Artificial Intelligence (AI) is now mature enough to get out of the research labs and disrupt many business models. Currently AI algorithms have the ability extract information and reason based on information that comes from images, written natural language and speech, and take adequate action depending on the data analysis. Moreover, their performance improve over time as the methods gather more information. Financial regulation is very much within the fields that will be impacted by the new AI methods.
We participated this week at the AIEurope Conference 2016 where the AI related business community got together to discuss new possibilities that AI is opening. As 2015 was the year of blockchain, 2016 has been the year of AI and more specifically, machine learning. In fact, AI related business have received enormous amount of financing by industry leaders from sectors as divers as the banking (BNP, J.P. Morgan), retail (MACYs), law firms (REED SMITH) or locomotive companies (SNCF, UBER). There has also been a massive increase in patent submissions in all the world, specially (and surprisingly) in China. So, as the hype on these technologies increases, the need to analyze the successful cases augments. As all deep learning practitionairs know, it can be tricky to use those algorithms successfully. In this post, we want to outline some of the ideas that were shared at the conference for a better AI implementation at a business level.
AI within the business environment
The impact of AI on business goes beyond ‘data science’ projects, it has to become a tool that allows human users to focus on decision making instead of information gathering. It has proven by user experience that the best results are obtained when we allow machines to manage systematic and repetitive processing of information, and humans control the overall process. This is normally referred to as Enhanced Intelligence. What we want to implement are methods that put humans at the center of the process and algorithms as a useful tool in their everyday use.
In this context, a talk that sticked out ‘Legal:identifying valuable information buried within a mass of unstructured document data’ by Lucy Dillon (Reed Smith) and Peter Wallqvist (Ravn Systems). They explained their proof of concept project on automatic summarisation for legal document processing. They verified that machine learning algorithms could obtain as good results as a group of junior lawyers, at the task of summarising and recollecting data from big amounts of contracts. The automatic method was also orders of magnitude faster. No need to underline the importance this has for law firms who are always tight on schedule.
Another important aspect of the application of AI is to improve user interaction. AI needs to be transparent, and work by delivering to the user only the information she/he needs at the precise moment. The main goal of AI for and in general, to simplify user intervention. This could also mean for the system to understand the information that the user needs at each point using inference, and set it out in a meaningful way, so that it can be easier to manage (knowledge representation or data visualisation).
In general, the recommendation for business was to stick to small scoped IA projects with a well defined methodology, in an environment that is well understood. Regarding IA companies, the recommendations were to focus directly on the added value to offer with special focus on the user experience and being very focused on an industry specific product innovation (Robert Golladay, CognitiveScale).
Annika Schröder (UBS) pointed out that current AI methods need to be more transparent in their decision making, so that a human can inspect what specific data made the machine decision for a certain option.
Putting to practice AI in organisations
Innovation: AI must always be implemented to answer to a business need, but in order to define it correctly we need some focus research at the business level. Although the final algorithm is important, what really needs some reflection is how to make it useful for the organisation, and this can only be solved by putting money and resources on innovation. Annika Schröder (UBS) underlined the need to have focused research projects with clear goals and avoid ‘random experimenting’, although she highly stressed the need of innovation. In fact, she clearly stated that there is a great proportion (80%) of the research projects that will fail, but that is ok, because its the cost of being innovative and the other 20% will justify the expense.
All the organisation needs to be on the loop: Ai implementation goal’s is to transform the organisation by improving everyday work, thus the project needs support from the higher ranks of the company, and to integrate data from all departments, experts in the domain, and users in general. Several presentations underlined the need to have a fast prototyping and testing scheme, to evaluate constantly with the final users the usefulness of the project at hand.
‘Garbage in, garbage out’. ML projects rely on good learning data, which means that the data set that the method uses must be reliable (good quality) and meaningful for the task to solve. Thus there is a need of constant communication with the experts on the domain. Data quality also involves how to evolve to information available to be able to take advantage of it.
Summarising, a successful ML implementation will always need to loop the following phases (Heloise Nonne, Digital direction):
Exploration+data transformation+machine learning+expert confrontation
Fields to focus on for the regulatory and compliance sector (Annika Schröder, UBS)
AML, Compliance advisory, Rogue trading prevention, automated compliance monitoring, KYC, Contract due diligence, and Information governance.