Machine intelligence (MI) is fast outgrowing its cradle in the technology department. This calls for a new, high-level executive function and a strategic roadmap. In this article I’ll outline some areas for your new CMIO to crack during 2017 and beyond.
At a time when machines are learning how to run brick-and-mortar retail stores, starting to beat doctors in diagnoses, and even inventing their own secret languages, it’s natural for machine intelligence to raise questions and issues. Having had the privilege to work on and oversee many MI* projects already, and diving deeper into the topic with each passing day, I see the future bringing more boon than doom for companies and individuals alike, as long the benefits and challenges are tackled with seriousness.
One of the strongest predictors of a society’s well-being is its productivity, and machine intelligence is bound to be a productivity beast: IBM Watson handles in seconds or minutes what would take a doctor, a lawyer, or their back office teams weeks or months of work; Amazon’s Go store uses MI to enable a lightning-fast retail experience without checkouts; Google’s MI already translates from and to languages it hasn’t encountered before; and Salesforce’s enterprise MI can dramatically cut administration, reporting, and coordination efforts, which now suck up over 50 percent of management’s time. All this will give people more time to think, manage other people, use judgment, and simply work with things that bring them joy.
Undoubtedly machine intelligence, like all technological disruptions, will bring also challenges, and will likely make the internet revolution look tiny in comparison. Still, while it’s widely seen as key to future success, details and comprehensive roadmaps are often missing, as is the sense of where MI should even live within the organization. In my point of view, it should be key to any large company’s strategy, and it should be led by a senior leader who reports directly to the CEO. These chief machine intelligence officers, or CMIOs, will be tasked to bring all business departments, as well as the company as a whole, to the machine intelligence age. If you are one of these people, or looking to hire one, below you’ll find some key CMIO-level challenges to tackle.
Where are the hidden connections awaiting discovery? What are the key things you should be able to diagnose and predict in real time? Which routine but massive tasks should you automate? Where should you use fast, contextual customization and real-time optimization at scale? These are some of the most promising areas of MI’s value in business use. The CMIO will need to have a thorough understanding of the company’s environment, its business processes, and their relationships to assess the biggest opportunities as well as the most acute urgencies.
In some areas, MI won’t even be bringing competitive advantage — it’ll be the price of doing business. The adoption rate of new technologies has accelerated dramatically throughout history, and we now have the mobile internet, accessible cloud services, and much of the Internet of Things already in place. MI can in this setting take over entire industries faster than preceding technologies did, and will lead a deep transformation of the organization’s operating models. How will, for example, Google make money when we don’t browse through search results anymore, but simply pose a question and expect one right answer, as well as instant execution? Laggards will face hardships and even extinction, as did stubborn incumbents facing internet-born companies. Only this time, the effect is bound to be even more dramatic.
Alongside machine intelligence, design thinking and customer experience management will be critical concepts for business success in the near future. People are not rational calculators, and they often value soft experience factors even more than hard, objective value. MI holds vast potential in providing more valuable experiences to people, and the CMIO needs to work with design departments and agencies to unearth these opportunities.
MI can streamline the everyday, remove tasks to get jobs done, make the previously impossible possible, empower and educate people, tailor and personalize products and services, enable important connections with people and things, make experiences safe and secure, and enable completely new forms of storytelling and entertainment. These are just some areas for innovating experiences, but across all of them, seamless transitions between platforms and devices will be key.
And MI will not touch just the what. My friend Alexander Manu’s upcoming book will revolve around the insight that in the behavior economy, the delivery (the how) is indeed the whole value proposition (the what). Chatbots and cognitive personal assistants are already making many common tasks smoother and more automatic, and learning about their masters as they go.
Mind also that you have both internal and external customers. Externally facing applications are usually instantly viable where problems are simple and ample data available. Where problems are murkier and data exploratory or scarce, it’s still often better to build decision and support systems for your staff. Their experience in dealing with customers and customer data can thus be greatly enhanced.
Autonomy and interaction
How much autonomy to give the MI solutions working with and for your customers, staff, and other stakeholders is a key question. Ideally, intelligent systems should take care of mundane tasks all by themselves, but even where they already can, people might not yet trust them. The fact that Google’s AIs Bob, Alice, and Eve figured out how to speak to each other in a language humans can’t decipher made some people worried. And as already happened years ago, haywire trading bots have caused flash crashes and millions of damage.
In his book Superintelligence, Nick Bostrom divides machine intelligences into three levels: Oracles tell their users what they need to know, like Google Search has done for a couple of decades now; Genies can carry out their masters’ wishes, but only upon request, the way Siri now works; and Sovereigns require only a few key principles to operate independently on behalf of their hosts. While true sovereigns don’t yet exist, robot portfolios by brokerages can be considered their present-day relatives.
Whichever level of independence you set on, CMIOs also need to lead the development and monitor the performance of the company’s MI interaction practices. MI systems increasingly interact with users via text and voice, sometimes also called the “interface of least resistance.” Gartner predicts that within four years, 30 percent of searches will be done without a screen; first via voice, then maybe via brainwaves.
While conversation is the most human of interactions, it’s not always the most natural medium for brands. Instead of logos, tag lines, and other common brand assets, the brand’s essence will distill down to questions, answers, micro-interactions, sentiment, and style of conversation, even pauses. These will become important business and brand factors, and metrics and analytics will need to evolve to reflect these nuances.
