Bird & Bird’s Toby Bond spoke at the AI.Tech World conference held at Olympia, London on 29 November 2017, together with speakers from Dixons Carphone and RainBird Technologies in a panel session exploring what artificial intelligence (AI) means to different industries.

Toby explored some of the legal issues which will affect the way different industries draw the boundary between tasks performed by a human and those performed by AI. The controls on solely automated decision making contained in the forthcoming General Data Protection Legislation (GDPR) suggest a greater emphasis will be placed on using AI to augment human decision making, rather than removing the human element from the loop entirely. Current uncertainty regarding liability and risk allocation for AI systems will also prove to be a challenge when using AI to replace humans in roles which involve interaction with the physical world; at least in the short term.

James Duez, the co-founder and chairman of RainBird Technologies, took as his theme the fact that all industries needed to do more with less. AI would be transformative for many businesses; it would bring costs down, allow more transparency and improve customer services. Although AI would take over from rote tasks, he preferred to think of AI as augmenting processes; with humans and machines making decisions together. He gave as an example AI being used to identify credit card fraud where humans were still needed to identify false positives. Transparency in why the AI was making a decision was also key; decisions had to be explainable and understandable.

Anthony Morris, Director of Strategy and Analytics at Dixons Carphone, said that AI already made his business more efficient; for example, bots working alongside staff helped find products more quickly thus improving the customer experience. He also noted that through mining customer data, retailers could anticipate customer demand; he even foresaw a time when retailers would know their customers’ habits so well that they would ship products to them before any request was made, dealing with the returns if and when they were made.

In all of this, ownership of data was an obvious concern. The panel agreed that access to large, high quality datasets was often a key enabler to develop many types of AI based solutions. In some industries, practices for ownership and licensing of data are well established. In others, there will need to be some learning in order to navigate new business and licencing models, plus the patchwork of IP rights protecting data.

There were multiple concurrent sessions throughout the day and this note only picks up on a few of them which were of particular interest to Toby and to Katharine Stephens who also attended.

AI in Financial Services

The immersive lab on the role of AI in the future of the financial industry was one such session. The panel was made up of speakers from Barclays, Smith and Williamson, IBM Watson and Accenture. The last couple of years have been setting the scene for AI in financial services, showing that neural networks are stable and can work. In the next two to three years, there will be greater implementation of AI techniques and new opportunities for their use will be identified. Asian markets are ahead of Europe in their use of AI, but nevertheless there is a lot of activity in Europe and many options to chose from. Data privacy was seen as one of the key issues that financial institutions, particularly banks, had to get right; they had to show their customers that they were a “trusted band”. Another aspect of this was traceability, that is the ability to show why a decision was made, both from the point of view of customers, but also regulators. When asked what developments the panel saw coming in the next year, they named greater use of natural language processing, IoT and visual recognition, much of this resulting in the customer’s experience being made more intuitive and engaging.

AI in Healthcare

In the immersive lab on AI in healthcare, data privacy and security were again a significant topic for the panel which was made up of speakers from the NHS Institute for Innovation and Improvement, Babylon Health, Diabetes.co.uk and myrecovery. With wearable tech and IoT, more and more information was potentially available regarding an individual’s health. The GDPR was seen as being very important in protecting that data, allowing individuals to control the way in which it will be used. The panel also considered the early successes for AI in healthcare, for example, in image recognition in oncology and cardiology (reference was made to Arterys imaging system which had recently won FDA approval, see here). However, application was limited at present because of fragmentation of datasets and the difficulty of training AI systems. As Mark Tsimelzon from Babylon Health explained, when training AI in diagnostics, the first step was to capture a doctor’s knowledge using semantic networks, which was not an easy task, then once the probabilities of a symptom relating to a disease have been analysed, the machine learning aspect could be added. Such diagnostic tools, once built, were not seen as replacing doctors, but rather complementing their work. The topic of chat bots was then discussed, the general view being that, at present, patients did not like interacting with a chat bot, instead they much preferred talking to a person. However, it was thought that patients would get used to them as the chat bots improved and that, certainly where resources were limited, they could significantly assist in the provision of healthcare.

AI and Business Problems

The immersive lab on applying machine learning to business problems also gave a fascinating insight into the strategic issues which affect how AI can be used to generate business value. The panel, made up of speakers from Warwick Business School, Prudential, Wandera and BT, explored the challenges of applying perceptual and statistical AI techniques in an enterprise context. The difference between AI techniques and existing data based modelling is "closing the loop" such that the system can adapt itself over time and take into account the effect that its own decisions are having on the system it is modelling. A key skill is also getting the input data right, which often evolved combining datasets from multiple sources, i.e. "data fusion". There are also challenges around deploying AI systems within existing enterprise architecture, as those building the AI need to understand existing business processes. The panel agreed that AI can bring higher quality, efficiency and performance to things businesses do already, but also provides the opportunity to extend services currently available to only the very top end of the market to a much wider customer base. An AI based private banker might well find a market amongst the emerging middle class in Asia.

The role of Natural Language Processing

The session on the role of natural language processing in enhancing business process and customer experience also provided an insight on the current opportunities and limitations of this core AI technology. Speakers from Web3//IOT, London Luton Airport, First Utility, Rainbird Technologies and Simply Business discussed the areas and applications where chat bots are currently best able to succeed and those areas where they fall down. Thinking in terms of the functional role the chat bot will play is important as it allows you to avoid negative customer experience where the chat bot is asked to handle situations beyond its core competence. An intriguing possibility to move beyond the limitation of current systems would be to merge human and chat bot based responses into a single conversation flow, with the chat bot handing over to a human when the conversation moved beyond its capabilities. The panel also emphasised that natural language processing is a key enabler for other types of AI technology as it can be used to interrogate unstructured datasets and provide the raw input for machine learning techniques. This will greatly help many businesses to harness the value of their existing business process data.