The Executive Guide to Artificial Intelligence

A Real Look at Artificial Intelligence  

Artificial intelligence (AI) applies computer systems to tasks that once required human intelligence. A long-standing debate within the AI community asks if AI should augment the human mind or replace the work it does. Either way, AI and automation will fundamentally reshape the workforce.

“Artificial intelligence is being used in businesses today to augment, improve and change the way that they work.”

AI can develop its abilities through supervised or unsupervised learning. In supervised learning, which is more common, people train AI systems using data and guide the system through making distinctions – like between pictures that show dogs and pictures that don’t. In unsupervised learning, systems start with data that mean nothing to them and identify patterns on their own.

“The biggest barrier to AI achieving escape velocity…is the overinflation of expectations.”

AI isn’t a hypothetical development that might appear sometime in the future. Businesses utilize AI today, and it transforms how they work. Many consumers experience AI today in the form of virtual helpers like Siri or Alexa. 

An “AI Framework”

AI has eight core capabilities. In this framework, four capabilities focus on capturing information and four focus on figuring out “what is happening.” The capabilities in the first set are: “speech and recognition, image recognition, search” and “clustering.” Image recognition involves tagging images and making distinctions among them. When machines capture information, they convert unstructured data (big data) to structured data. This requires speedy processors and a lot of training. Certain capabilities make AI immediately useful. For example, speech recognition lets people give machines direct commands.

“It’s impossible to get value out of something if it is not understood, unless it’s by some happy accident. In the world of AI there are no happy accidents; everything is designed with meticulous detail with specific goals in mind.”

The capabilities in the second set are: “natural language understanding (NLU), optimization, prediction” and “understanding.” The first three have applications in daily life. NLU goes beyond voice recognition. It includes a degree of understanding, AI’s ultimate capability, and it requires cognition. In optimization, an AI system transforms data from one form to another. Optimization requires the system to reach a goal, and it often applies algorithms and “cognitive reasoning” to solve problems. Prediction uses historical data to assess new data, for example, in making restaurant recommendations or in analyzing the risk factors in a loan application. Understanding, which isn’t yet commercially available, involves the machine’s ability to be consciously aware of what it does or thinks. 

“The first driver for the explosion of interest and activity in AI is the sheer volume of data that is now available.”

These eight capabilities work sequentially and synthetically. For example, speech recognition might recognize someone’s words, a prediction function might complete the requested search and optimization might solve a problem. 

The Rise of AI

AI’s development stretches back to the mid-20th century. Early work focused on so-called expert systems. Programmers mapped knowledge of a topic in a set of branching choices. User choices would guide the system down one branch or another – an approach still used today in applications like chatbots.

“Chatbots come in all shapes and sizes, which is a rather polite way of saying that there are really good chatbots but also very bad ones.”

AI passed through two “AI winters” – from 1974 to 1980 and from 1987 to 1993 – during which progress stagnated. Both winters occurred thanks to too much hype and too little funding. 

Several factors contribute to the contemporary rise of AI. The first is big data. Artificial intelligence needs “millions of examples” for training. Today’s continual use of social media and the Internet provides that data. Cheap storage, constantly increasing computing speed and ubiquitious connectivity drive AI and fuel the growth of cloud AI. Still, AI faces several barriers, including hype. People claim too much for AI. Excessive claims make people fear how AI might change business and the economy, or make their jobs obsolete. Most of AI’s tasks remain hidden from observers, and regulation could be a potent barrier to implementation.

Deep Neural Networks

AI depends on machine learning, that is, machines carry out difficult conceptual work, not people. Deep neural networks (DNNs) provide AI architecture. These networks have multiple layers – the more complex a problem, the more layers. DNNs have an input layer, an output layer and hidden layers in between where the difficult work gets done. Nodes in one layer connect to nodes in others. Each connection is weighted, which creates both weak and strong links. Weaker links produce undesired answers during training and don’t pass along as much information. As developers train networks, the weights adjust to reach an optimal level.

Associated Technologies

Practitioners can use AI alone or with other technologies. Cloud computing uses multiple remote servers linked via network. These servers store data. Cloud computing gives AI access to large, public data sets. Analysts then use cloud computing to process the data. Technicians use AI with robotic process automation (RPA), which employs technology to replace a series of human actions. RPA performs transactional work much more cheaply than people can, especially repetitive processes – like reading similar documents – and rules-based processes – like answering IT service requests.

“The reason machine learning is called machine learning is, rather obviously, that it is the machine, or computer, that does the learning.”

Robotics uses AI. Autonomous vehicles depend on AI to sort the information their sensors gather. Some firms use service robots to greet people. AI also comes into play in the Internet of Things (IoT) when devices transmit data directly to each other. When billions of devices transmit data, this generates massive big data – a natural place to implement AI. When AI can’t complete a task, humans intervene, such as in crowdsourcing or in cases when a task exceeds a system’s capability – say, reading handwritten text.

AI in the Real World

Some organizations use AI to improve customer service, for example, via chatbots. Simple chatbots can answer only yes/no or multiple-choice questions. But chatbots that receive extensive training through thousands of human-to-human chat conversations can answer questions and help customers make orders.

