Barriers to AI value creation have changed a lot in the past two decades. Twenty years ago, they were computation and storage but cloud computing made those practically free. Then clean, labeled data was the challenge so we spent years developing data warehouses, Hadoop data lakes, ETL, ELT, data cleaning, and data harmonization.
Next, human technical skill was the bottleneck. Everyone tried to hire data science unicorns who knew how to train AI and machine learning (ML) models, run statistically valid experiments, develop MLOps, and understood business goals. We soon learned that unicorns don’t exist, so we built unicorn teams instead. Today, the biggest barrier I see when working with our customers is human skills again, but this time it’s non-technical — it’s vision and imagination.
The biggest blocker to AI ROI today is vision and imagination.
Organizations aim too low. Too often, AI development teams prioritize point solution use cases that have little value and less frontline user support, use cases that cost $1x to develop, $5x deploy and adopt, and generate $0.5x in value yielding a whopping -92% ROI. (See examples from medicine and insurance.)
3 Levels of Technology Value Creation
Disruptive technology use cases have fallen into three stages the past 300 years. In ascending order by value, they are:
- Point solutions, such as demand forecasting, churn prediction, and predictive maintenance
- Apps that redesign processes, and
- Systems that introduce new sets of processes
A good example of the second stage is John Lee’s work at NBCUniversal. Objectivity caveat: I’m a fanboy. John was my boss’s boss a few years ago when I helped Merkle with cloud and AutoML migration. (We moved from one big, shared SAS server to many AWS accounts and elastic Spark clusters. It was like dropping a 17th century Yorkshire dairy farmer into New York City’s Times Square today. What a trip!)
John has vision. He’s not applying AI to dozens of tiny point solutions but rather is using it to redesign major processes, processes that took floors of people to execute a generation ago. Consider this chart of advertising analytics disciplines by video media types:
Briefly, first-party data is data that media companies, ad networks, or advertisers collect directly such as NBCU knowing that I watch Yellowstone on Peacock or Nike knowing that I buy green polo shirts. Viewer identity management is stitching together all the IDs and devices I use to access NBCU and Nike. Media planning is deciding where Nike should show ads, such as on Yellowstone streaming, and activation is delivering the ads. Media metrics that have been used since "Mad Men" days include how many people see an ad (reach) and how often they see it (frequency). Reach and frequency are but means to an end and the end is Nike’s business goals, which might be brand awareness, brand favorable impressions, website visits, website sales, website revenue, or incremental sales at Dick’s Sporting Goods.
There are 18 cells in this table. Two generations ago, each column was its own company and, a generation ago, its own floor in a big agency. Today, they’re still largely disparate, siloed teams with their own data, tools, and jargon. That kind of specialization was needed because of the complexity of the columns. John is leveraging big data and AI to drastically reduce their complexity and collapsing 18 silos to one integrated process. The Henry Ford of TV ads, if you will.
This kind of AI-driven collapsing of processes is happening across the digital landscape. Epsilon is another martech customer of ours. They developed a real-time, omnichannel hyper-personalization platform covering connected TV, streaming, web display ads, audio, SMS, email, web traffic, and more. Unilever (also a Dataiku user) applied Epsilon’s platform to website personalization and doubled the number of identified customers.
What to Do?
Take a look at your AI product backlog. Are they all point solutions? For the most valuable ones, what would an app that redesigns a process look like? For the apps, what would a new AI-driven system be like? We’ve helped hundreds of companies with this through data-thons, hackathons, use case ideation workshops, and use case scoping workshops. Please reach out to learn more.