Mostly driven by growth in the IoT, and the widespread use of internet, social media and mobile devices to perform search, send text, email and capture images and videos, the amount of data that we are producing on a daily basis is startling.
Consequently, companies are turning to data analytics and AI technologies to help them make sense of all the data at their disposal, predict the future, and make informed decisions that drive enterprise performance.
Although adoption of analytics and AI systems is increasingly extending in more mission-critical business processes, the implications of these emerging technologies on busines strategy, management, talent and decisions is still poorly understood.
For example, the single most common question in the AI debate is: “Will adoption of AI by businesses lead to massive human job cuts?”
Borrowing lessons from historical technological advances, yes, certain jobs will be lost, and new ones also created. However, machines are not taking over the world, nor are they eliminating the need for humans in the workplace.
Jobs will still be there albeit different from the traditional roles many are accustomed to. The majority of these new roles will require a new range of education, training, and experience.
For instance, nonroutine cognitive tasks demanding high levels of flexibility, creativity, critical thinking, problem-solving, leadership, and emotional intelligence do not yet lend themselves to wholesale automation.
Analytics and AI rely on data to make predictions
As more and better data is continually fed to the machine learning algorithms, the more they learn, and improve at making predictions.
Given these applications search for patterns in data, any inaccuracies or biases in the training data will be reflected in subsequent analyses.
But how much data do you need? The variety, quality and quantity of input, training and feedback data required depends on how accurate the prediction or business outcome must be to be useful.
Training data is used to train the predictive algorithms to predict the target variable, while the feedback data is used to assess and improve the algorithm’s prediction performance.
Undoubtedly, advanced analytics and AI systems are only as good as the data they are trained on. The data used to train these learning algorithms must be free of any noise or hidden biases.
You therefore need to understand how predictive technologies learn from data to perform sophisticated tasks such as customer lifetime value modeling and profitability forecasting.
This helps guide important decisions around the scale, scope and frequency of data acquisition. It’s about striking a balance between the benefits of more data and the cost of acquiring it.
Humans and machines both have shortcomings
In the context of prediction, humans and machines both have recognizable strengths and weaknesses.
Unless we identify and differentiate which tasks humans and machines are best suited for, all analytics and AI investments will come to naught.
For instance, faced with complex information with intricate interactions between different indicators, humans perform worse than machines. Heuristics and biases often get in the way of making accurate predictions.
Instead of accounting for statistical properties and data-driven predictions, more emphasis is often placed on salient information unavailable to prediction systems.
And, most of the times, the information is deceiving, hence the poor performance.
Although machines are better than humans at analyzing huge data sets with complex interactions amidst disparate variables, it’s very crucial to be cognizant of situations where machines are substandard at predicting the future.
The key to unlocking valuable insights from predictive analytics investments involves first and foremost understanding the definite business question that the data needs to answer.
This dictates your analysis plan and the data collection approaches that you will choose. Get the business question wrong, conclusively expect the insights and recommendations from the analysis to also be wrong.
Recall, with plentiful data, machine predictions can work well.
But, in situations where there is limited data to inform future decision making, machine predictions are relatively poor.
To quote Donald Rumsfeld, former US Secretary of Defense:
There are known knowns. These are things we know that we know. There are known unknowns. That is to say, there are things that we know we don’t know. But there are also unknown unknowns. There are things we don’t know we don’t know.
Donald Rumsfeld, former US Secretary of Defense
Thus, for known knowns, abundant data is readily available. Accordingly, humans trust machines to do a better than them. Even so, the level of trust changes the moment we start talking about known unknowns and unknown unknowns.
With these situations, machine predictions are relatively poor because we do not have a lot of data to ingest into the prediction model.
Think of infrequent events (known unknowns) that occur once in a while, or something that has never happened before (unknown unknowns).
At least for infrequent events or happenings, humans are occasionally better at predicting with little data.
Generally so because we are good at comparison and applying prudent judgement, examining new situations and identifying other settings that are comparable to be useful in a new setting.
We are naturally wired to remember key pieces of information from the little data available or the limited associations we have had in the past.
Rather than be precise, our prediction comes with a confidence range highlighting its lack of accuracy.
Faced with unknown unknowns, both humans and machines are relatively bad at predicting their arrival.
The simple truth is that we cannot predict truly new events from past data. Look no further than the current Brexit conundrum.
Nobody precisely knew the unintended consequences of the UK leaving the EU. Leavers and Remainers both speculated as to what the benefits and disadvantages of leaving the EU maybe.
Of course, nobody knows what will happen in the future but that doesn’t mean we can’t be prepared, even for the unknown unknowns.
In their book Prediction Machines: The Simple Economics of Artificial Intelligence, Ajay Agrawal, Joshua Gans, and Avi Goldfarb present an additional category of scenarios under which machines also fail to predict precisely – Unknown Knowns.
Per the trio:
Unknown knowns is when an association that appears to be strong in the past is the result of some unknown or unobserved factor that changes over time and makes predictions we thought we could make unreliable.
PREDICTION MACHINES: THE SIMPLE ECONOMICS OF ARTIFICIAL INTELLIGENCE
With unknown knowns, predictive tools appear to provide a very accurate answer, but that answer can be very incorrect, especially if the algorithms have little grasp of the decision process that created the data.
To support their point of view, the authors make reference to pricing and revenue analysis in the hotel industry, although the same viewpoint is applicable elsewhere.
In many industries, higher prices are analogous to higher sales, and vice versa.
For example, in the airline industry, airfares are low outside the peak season, and high during peak seasons (summer and festive) when travel demand is highest.
Presented with this data, and without an understanding that price movements are often a function of demand and supply factors, a simple prediction model might advocate raising different route airfares to sell more empty seats and increase revenues. Evidence of causal inference problems.
But, a human being with a solid understanding of economics concepts will immediately call attention to the fact that increasing airfares is unlikely to increase flight ticket sales.
To the machine, this is an unknown known. But to a human with knowledge of pricing and profitability analysis, this is a known unknown or maybe even a known known provided the human is able to properly model the pricing decision.
Thus, to address such shortcomings, humans should work with machines to identify the right data and appropriate data analysis models that take into consideration seasonality and other demand and supply factors to better predict revenues at different prices.
As data analytics and AI systems become more advanced and spread across industries, and up and down the value chain, companies that will progress further are those that are continually thinking of creative ways for machines to integrate and amplify human capabilities.
In contrast, those companies that are using technology simply to cut costs and displace humans will eventually stop making progress, and cease to exist.
To be highly effective the implementation of AI is a multi-discipline exercise that will require resources from many parts of the business. A good example of this can be illustrated when using AI to assist in auto invoicing and payment applications. The sales department, manufacturing and shipping departments will provide data that allows these two functions to operate autonomously. The data from these departments will be incorporated into algorithms that function to determine how much, when and to whom to send an invoice; and, how to apply payments when the bank reports them as received.
In addition to the speed factor, AI now is being used to replace repetitive, linear tasks and increase our information output capacity. Both uses have wide implications for the CFO, including;