“One aspect where AI projects are generally trickier than ‘normal’ IT projects is with the dependency on data, and this challenge is particularly acute during the prototyping stage.”

Recommendation engine AI – such as Amazon’s and Netflix’s – applies data from customer purchases to suggest future purchases. AI processes claims quickly and improves functions customers will never see. British retailer Tesco sends robots  through its stores filming the shelves. The system uses image recognition to identify product gaps and lets staff know where to restock. The Israeli tech company Nexar uses information from a dashcam app to help people become better drivers. Business leaders who want to work with AI should identify the challenges their company faces and ask how AI can help. Leaders should consider AI and automation together and decide what they want such systems to accomplish. They can try a solution or application on a small scale, test it and then apply it more broadly. Businesses should align their AI strategies with their overall strategies.

The “AI Maturity Matrix”

Companies can adapt a Maturity Matrix – as originally developed by Carnegie Mellon University for use in IT – to evaluate their current level of AI integration. Traditional maturity matrices have five levels, but an AI matrix should have six, with “Level 0” referring to firms that still do everything manually. Companies, or individual departments or divisions may operate at five levels:

1. The firm applies traditional IT applications to specific tasks – like processing invoices – but hasn’t assessed AI’s impact or applied automation more broadly.
2. Most people still do most things manually, but at least one team has automated a task using scripting or macros.
3. A firm starts applying automation tools tactically to meet distinct goals.
4. Firms use a range of automation tools to apply AI to multiple processes.
5. A Level 5 firm applies AI and automation throughout its operations.

“AI Heat Map”

Organizations can create an AI heat map to identify the areas of their operations where applying AI is “desirable, economically viable and/or technically feasible.” Firms should start with their strategic objectives, and identify pressing challenges and places where sufficient data is available to enable AI-based solutions.

“If there is trust and transparency around the data that consumers find useful, then they are more likely to allow businesses open access to that, therefore increasing the utility even further.”

For your firm, list the possibilities and rate them by desirability and how feasible or viable they are. Rate all possibilities using the same scale, say 1–10, so your firm can compare rankings from different areas. As a firm chooses AI projects, it can develop a business case for each one. Calculate a project’s “hard benefits” – like reducing costs, mitigating risk, increasing compliance and customer satisfaction, reducing losses, and generating revenue. Also assess its “soft benefits,” such as its impact on the firm’s culture and its marketing.

“Prediction employs one of the core ideas of AI in that it uses lots of historical data in order to match a new piece of data to an identified group.”

Consider your options before implementing AI. Buying off-the-shelf AI software is simplest. Firms with special needs may build their own platforms and applications for greater control and flexibility. Only build a customized corporate system when your firm has large-scale, pressing needs. AI platforms fall somewhere between those two options. Huge companies such as Google and Amazon use platforms because they can train customized algorithms to handle specific tasks.

Implementing AI

As many firms implement AI, some are ready for the next level – “industrializing” AI. A successful firm will develop an “ecosystem” to support its AI and automation projects. Within that system, all vendors and technology should align with corporate strategy. Vendors should demonstrate technological expertise, experience and a cultural fit. The firm should form architecture teams to guide AI-related options through development and implementation to operations. AI-driven firms may add new leadership positions, such as a “chief data officer” and “chief automation officer.”

“Creating your first AI build, however small, is a key milestone for any AI program.”

AI’s primary challenge is dealing with poor data. With AI, accuracy isn’t as important as in traditional computing. “Data fidelity” matters more. Biased or inappropriate data can disrupt AI performance. Users can improve data by crowdsourcing or “cleansing” it to remove inaccuracies. AI also must cope with its own “bias and naïveté.” AI systems don’t understand social norms and may learn incorrect or inappropriate behavior. They need training via human intervention so they don’t find correlations that lack meaning.

“Sometimes AI simply isn’t up to the job. Sometimes you will need to pull humans into the loop to help complete the process.”

Choosing the “wrong technology” is also a risk. However, AI uses specialized applications that do just one thing and do it well. If a business assembles an AI system out of multiple components, it should be able to replace any single component to improve overall system function. As businesses adopt AI, they could become overly dependent on it. This overdependence can be practical (can users tell if answers are correct?) or philosophical (will humans forget how to think if the machines think for them?). There’s also a risk of “malicious acts.” For example, if a bank implements voice recognition as part of its security system for account holders, an AI system could mimic those voices to gain access to the accounts. Some users have directed AI to  “socially engineer people’s behaviors.” For example, bots can post messages on social media to redirect political conversations.

AI’s Future

Image recognition will continue to improve – with better image tagging and facial recognition. The use of voice recognition use will expand more into business-to-business interactions. Improvements in microphone technology and algorithms for speech recognition will make real-time voice transcription more accurate and efficient. Search software will improve. NLU will gain capability, especially in real time. The lack of “properly labeled, high-quality data sets” will continue to be a constraint. As proper data become more available and AI improves – through reinforcement learning – at using unlabeled data, optimization will continue to develop.