AI

By Robin Bordoli, August 2, 2016

CEO to CIO: “What’s our AI strategy?” – Part 3

This post is the third in a three part series.  In the first post 2 weeks ago, we gave some insight and context into why your CEO is asking this question, why now, and why you the CIO.  In the second post last week, we gave you a foundational framework to think about AI so you could give your CEO a thoughtful response.  This week we will discuss how you can engage the business on the topic of AI and consider important criteria when evaluating AI vendors.

Engaging the Business

So now you have the AI = TD + ML + HITL conceptual model to apply to your business, it’s time to start engaging with your executive counterparts in Marketing, Product, Sales and Customer Support.  

Focus on Outcomes not Technologies

Avoid the trap of taking the conversation with your executive counterparts in Marketing, Product, Sales and Customer Support too quickly into techno speak.  Focus the initial conversation on the outcomes that matter to them.   A good way to focus the conversation is to start with 2 questions:

  • What are you already doing that you want to do faster/cheaper/better?
  • What are you not currently doing that you want to start doing?

For example, you may have the SVP of Customer Support say that s/he is not happy with the quality and consistency of how they classify the severity level of support tickets since it’s the individual customer support reps doing the classifying and the incentives are misaligned.  Or you might have the VP Product say they need to start collecting relevance data for search queries and results to understand how well they are meeting customer demand.  Or you might have the CMO say they need to start collecting sentiment data from Twitter to establish a baseline for how customers view their brand and products so they are not flying blind when they embark on expensive campaigns.

Explaining AI

Once you’ve established the outcomes that matter to your executive counterparts, you can then introduce the concept of AI.  Don’t lead with it.  The goal is not AI for the sake of AI, but AI that helps the business meet its business objectives. With the clarity of the business outcomes you want to help deliver against, you can now introduce the conceptual framework with the AI = TD + ML + HITL equation.

Step through the equation and translate what each concept mean in terms they understand to them.  For example, if the CMO wants to start doing sentiment analysis at scale then walk through the following:

  • Training Data.  Download the tweets from Twitter based on the #hashtags and @mentions that matter.  Define the questions and sentiment scale you want to use.  Find humans to apply the classification.
  • Machine Learning. Create a model that can predict sentiment based on the Training Data.  Apply the predictive model to new tweets.  Model assigns a confidence level to each individual level sentiment prediction.
  • Human-in-the-loop.  Use humans to override the model predictions with low confidence.  Add the new human labeled data to continue to train and improve the model.

Systems of Record with Unstructured Data

Once your executive counterparts in Marketing, Product, Sales and Customer Support have a solid foundational understanding of how AI could apply to their business priorities, then you can move on to taking stock of the nature of the data in the mission-critical customer and product systems of record.  Specifically, you want to understand what are the sources of unstructured data (text, images, audio, video) that if structured could be the initial training data.

If you’re like the typical enterprise then probably your unstructured data volume is 4X your structured data volume

A graph showing that you need more structured data as part of your AI strategy.
Examples of unstructured data could include:

  • Tweets from customers discussing your products or brand
  • Customer attributes written into a sales opportunity in Salesforce Sales Cloud or Microsoft Dynamics CRM
  • Product descriptions and images in your product catalogue system of record
  • Relevance labels in your search algorithm
  • Email conversations between customers and support reps in support tickets in your Salesforce Sales Cloud or Zendesk systems

Once you understood how and where your business is collecting unstructured data, and the path to turn that into structured training data, then machine learning is a possibility.  

Evaluating your Vendors

The third and final part of how to respond to your CEO’s question “What is our AI strategy?” after formulating a framework and engaging the business, is evaluating your vendors.  Managing vendor risk has been part of the CIO’s mandate since the role emerged.  Developing this competence has been more important as the number of vendors exploded with the emergence of the cloud wave.  

Rather than rehash all the typical considerations when evaluating vendors, let’s call out three factors worthy of extra focus specifically for vendors claiming to deliver AI solutions.

Single Domain vs General Purpose

Some vendors such as Wise.io have chosen a single domain (customer support) while others such as CrowdFlower and Sentient have chosen to build a general purpose platform that can fulfill many machine learning use cases for text, images, audio, and video data.

A screenshot showing all of the options to create CrowdFlower jobs as part of your AI strategy.

While one approach is not inherently better than the other, there is a long-term trend and desire for CIOs to rationalize the number of vendors they manage so there is a slight bias towards a general purpose platform, if (and it’s a big if) that general purpose platform can meet the business needs of the different functions.

Black Box vs White Box

With the emergence of 180,000 data scientists in North America over the past 5 years, the organizational model of where they fit is still developing.  Some more mature organizations have a Chief Data Officer with a centralized data science function, but the more common model seems to be a decentralized model with data scientists sprinkled across an organization inside different functions.

In the decentralized scenario there will inevitably be a broad spectrum in the skill set of the data scientists within a single company.  Some will have advanced machine learning backgrounds, while some will have progressed to the data science role from being a software engineer or a data analyst.  Those with the advance machine learning chops probably won’t be satisfied with a black box approach where it’s not transparent how an algorithm works and they can’t fine tune the parameter weightings.  Bear this in mind when considering solutions that are black box only.

Point Solution vs Integrated Platform

Many vendors claim to be AI solutions, but as you now know they are conflating AI with Machine Learning.  Commercially viable AI needs the 3 critical components of Training Data, Machine Learning and Human-in-the-loop to be integrated together.  So you have a choice.  Either select 2 or 3 different vendors for Training Data, Machine Learning and Human-in-the-loop and integrate them yourself.  Even with well defined RESTful APIs there is still a cost to do that.  Or look for a solution that has Training Data, Machine Learning and Human-in-the-loop integrated into a single platform.

A formula for AI strategy is Training Data + Machine Learning + Human in the Loop (AI = TD + ML + HITL).

Now that you’re armed with an approach for how to engage your business on the topic of AI, why wait for your CEO to ask “What’s our AI strategy”.  Why not be proactive and tell your CEO “Here’s our AI strategy”.  The days of CIO meaning “career is over” are long gone.  It’s time for the best CIOs to seize the initiative and ride the AI wave.