by James Duez, Rainbird
4 February 2020
Gartner have identified AI augmentation as one of the top five emerging technology trends for 2020 and PwC are predicting a potential $15tn boost to global GDP from AI by 2030.
AI adoption continues to be scrutinized and more work is required to deliver both the transparency and the transformational value demanded by today’s enterprise. To achieve that, humans must continue to work in close tandem with machines.
While the skyrocketing use of algorithms might lead us to assume that we are less dependent on humans, the opposite is true. Effective results require a nuanced collaboration between data scientists and subject-matter experts. The value is still in automation that effectively leverages human expertise.
The limitations of ‘data up’
Data offers us a powerful way to scrutinize the past in order to forecast the future, but there are serious limitations. The majority of organizations are not data-ready, lacking a volume of good clean data, or the skills to work with the data optimally. Effective statistical outcomes require not just data scientists but also domain experts.
Transparency is the biggest barrier to the growth of machine-learning, especially in regulated markets. Their murky inner workings can expose firms to cost and risks. It can be a legal and ethical minefield. Because machine learning is statistical, their models can only be understood by a handful of data scientists – at best. Machine learnt models tasked with making predictions in areas such as fraud prevention and drug development operate on the basis of complex probabilities and correlations beyond the comprehension of most laypeople. If they are not careful, innovators run the risk that these algorithms may operate in illegal or immoral ways. This may well be without their explicit knowledge. It can be hard to defend criticism or litigation when you cannot explain the methodology used. Such models can create a legal ‘grey area’ and significant liability.
For example, much is made of the fact that machine-learning systems often develop irrational prejudices, from gender-recognition cameras that only work on white men to algorithms that display ads for lower-paying jobs to women. This is because AI training data often reflects the biases of its human compilers. They are molded entirely by their learning environment, and make use of hidden variables to make their decisions.
Results start with human knowledge
A more pragmatic and arguably ethical alternative is to build enterprise intelligent automation around the expert knowledge that businesses already possess – from their best people. While statistical models can be powerful, something is lost if the corpus of human expertise on a domain is not considered as an effective starting point.
Distinct from data-driven statistical machine learning models, it is now possible to inject our automation models with the crucial ingredient that has so far been missing: the best of human expertise.
Most organizations care about three things; efficiency (containing costs), effectiveness (improving customer outcomes) and innovation (generating new products and services to increase revenue), and AI has promised to transform all three.
In nuanced, high-volume roles like credit risk analysis, tax and audit, underwriting and claims – skilled human expertise is often in short supply. What’s more, it is well established by the research of Daniel Kahneman and others that bias and noise significantly disrupt human-decision making. More often than not, a number of factors including lack of expertise, time pressures and poor data hampers the quality of this kind of decision-making at scale.
These reasoning-intensive decision-making tasks are too complex to be automated with Robotic Process Automation (RPA) as there are too many permutations and variables, while the regulatory oversight in these sectors makes machine learning poorly suited due to its lack of transparency.
So how then to find the best balance of human and machine? The key is to find ways of synthesizing a company’s human expertise and scale it, making it consistently available across the enterprise. This can only be achieved with a knowledge mapping approach that goes well beyond what has been possible with traditional decision-trees. The next generation of intelligent automation technology can enable the expert themselves to create holistic models of knowledge. Using probabilistic methods, these can be turned into human-like machine intelligence models that can get the best out of data – looking at each case through the same lens as the most experienced domain expert would, only faster, with more accuracy and with an audit trail.
This next generation of intelligent automation is becoming essential. There is huge value to increasing the speed, quality and transparency of decision-making processes enabling organizations to become more compliant, more trusted and ultimately more profitable.
The same can’t be said for machine-learnt models, which produce potentially biased unethical or incorrect results that can’t easily be explained to regulators or consumers. Sustainable, transparent automation should be the goal, and we won’t get there by continuing to opt for short-termist ‘black box’ models that simply mine data for quick and opaque results.
Without the human element, statistical techniques can be limited to providing insights but not the exponential value via truly complex decision-making that are required by today’s regulated companies.
Generating machine intelligence from human expertise
Gartner estimates that AI augmentation will create $2.9 trillion in business value in 2021 alone. To achieve this value, enterprise must wake up and recognize what is now possible. They must look beyond the pure machine-learnt approach and also digitize the best of their human expertise.
Over the past few years, the early adopters have already deployed numerous solutions delivering high quality and transparent automated decisions in credit, tax, fraud, law, insurance and healthcare. The result is exponential – improving quality, reducing costs and heralding a generation of new digital products and services.
James Duez is CEO at Rainbird, a British company that develops an AI-powered, automated decision-making platform