A key task for the CMIO, together with the CTO, COO and/or CFO, is to figure out the technical architecture of machine intelligence, along with a roadmap and investment plan. The MI ecosystem is already a flowering and fast-growing garden with a mix of cloud services and on-site solutions, as well as hardware and sensors available for companies large and small, for core and support functions alike, and customized for different industries and sectors. The technology stack also includes several levels vertically, from understanding natural language and abstract concepts to data collection, data sciences, and machine learning, as well as general-purpose libraries and MI training grounds. There are already many vendors of ready-to-use capabilities and APIs, and also lots of free or cheap resources to build your own solutions from scratch.
While open source, free-to-use libraries like Google’s TensorFlow might sound great at the outset, think before embarking on a quest to build MI in-house. Even though the basic tools might be free, you’ll need a cadre of experienced scientists and specialists to develop the self-improving machine intelligence models and algorithms. This is not easy: Capable people are hard to find, not to mention expensive. In StackOverflow’s survey, only 0.1 percent of respondents recognized themselves as machine learning developers, and 1.9 percent had a background in math or statistics, key skills for MI development. Even with easy-to-use interfaces and options to train MIs by talking to them, you’ll still need the experts when something needs fixing or tweaking.
So, rather than trying to best Google, Salesforce, Amazon, Microsoft, IBM, and Facebook in a global talent hunt, partnering with these companies is a viable option. It will also bring in the added benefit of combining the big players’ MI stacks with the tools you already use daily. Especially for niche, specialized purposes, there’s an ever-growing selection of small but potentially very suitable providers. But remember also that in machine intelligence, scale is usually a benefit, since bigger data means faster and better learning opportunities.
An MI system needs to learn constantly, because it’s just a baby at its release. In fact, researchers are now creating virtual babies to help them understand how babies learn and teach computers to learn faster. Charting out learning goals and strategies for the company’s MI initiatives is a critical task for the CMIO. Constant streams of data play a critical part in learning, and a rule of thumb for data for MI is “the more the better.” A key strength of MI is that it doesn’t need to take samples — it can learn directly from all available data, making insights and predictions more solid. While it’s great to have a theory (or even just a guess) about how the data and the problem might link, remember that MI can help make novel connections, find new opportunities, and answer the questions you haven’t even thought to ask, so long as there’s troves of data available.
Different problems will require different learning paradigms, with reinforcement, supervised learning, and unsupervised learning being a few of the most well-known methods of machine learning today. For training purposes there’s already an ecosystem: Elon Musk’s OpenAI just opened Universe, which provides tools for everyone to train, measure, and evaluate intelligent systems; OpenAI Gym is used to train reinforcement algorithms; and Microsoft has open-sourced its Microsoft Cognitive Toolkit to speed up development of MI applications and is gaining traction. If even the most secretive tech company, Apple, is opening up its MI research to the wider academic community, it should be clear that the future of MI leans on collaboration and collective learning.
A good way of thinking about training your intelligent system is making it a game, with triggers, actions, failures, and successes, as well as reward and feedback loops, that can be easily understood by people and machines alike. The research company L2 aptly sums the importance of learning by outlining the formula for value in the future as: amount of data sensors available × actionable intelligence that can be fed back to the system = value proposition.
Last but not least, the CMIO will need to tackle the human component of intelligent systems, as MIs don’t yet manage on their own. People are required simply to behave in ways that leave a data trail that the MI can measure, learn from, and apply to new situations. But humans are also needed to intermittently check data, tweak algorithms, provide training sets, and coach the MIs on their correct and incorrect judgments, as well as manage situations that the MI isn’t able to independently resolve.
Reinforcement learning and supervised learning require more of the human component, while unsupervised learning lets the machine learn on its own. That last one, however, is still the most experimental of the three, and even unsupervised systems don’t (yet) build themselves. Google AI Experiments is an engaging example of how everyone on the internet can be harnessed to play with intelligent systems, and train them in an open learning environment.
A thing to keep in mind is that MI systems are still built by people, so they can perpetuate their creators’ biases if this is not properly controlled for. This does not just concern the scientists creating the algorithms, but also the annotators labeling the training sets. As a practical example, when measuring on a scale of 1 to 5, women rate the same puppies 0.16 stars cuter on average than men do. This is a statistically significant divergence, and it matters if you’re teaching your MI about the concept of cuteness, for example.
Even with a seemingly perfect system, the CMIO still needs to decide how much to trust the machines, how much control to retain, and how to establish checks and balances with people. As already mentioned, one of MI’s brightest promises is that it will allow your human components to focus more time on human work, such as judgment, socializing, and networking, as well as creativity. Over time, the need for control should diminish.
At Havas, we currently have many marketing and product initiatives utilizing MI underway, worldwide and across industries. Perhaps the one closest to my heart is still the “Most Confident Fan” initiative, which we created with our clients TD Ameritrade and IBM Watson, to provide a new take on TD Ameritrade’s NFL sponsorship. For the football season, we created an engaging web application that measured how the fans’ confidence about their team as expressed in social channels correlated with the team’s success. The first-of-its-kind solution fared surprisingly well. For example, it predicted every game of the NFL champions Denver Broncos correctly, including their winning the Super Bowl, in which many considered the Broncos the underdog. But no, unfortunately I didn’t bet on the Broncos.
Machine intelligence will take over many sectors and verticals in the coming years. And the question is not if, or even when, but instead how are you preparing for the machine intelligence era? Considering machine intelligence a key strategic office in the company is a great first step.
This article appeared originally on LinkedIn.