From data-driven to AI-driven decision making. Part 9 of the series ‘Perspectives on Artificial Intelligence’.
Of all the current technological developments, artificial intelligence is both the most profound and the least understood. We are witnessing impressive new applications, but can hardly foresee their impact on people, organisations and society. In this series of blogs – Perspectives on Artificial Intelligence – we investigate not only the opportunities, but also the intended and unintended consequences.
Decision making when AI systems are smarter than people
According to many managers and executives, the phenomenon of data-driven decision making is little more than pouring old wine into a new bottle; their decisions have always been based upon facts and figures. If data analyses are limited to historical data trends and their extrapolation to the future, this is indeed nothing new.
But new information systems that use artificial intelligence (AI) and machine learning go beyond supporting the decision-making process. When they come up with better solutions to problems than humans, acceptance of an AI system’s larger, independent and autonomous role in the decision-making process quickly escalates. Just as doctors are increasingly using AI systems to determine diagnoses and treatment plans, managers can do the same.
The Google DeepMind AI system that beat the Go world champion was also able to reduce energy consumption by cooling Google’s data centres 15% more efficiently than attempts made by human experts. Some financial institutions use AI to give personal advice to customers and prevent money laundering. At insurance companies such as Lemonade both underwriting and claims handling are managed by AI as often as possible, allowing employees to concentrate on less generic cases. Credit application assessment is also increasingly being carried out by AI, while manufacturers leave logistics and stock management to smart systems.
From descriptive to prescriptive analysis
Thanks to AI, and in particular to machine learning, business information systems are being created that not only analyse the past, but also look ahead. The following figure by McKinsey illustrates the development from descriptive to prescriptive analytics.
Descriptive analytics concerns historical and current results and is used by controllers for performance management purposes; what happened and why? This type of analysis is descriptive and diagnostic.
Predictive analytics investigates what might happen and suggests alternative outcomes using probability calculations and risk assessments. This provides better insight into product sales, customer behaviour and production developments. Thanks to machine learning and neural networks, both structured and unstructured data can be analysed.
Prescriptive analytics goes a step further and indicates which decisions need to be taken and what needs to be done to achieve an objective. With this type of analytics, it is also possible to automate decisions and minimise human intervention.
This analysis is not based on pre-programmed instructions (if this, then that). Through machine learning, an AI information system can detect data patterns and define predictive and recommendation algorithms autonomously.
Application at strategic and operational levels
I do not advocate robotic leadership, but I do believe we should apply more intelligence in decision-making through the use of AI information systems. This can be advantageous both at a strategic level (capital allocation) and at an operational level (optimisation). When, for example, you are allocating budgets for research, product development and marketing, AI can help with analysis of your firm’s historical data as well as competitor decisions and results.
AI can assist a CFO to optimise the capital structure of the company by taking into account not only internal financial metrics but also financial market developments, investor sentiment and generic economic indicators.
Internal email traffic analysis can say more about employee engagement and resignation risks than an annual satisfaction survey, with due observance of the privacy rules, of course. And analysis of social media provides more insight into customer perception than more traditional market research methods.
Other applications are margin maximisation and stock management optimisation through the analysis of sales trends, price sensitivity, competitor actions, marketing budget, sales force size and customer sentiment.
During the financial crisis we often heard about top executives who were unaware of decisions made at lower levels within their organisation, resulting in financial risk-taking without sufficient oversights and costing the company involved a great deal of money. With the correct deployment of Data Analytics, this scenario could have been avoided.
Data Quality Management
The ultimate AI system creates a digital model of the organisation; this requires a significant amount of data. Based on information concerning input (the deployment of people and assets) and output (results), patterns can be identified and recommendations made.
The amount of data collected and stored worldwide is growing at an exponential rate. According to SINTEF, 90% of worldwide available data is less than two years old. For organisations, however, it is not easy to make relevant data accessible in a relatively structured manner. Data Quality Management is a critical function. In this context, a new senior post, Chief Analytics Officer, is being created.
Why we still need executives
Although AI systems are designed to support executives, it is easy to foresee that those executives who deviate from advice given by a prescriptive information system will have to justify their actions. This would add a new dimension to the Corporate Governance principle ‘comply or explain’.
Of course, decisions remain that AI systems are unable to take or prepare. To design an algorithm you need countless examples and these are simply not available for less generic situations. Furthermore, executives and directors might possess information that alters existing circumstances, such as new laws and regulations or technological breakthroughs that are not included in the – by definition historical – dataset. We therefore need managers to make decisions when information is incomplete and when future results can not be deducted from historical data.
If routine decisions are taken over by AI, competencies such as responsibility, empathy, creativity and originality will become more important in the field of management than operations and finance skills. Asking questions becomes more important than giving answers.
Mind the HiPPO
The biggest obstacle for AI-driven decision making within an organisation is the HiPPO phenomenon, an acronym of the ‘highest paid person’s opinion’. As a result of the Authority Bias, many people find it difficult to contradict the highest paid member of the group.
On the other hand, it is bad management to base decision making solely on the subjective opinion of a person, whoever that may be, when data and AI information systems are available.