An Executive’s Guide to AI and Machine Learning

Intro

There have been a number of companies with significant budget and resources, who have attempted artificial intelligence and machine learning projects, only to have them fail to deliver a return on investment. Prof. Dr. Thomas Hofmann has a unique background in artificial intelligence that spans from academia (ETH) to multinational companies (former director at Google) to startups (CTO of 1plusX, co-founder Recommind). He shares his thoughts on what traits companies have that enable them to get a return on investment for their AI initiatives.

STRATEGY

How important is it for executives to be familiar with AI and how it works?

AI technologies have shown to be widely applicable across many domains and industries. However, this does not mean that they will always work out-of-the-box. One thing that AI systems need for sure is data: the right data and a lot of it. Otherwise it is garbage in, garbage out. Gathering data across the enterprise and possibly enriching them through data alliances and third party data is thus one aspect that is important for executives to understand, because it ties into strategic questions, partnerships and the like. Another aspect that is important to understand is that building AI solutions is not a one-off project, but an ongoing endeavor that requires continuous work. Executives should have enough understanding of the workings of AI to create and maintain healthy teams and practices.

Are there business models that are particularly well-suited for AI companies / projects?

A real AI company is driven by meaningful metrics, which provide guidance all along the development process and operations: from idea sourcing to deployment. In a B2B context, this means that we will see more and more business models that are metrics and performance oriented. This also sets the right incentives for providers of AI services and allows them to tune their systems for maximal performance.

Should companies treat AI projects as experimental research and development, where they must set and accept some degree of risk of failure?

Yes and no. If one reads “experimental” as in: not committing significant resources to the project, then “no”. One can for instance, not easily test AI at a small scale and then scale it up. Without a critical amount of data and computational power, one may underestimate its true potential. On the other hand, it is true that the success of AI projects are often hard to predict. In this sense, one cannot always expect a positive outcome, but may have to experiment with different setups and methods. In general, experimentation for incremental and continuous improvements are a key feature of data-driven methods.

ORGANIZATION

Are there executives who are best-suited for overseeing AI and ML initiatives? CTOs? Chief Innovation officers? Chief Product Officers?

Suitability depends not so much on the designation, but rather on the experience and skills of the person, and at a second level, their ability to exert the necessary influence within the company. It is not easy to make good judgement calls, one needs deep analytic capabilities to get to the bottom of issues. Additionally, leadership skills are essential in convincing others to make — often disruptive — changes to business models, product development, and company culture.

Should all companies have AI departments? Are there reasons why a company should *not* build and instead go the buy route?

For some companies, a separate department is needed, where AI is tied to the future of the business and its deployment spans many products and services. In other cases, it may be more important to directly embed the AI experts with the product owners and developers. It is often a good practice to have the product owners define metrics, which the AI team can optimize on. Hiring the right people (that are not just skilled “on paper”), establishing the right practices and organizational structures is highly non-trivial. Also, the ongoing monetary investments and the profound change in culture needed are often grossly underestimated. Everyone can hack-up something in Spark or Tensorflow (say), but that is merely a tiny step towards a product or service of commercial value. I thus predict that many such attempts will fail and that the right model is to outsource some of the complex AI challenges to specialized companies by buying such services.

What type of engineering / technical profiles do you look to hire?

The largest group of core engineers at 1plusX either works on large scale data processing infrastructure or on machine learning algorithms and analytics. In addition, there is a wide spectrum of more specialized profiles from UI and customer integration to dev-ops and security. In general, we are looking for people that are self-driven and eager to learn and grow. Knowledge and skills of our engineers are a huge factor in increasing productivity, quality and ultimately success.

EXECUTION

How important is it to define specific problems to solve from the get-go?

It depends. In some cases, the use case and value add of AI technology are clear and can be addressed in targeted projects. However, the true potential and value-add may only become apparent as things evolve and constraining projects too strongly from the start may not be adequate. This is also where it helps to work with strong technology partners who have more experience and a larger suite of services that can be explored at low costs.

How do you measure the quality of your AI solutions?

In general, it is always mandatory to establish ground truth. This can happen through panel providers or other high quality sources of user data, say. In other cases it may suffice to use cross-validation and hold-out data. Addressing statistical biases as well as shifts in data distributions (e.g. over time or in different markets) is also crucial.

How important is it to stay up to date with the latest AI research developments?

It is *very* important at the current phase. While AI technologies has shown impressive successes, they are not sufficiently well understood from a theory point of view. This deficit in understanding does not challenge that real world success, yet it may happen that new basic innovations come along, which provide significant benefits that then mandate replacing existing systems. The pace of progress is very high and so is the pressure to stay on top of the game. Here is where a strong connection to a technology partner or academic institution can be invaluable.

About 1plusX

1plusX is a predictive data management platform that uses an innovative, holistic approach in generating anonymized user and page profiles. By connecting information from many touch-points, the 1plusX software detects behavioral patterns and turns them into predictions in real time. These predictions enable digital agencies, media, telecommunications and e-commerce companies to optimize their advertising, sales, marketing and content.

1plusX was founded by CEO Dr. Jürgen Galler, Prof. Dr. Thomas Hofmann (ETH Zurich) and Joachim Schoss (founder of Scout24) in 2014. The team is composed of world-class product and engineering talent from leading internet firms.

Please visit www.1plusx.com for more information.