.
The ideas that come from these activities are then pitched Dragons' Den-style to the central UBS group innovation board, to gain funding to run proof of concepts (POCs). Schroeder said that of all the innovation projects the bank is running, over a third now hinge on AI technology.
For front office, customer facing projects UBS is already trialling AI-driven robo advice and conversational interfaces, like chatbots. Schroeder is aware of the pitfalls of chatbots though, saying that "if a conversational interface goes wrong we are pretty screwed. So we are currently just trialling internally to see how users take these up, as we see a long path ahead to good user acceptance.
"The trick here is natural language understanding. We are currently at 60-85 percent, which isn’t good enough to let it loose to customers", she said.
For the back office UBS is looking to automate some of the know your customer (KYC) process for regulatory staff. Schroeder said: "So probably the negative news search part is easy but there are other parts of the work of a due diligence officer which are very manual, that require a lot of experience and gut feeling and reasoning, so this is something we are trying to part automate."
In operations the first AI project to go live at UBS was a machine learning system for dealing with IT tickets. Schroeder explained: "The system works probabilistically to understand structured input from IT application support tickets, find the best way to resolve requests by teaching it a number of knowledge items which it is using on its decision path to reach the perfect solution to a problem.
"Obviously it is getting better over time and is already handling thousands of tickets and will learn to anticipate typical faults that lead to tickets and significantly lower that number."