How to use AI for strategic decisions

AI is rapidly entering new markets and is quickly becoming an essential tool for business strategy decisions. Although many leaders struggle to understand how to use AI, there is a simple process flow that can be used for any AI initiative.

digital brain connected to circuits

Artificial Intelligence (AI) is quickly becoming an important buzzword in business today. Just like other technologies like Cloud and Blockchain, it is spoken of often and frequently with little true understanding of uses and applications. However, unlike many other buzzy technologies, AI actually has a huge set of potential applications. It actually is relevant to most industries and functions and it is rapidly permeating the market place. This fact makes it even more problematic that few people really understand how to apply artificial intelligence to strategic business problems.

So, let’s work together to clear a path through the haze of buzzwords and understand what a company has to do in order to develop a differentiated, valuable, AI application. At OpenMatters, we have worked with clients many times to develop AI applications to support better decision-making in key areas like capital allocation, leadership composition, and talent sourcing. We use the flowchart below to explain to our partners what the steps are from ideation to application.

the process of building ai OpenMatters

The process of building AI

4 important teams for AI initiatives

We have found that there are four key groups that must be involved for success in AI endeavors:

  1. First comes the leadership team—the board and executives. Buy-in and support from leadership is essential for success. New AI initiatives require time, capital, and new skills, all of which are difficult to come by without leadership support. Although the leadership team would never be expected to manage the technical details of an AI initiative, the team should be involved in ideation—envisioning where and how AI can help the firm.
  2. Second comes the business implementation team. This group is responsible for carrying out the vision of the leadership team and managing the technical team. AI without a business purpose is obviously a waste of time, and the business implementation team will need to be deeply involved in gathering data, creating insights, and iterating on algorithms. This is a full-time, not a part-time job.
  3. Third we have the AI technical team. This group brings the AI expertise needed to turn data into insight. The AI team should understand the business goals and applications but will be focused on identifying the tools and techniques that will make the AI a reality. This team will be making decisions about the programming environment, supporting tools (like TensorFlow), and techniques (like neural networks).
  4. Finally, we have the infrastructure team. The cornerstone of AI is data, and a big pile of data quickly becomes a big mess. The infrastructure team will manage and store the data so that it is accessible and easy to use. This team will deal with irritating but important problems like identifying a good symbology to join together datasets and managing concurrency for situations when there are multiple use cases and updates happening simultaneously.

Each of these groups is essential to successfully using AI for business decisions. The process won’t work without all of these functions (we have tried it and we know) and each group really does need to be a “team” comprised of multiple skilled individuals. Many of the functions listed above are quite technical, and the tools and techniques are rapidly developing. One individual working without the support and oversight of a peer can quickly fall behind or gloss over errors.

Key steps in an AI design process

Once your team is assembled, the process for creating AI is actually fairly clear and replicable. This doesn’t imply that it is easy, but it does mean that the same process can be applied to myriad business strategy decisions.

The first step is for the leadership team to set the goals. The key questions for this group to answer are (1) what is the unique topic or application that we have data on, (2) what is the hypothesis or perspective we can add, and (3) how will these new insights be useful. Notice that these questions begin with the data. Data is the foundation of artificial intelligence. The leadership team does not need to review the data and understand it intimately, but it does need to have a sense of what data the organization has access to that is differentiated and useful. Often this is data that has been warehoused for years and is used sporadically by human analysts to support business decisions.

For example, a retailer will probably have access to point-of-sale data for many stores, years, and customers.  The leadership team may have a hypothesis that a certain customer groups have different seasonal shopping habits.  Their goal might be to use AI to understand season shopping habits of various customer segments in order to better target ad mailers.

Once the goals are set, the next—and very important—step is to build the data set. Getting the data right is the hardest and most important step in an AI initiative. Often an organization will need to use multiple data sets from different sources internal and external to the organization. The data will need to be combined and cleaned, which is a fiendishly detailed and painful process. Outliers and other bad data must be removed, and datasets matched together using identifiers, such as a CUSIP for a company or an address for a customer. There are many valuable external data sources, but combining the data usefully is always a challenge.  For example, the retailer mentioned above might need to combine their point-of-sale data with a vendor-supplied marketing database that provides demographic information on customers. 

Training data is often an important complement to your data set.  This is where your unique insight or hypothesis comes into play. In order for an AI to “learn” a process, it needs a guide to help it understand what good looks like. For example, our hypothetical retailer will need to define exactly what type of shopping habits it wants to identify.  For example, spending $50 or more in a specific product category might be a useful event indicating the type of pattern the AI should uncover.  Defining or creating training data is a job for the business team because it requires an application of the human insight that you want the AI to replicate.

The data set, and training data are never truly complete—they get added to and updated over time. However, each version should be documented and stored safely by the infrastructure team. Data is a valuable asset and should be treated as such.

With the data in place, it’s time to hand things off to the AI team. The AI team must consider the goals of the business and determine which tools and techniques will best help achieve them.  There are a variety of pre-existing tools on the marketplace today with different target applications.  TensorFlow from Google, Einstein from Salesforce, and Pytorch from Facebook are just a few of the largest.  Another important consideration is which AI methodologies are appropriate for your application.  For example, neural networks are a powerful tool, but one in which the results are difficult to interpret and explain to human users.  For many business applications this black box situation is unacceptable and other techniques, like logistic regression, are more suitable.

The AI team will begin generating algorithms and “grading” them based on how well they are able to replicate the training data.  When the team deems the system sufficiently “smart”, they will bring the results back to the business team to review.  The algorithms generated by an AI will need to extend beyond the training data and applied at scale—and it takes business review to confirm that an algorithm extends well beyond the data it was trained on.  Further, there are often additional considerations that the business team will need to examine carefully.  For example, a specific group of customers that is more important than the rest to identify correctly. 

In many cases the AI will need iteration before it is ready for prime time.  Changes to training data, different AI techniques, and new data sets are all potential paths forward for an underperforming AI.  Here we see another important role for the infrastructure team: documenting each version of the AI.  This documentation should include not just the features used and formulas of the AI, but the overall performance against training data, techniques and products employed, and anything else necessary for a business evaluation of the AI.  It isn’t uncommon to spend a month iterating only to conclude that a previous version was superior, and it’s essential for the team to quickly be able to compare and contrast versions with complete and standardized documentation.

In the happy case that everyone is satisfied with your AI, you can move into the happy space of application—using your new artificial intelligence to help complement and expand the existing human intelligence on your team.  That isn’t to say that development is done, as new data, techniques, and applications are always on the horizon.  But you will have taken the first important step towards embracing AI.

It’s impossible to know all the ways that AI will affect our jobs and the marketplace, and to what degree.  But there is no possible future where AI doesn’t become an essential tool in the executive’s toolbox.  AI generates new insights, quantifies strategy, and extends at zero marginal cost—and in 10 years no one will be operating without it.

This article is published as part of the IDG Contributor Network. Want to Join?

NEW! Download the Spring 2018 digital edition of CIO magazine