Everyone from Elon Musk to Vladimir Putin suddenly seems to be talking about artificial intelligence (AI) and how it will change the world – for better or worse.

But what does AI really mean once you set aside the hype and fears of "Terminator"-style killer robots? What will it mean for business? How quickly will it arrive? Where can it be applied? And how should companies be responding to it?

What exactly do people mean when they talk about AI in 2018?

Artificial intelligence is a very broad term that dates back to 1956 and describes efforts to get computers to do things that could previously only be done by humans. It occurs frequently in science fiction, yet today's excitement is focused on one particular subfield of AI known as "machine learning," which involves teaching machines to do things by example – as opposed to, say, "expert systems," which rely on rules and knowledge distilled from human experts. The excitement, in fact, is mostly centered on a specific machine learning technique known as "deep learning," in which software simulations of simple models of the human brain are trained to do things by showing them large numbers of examples. "Neural networks," as these simulations are called, have been around for a while, but "deep" networks, which are more sophisticated and can be trained to recognize more subtle differences, have become far more capable in recent years.

In other words, deep learning is just one very specific example of a subfield of AI, but the excitement around this particular subfield stems from its ability to handle a very wide range of problems, from image recognition to language translation to transcribing speech. "Machine learning is now superhuman in its ability to do certain classes of very specific tasks," says John Giannandrea, head of Google's machine learning team. But he stresses that deep learning's success does not mean AI is a solved problem. "The idea that one corner of computer science is good for speech recognition, image detection and self-driving cars has got people really excited – perhaps overly excited." The more fundamental questions regarding the nature of consciousness and intelligence remain as impenetrable as ever.

Why has AI suddenly taken off in the past couple of years?

The recent rise of AI is the result of rapid progress in machine learning, and deep learning in particular. There are three reasons behind this. First, the internet – and, more broadly, digitization – has provided an enormous volume of data that can be used for training. Second, researchers figured out more efficient training algorithms for use with larger – "deeper" – neural networks. And third, they figured out how to get graphical processing units – the specialist chips found in video game consoles – to run deep learning software instead.

This provided a hundredfold boost in performance when it was first done in 2009, and chipmakers have since devised new chips optimized for deep learning; the share price of Nvidia, the leading maker of GPUs, has risen more than tenfold in the past four years as a result. Deep learning's power was first apparent in 2012 at the annual ImageNet competition, which pits image recognition systems against each other. That year a deep learning system demonstrated an unprecedented improvement in accuracy, easily winning the competition. That prompted widespread adoption. "The applications are so broad. There are so many ways to use it," says Nvidia CEO Jensen Huang from his self-driving car, which contains two of his company's chips, as it drives him to work in Silicon Valley.

There's been excitement about AI before. Is this time really different?

Yes. Building software systems that learn from data, rather than following explicit rules hand-coded by human programmers, is not new. Previous bursts of enthusiasm around AI always ran out of steam as the technology proved difficult to scale up or deploy, however, resulting in fallow periods known as "AI winters." This time around, deep learning systems are powerful enough that they have been deployed on a large scale by internet companies, and are now used on a daily basis by billions of people, most of whom are entirely unaware of it. Deep learning systems underpin Google's search engine and translation services, suggest replies to emails and recognize speech for its smartphone assistant. Facebook uses deep learning to recognize and help tag people in uploaded photographs, and to figure out which posts and advertisements to show to users. Deep learning powers Apple's Siri and Amazon's Alexa. Chinese internet giants Baidu, Alibaba and Tencent are all using it too. In short, and in contrast to previous AI technologies, it has proven to be applicable to a wide range of tasks and is reliable enough to be embedded in systems that people use every day. "You're using it when you talk to your phone, when you search for something on the internet. You're using it already in so many ways," says Richard Socher, an AI researcher who is now chief scientist at Salesforce. Even if the technology failed to advance any further, says Yoshua Bengio, a pioneer in the field of deep learning, there are still lots of areas where it can be usefully applied. Another deep "AI winter" seems unlikely.

