Why is it so hard to deploy AI?

We all know AI will transform everything

There might be a corporate executive somewhere who hasn’t yet concluded that over the next few years, artificial intelligence (AI) will transform their organisation and their industry. If there is, they are very unusual. However, despite this general acknowledgement of its importance, many companies are struggling to deploy AI. How should they do it?

Daniel Hulme has some suggestions, which he explains in episode 8 of the London Futurist podcast. Daniel is the founder and CEO of Satalia, which until recently was London’s largest independent AI consultancy, helping companies solve hard problems with data and algorithms. Earlier this year, Satalia was acquired by WPP, the global communications group, and Daniel is now WPP’s Chief AI Officer.

Deep learning is rare in practice

Perhaps surprisingly, given what he does for a living, Daniel argues that deep learning – the branch of AI which has swept all before it in the last decade – is not being used effectively in many companies. The definition of AI hints at one of the reasons: it is “goal-directed adaptive behaviour”. Executives are understandably wary of systems that will adapt themselves in operation: they don’t want to implement systems that they cannot understand or predict. They certainly don’t want systems which will make mistakes, but making mistakes is an essential part of the learning process.

Six types of deployment today

According to Daniel, there are six ways that companies are trying to deploy AI today. The first, automation, is usually not genuinely AI, in the sense that there is generally no learning going on. Instead, software just follows pre-programmed “if … then …” instructions. The second is generative AI systems, like the remarkable Dall-E and Midjourney systems which create photorealistic imagery from written instructions.

The third category of AI deployment is humanisation, which means using technologies like DeepFakes and natural language models to mimic the behaviour of humans. The fourth category is analysis, where deep learning is deployed to extract insights from data – finding correlations that humans could not. The fifth is the physical or cognitive augmentation of humans, and the sixth is complex decision making, or optimisation, which used to be known as operations research. Daniel claims that this is the most promising of the six categories, because “companies don’t have AI problems, they have decision problems”.

The unfair advantages of Big Tech

There is one part of the economy where AI is genuinely being deployed at scale, and that is among the tech giants: Google, Apple, Amazon, Microsoft, Facebook and Amazon in the US, and Baidu, Alibaba and Tencent in China. These companies have unfair advantages. They have access to vast amounts of data, and they can afford to hire and retain the rare and very expensive AI experts who know how to compile, manipulate and apply that data, using sophisticated algorithms. They have the financial resources to pay for the massive amount of computation power that modern AI requires. They also understand the limitations of today’s AI, and how to combine human decision-making with machine intelligence.

Clients in less exalted organisations often ask Satalia if it can deliver some “quick wins” using AI, which could build the organisation’s confidence to deploy the technology on a larger scale. Daniel replies that this is a false hope. Enormous returns on investment in AI are possible, but the sad fact is that implementing AI is not easy. Most projects fail because of under-investment, or because of mis-understandings about what it is actually capable of.

Data, data everywhere, but no insight in sight

A typical cause of failure is the idea that gathering as much data as possible in data lakes, and making it all mutually compatible will lead, almost by magic, to invaluable insights. These insights will in turn lead to increased revenue, reduced costs, and sustainable competitive advantages. Success generally lies in the opposite direction. Companies should begin by identifying the frictions in their business; the pain points. They should prioritise these and develop hypotheses about what information could help ease the friction.

Daniel offers the example of a low-cost airline which spent a considerable amount of money assembling all the data it could find about its customers: what day of the week they bought tickets, their seat preferences, the number of people they flew with, and so on. They hoped this would help them understand why they sometimes chose to fly with a rival airline. When asked why they themselves might switch, the executives cited the rival’s pricing levels, but this was not one of the data points they were collecting, mainly because it was harder to do so.

“Tech debt”

A more subtle example is a short-term loan company which wanted to predict which customers would fail to repay their loans. Eventually, they discovered that the best indicators were the use of a particular font in their application forms, and taking a longer time than average to complete the forms. The font was most commonly found on gambling websites, and it turned out that the customers taking a long time were drinking. Gambling and drinking are reliable indicators of poor loan performance.

Daniel argues that the people working with customers in those organisation knew about these indicators, but nobody working on the AI systems thought to ask them. The upshot of situations like this is that companies are spending substantial sums on data lakes and AI systems which they will quietly abandon in a few years’ time. He calls this unfortunate process the build-up of “tech debt”.

Getting it right yields huge results

All is not doom and gloom, however. Satalia has completed numerous successful projects for its clients, and Daniel is able to share some of the success stories. Satalia built a last-mile delivery system for Tesco from scratch within six months, which is reckoned to reduce the miles their vans drive by around 20 million miles a year. For PwC, Satalia developed a system to allocate consultants to jobs more efficiently, taking into account their locations, their skills, and their preferences. The maths underlying these projects is extraordinary. For instance, there are more possible allocations of 60 consultants to 60 jobs than there are atoms in the universe, and PwC has many thousands of each, not just 60. Within the systems designed for Tesco and PwC, there are multiple optimisation algorithms and multiple ML algorithms.

AI is not easy to deploy, but when organisations get it right, the results can be spectacular. Which is why, in the coming years and decades, it will transform every industry.

Related Posts