What kinds of tasks can AI actually do? How does it approach those tasks?

Nearly everything that deep learning is currently being used for, from image recognition to language translation to speech transcription, is actually a variation of the same underlying task, from a technical perspective at least. In each case a neural network is trained by exposing it to millions of examples (inputs) for which the correct answer (output) is already known. For image recognition, that means training a network with millions of labeled images (this is a dog, that is a cat); for speech recognition, millions of sound clips are used, each tagged with the correct transcription. Once the network has absorbed enough examples, it can correctly predict the right output for a previously unseen input. This configuration of deep learning, called "supervised learning," is the most widely used in business. "Most of the value of machine learning today is supervised learning," says Andrew Ng, a deep learning pioneer who has worked as head of AI at Google and Baidu. Spam filtering, credit scoring, handwriting recognition, analyzing medical scans or teaching a self-driving car to read road signs can all be implemented using supervised learning. More generally, any sufficiently large labeled data set (where the correct output for millions of inputs is known) can be used to train a deep learning system. So now ask yourself: What labeled data sets does your company have?

Where do I start if I want to embrace AI in my business?

Deep learning's dependence on vast amounts of training data explains why internet giants were its earliest and most enthusiastic adopters. They have access to enormous amounts of data that can be used to train systems. For companies that are used to processing a lot of data – for example, those in retail, telecommunications or financial services – moving from "big data" analysis to the adoption of machine learning is an obvious step.

For other firms, the adoption of AI techniques depends on first being able to gather, process and analyze internal data effectively; companies with poor analytics capabilities or flaky data management will struggle. But the opportunity is clear. "Today every company has processes that can be managed, optimized or augmented by AI," says Antoine Blondeau, co-founder of Sentient Technologies, an AI startup that has attracted more than $140 million in funding. Elevator companies have data about the reliability of elevators; carmakers have data on the behavior of cars. "I think you're going to see deep learning approaches being adopted in literally every single industry, from engineering to marketing, from sales to manufacturing," says Jensen Huang. But you need lots of data first.

Do big companies with lots of data have an insuperable advantage?

Because data is necessary in order to train machine learning systems, access to lots of data can give companies a big advantage. Google is the most popular search engine, which means it has the largest volume of search queries to analyze and the most raw material to feed into machine learning systems that make its search engine even better. The same dynamic may play out in other industries: The more data you have, the better you can make your product, attracting more users and thus generating more data. This is known as a "data network effect," and it may mean that first movers in some fields end up with a huge head start over their rivals that will continue to grow. But there is a counterargument: For many tasks, abundant training data can be found online. "The internet is full of data. There's enough data to build AI out there," says Yoshua Bengio. Opensource data sets that can be used for training exist in several fields.

Data can also be gathered from the real world or generated in a virtual environment: Self-driving cars are being trained using video from dashboard cameras or images of imaginary streetscapes generated by video game engines. DeepMind's AlphaGo system, which defeated the world's best players of Go, a board game popular in Asia, began its training by analyzing past games and then improved by playing against versions of itself. Future machine learning techniques may someday allow systems to learn from fewer examples. But for now, the more data you have, the better – and access to training data that nobody else has can give you an advantage, though not necessarily an insuperable one.

Will every company end up using AI eventually?

Yes. Just as all companies now use electricity and the internet, they will all end up using AI too. But as with those previous technologies, there are many levels of adoption: Only the biggest companies generate their own electricity. Similarly, for companies that have data processing at the heart of their business, machine learning expertise will become a core competence, requiring teams of specialists. Other firms will adopt the technology at arm's length as AI features are added to devices, software and services they already use, from smartphones to email systems to e-commerce engines. "A lot of what you'll see in the next five years is incremental, and imperceptible to most people, but things will work slightly better," says Demis Hassabis, co-founder of DeepMind. "You'll see more and more everyday things becoming more alert to your context, a little bit smarter."

The rapid rollout of AI technology by internet giants is not representative, in short. Most companies do not have vast troves of data, thousands of engineers and billions of users. A recent survey of corporate executives by IBM found that 38% of companies are still "observers" that have yet to adopt AI; only 11% of companies have made significant investment in the technology. As with electricity and the internet, adoption will take time.

What does all this mean for my employees, for hiring and training?

Many observers are gloomy about the long-term impact of AI on jobs. "Seventy percent of jobs will be destroyed, because they can be performed cheaper or better by robots and AI," predicts Bruno Maisonnier, a robotics entrepreneur best known as the creator of the humanoid "Pepper" robot. Others disagree, saying that technology has always created more jobs than it destroys. Either way, mass unemployment is not in the cards just yet.

The more prosaic reality is that AI technology cannot yet replace entire jobs: It automates or accelerates some tasks, instead. Most jobs consist of a mix of tasks, and that mix is likely to change more rapidly in the future. That in turn will require workers to acquire new skills and companies to take a more organized approach to regular retraining as skills become more "perishable." The spread of computers into the workplace since the 1980s already requires people to learn new skills and new tools every so often, and AI will simply accelerate the process. Automating some parts of jobs but not others will also emphasize the importance of soft skills such as empathy and social interaction – the things machines cannot do (not yet, at least) – and which employers will increasingly value. Rather than AI directly displacing humans, it seems more likely that jobs that involve using AI will displace jobs that do not. A report from investment bank UBS summarized the situation as follows: "Don't fear the robots, but be prepared to change jobs."

Where is AI going next, and what will the impact likely be?

All kinds of new machine learning architectures are being developed that go beyond what existing systems can do. "Unsupervised learning" can find data patterns even when you don't know what you're looking for. "Reinforcement learning" is an AI technique that allows machines to develop strategies to solve difficult problems, like driving vehicles, controlling robots or playing complex games. "Generative adversarial networks" can take examples of music, paintings or photographs and then generate a convincing imitation, or recreate one in the style of another. AI techniques will also provide increasingly rich new ways to interact with tools and data, through speech and conversational interfaces and augmented reality, where data is overlaid on the real world. But many researchers are already looking beyond deep learning. "Neural networks are not the ultimate AI – there will be other solutions, other approaches," says Antoine Blondeau. Some researchers are devising systems that can learn from just a handful of examples, just like humans can. That would make machine learning more widely applicable than today's deep learning systems, which require large amounts of training data.

Others are pursuing entirely different AI techniques. "The field of AI is dominated right now by a particular approach to machine learning," says Gary Marcus, former head of AI at Uber. He thinks long-term progress in AI will require more input from the field of developmental psychology and worries that with its current emphasis on deep learning, the industry risks getting stuck in a rut. Deep learning is just the beginning of AI's deployment – not the end point.

What are some of the things AI could do in future?

Like electricity and the internet before it, AI will power new tools that enable companies to do new things. Many researchers are excited by the prospect of being able to find patterns in huge data sets that are simply incomprehensible to humans. That might include spotting previously unknown connections in troves of legal, scientific or medical documents, or finding patterns associated with particular diseases buried in mountains of genomic data. "Having machine learning in the loop will turbocharge breakthroughs in science and health care," says Demis Hassabis. "It's like having the world's most relentless research assistant that never gets tired."

AI has the potential to transform an enormous range of tasks involving analyzing data, looking for patterns and devising strategies in response. What does it mean for your business if speech can be translated instantly from one language to another? If hours of video can be searched, transcribed and summarized? If vehicles can drive themselves and delivery costs for goods fall dramatically? We are about to find out.

Curious about the contents of our newest Think:Act magazine? Receive your very own copy by signing up now! Subscribe here to receive our Think:Act magazine and the latest news from Roland Berger.