I have used the terms digital and data technology interchangeably. And I have described them as things made of bits and bytes. Of ones and zeros. Instead of physical products and services.
And I have spoken quite a bit about how things made of bits and bytes have different economics. They have powerful but also dangerous economics. The scalability and zero distribution costs of digital goods are amazing. The ease of bundling can be powerful. But the competitive barriers can also be very low and most software ends up being free. I went through this in detail in Software and the Sexy but Dangerous Economics of Digital (pt 1 of 3).
But, in truth, I was mostly talking about software – and not about all digital and data technology. In particular, I was not talking about machine learning and AI, which increasingly appear to have different economics than software engineering. The attractive economics may not exist for AI projects and companies. The efficiency and ease of growth and innovation may not exist.
Overall, AI and ML look a lot messier, slower and more difficult.
Here’s how I think about it.
- Software is like doing engineering – so lots of designing and building things within business systems.
- Machine learning and AI are more like doing science – with lots investigating and testing of a much more complicated, real world.
Here’s another analogy.
Why I Don’t Like Hospitals
I did a stint as a turn-around CEO of a hospital in Saudi Arabia. The turn-around part was great fun. It was a lot thinking, diagnosing and then rapidly fixing a lot of problems. It fit my strengths as 90% a thinker and 10% an operator. But after the turn-around, I didn’t much enjoy running the hospital. It was mostly an operational exercise and I discovered I didn’t like the nature of the hospital business.
Because as CEO, I really didn’t have much control. I had to staff up 400-500 people of all types of specialties and skills. Everyone from surgeons to radiology technicians to cafeteria workers and security guards. I really didn’t have any choice as a secondary care general hospital. And I built lots of departments (pediatrics, ER, OR, ICU, etc.). Which I had to fill with expensive machines. Everything from x-ray machines to blood draw stations. Again, I didn’t have much choice in any of this.
Because as a hospital, you have to handle whatever comes in the door. You are there to serve the problems of health and society, which is a complicated reality. If a car crashed nearby, you needed surgeons ready. If someone with psychiatric issues came into the ER at 2am, you had to have psychiatrists and social work ready. If a pregnant woman randomly walked in and gave birth (this happens), you had to be ready. And if she also had heart problems while in labor, you had to have cardiologists and radiologists ready as well. Oh, and if there is a viral pandemic, you had to have ventilators and protective equipment ready for that.
Running a hospital basically means you have to solve whatever problems nature, society, and medicine might throw at you. You are on the receiving end of everything. You have to build an operation that can handle a sea of complexity.
Now, compare that to running a shopping mall.
Why Shopping Malls are Awesome
I also worked on a high end shopping mall that had anchor stores like Saks Fifth Avenue and Marks & Spencer. It had tons of brand stores like Coach and Gucci. There was a food court with everything from pretzels to kebabs. There were a few Starbucks. And there was a 5 star hotel upstairs.
And when the doors opened in the morning, we knew what was coming. It was shoppers coming into an environment we had designed and created. We decided what was to be sold. Which stores. Which food outlets. We choose the lighting. We selected the inventory for some stores. We also decided when we opened and closed. And we could even remove people who weren’t behaving appropriately. It was a service business with lots of real estate, just like a hospital. But in this case, we were solving problems we chose to solve. With a system we created. And we overwhelming chose to to solve a narrow niche of business problems. We did not try to solve problems in human biology. Or in society. Or in nature.
This is similar to the difference between software engineering and AI / machine learning. Like shopping malls, software is designed to solve a specific business purpose. And a handful of use cases. It is made of components that other engineers have made. It can be tested and debugged. Creating software is a lot like developing a shopping mall, but with much better economics.
Machine learning is more like running hospitals. You aren’t designing a system you control. You aren’t focused on a few controlled business problems. You are in the predictions business. You tell it the outputs you want and then feed it training data. It then creates the algorithms that can make hopefully accurate predictions. So the Netflix and TikTok AI try to predict which videos you will most want to see next. And Didi’s AI predicts where a driver should go wait for the highest probability of finding a rider in the next 15 minutes. Note: I spoke about AI as cheap prediction in Products with Personalities? My Interview with JD About Conversational AI (Jeff’s Asia Tech Class – Podcast 31).
But you really don’t control the inputs, anymore than you can control who walks into the hospital. The range of inputs can be vast.
- When Elon Musk trains AI to drive cars, it has to be ready for any crazy thing that it might encounter on the road. You can limit this by putting autonomous vehicles on straight highways or in specially designed business parks. But think of all the inputs a Tesla encounters when driving in downtown London.
- The Google search engine has to predict the right answer to any question anyone might type in. It might be good at common questions like what is the weather today. Or what is the current time or exchange rate. But things get very difficult when people ask obscure questions like who was the most famous nun in the south of France in the 1200AD.
People in AI always talk about the long-tail, which is their greatest benefit and curse. TikTok does well at presenting each individual niche taste with long-tail niche content. It’s why it’s so addictive. Google has maintained its search dominance because nobody can match its quality in long-tail search queries.
But it is also the big problem because the costs of effectively dealing with the long-tail is changing the economics of digital. We are not seeing the economics we say in traditional software. It is like going from the shopping mall to the hospital in terms of operations and costs.
I’ll go into the long tail more in another article. For now, I just want you to think about common use cases versus the long tail. And whether you are using digital / data technology to solve business problems (which humans design) or problems in nature, medicine, science or society. Because they have very different economics.
Ok. Let me tee up a related idea.
Can Prediction Be a Competitive Advantage?
In my SMILE Marathon, I listed machine learning / AI as one of the 5 operating dimensions you can compete on long term. And, as mentioned, the core capability you are building is cheap prediction. You can see how this would be very important in video services (i.e., TikTok) and in making loan decisions (i.e., Ant Financial).
But I also listed rate of learning as one of the SMILE marathons of competition. And I’ve said rate of learning can also be a competitive advantage. This idea goes way back to Henry Ford building the Model T. Every time Ford doubled cumulative production of cars, the costs of production went down 15-20%. Speed of learning and cumulative learning in an organization (and in people) can create real efficiencies and productivity gains. In factories making products, this type of learning advantage shows up as a measurable per unit cost advantage.
It can also show up as a competitive advantage in services. Companies like Goldman Sachs, McKinsey & Co and IBM are capable of learning about rapidly changing environments and topics. They have unique cultures and systems that transmit knowledge throughout the company and its staff. Similarly, Zara, H&M and other “fast fashion” retailers are good at learning rapidly about fashion trends and then changing their inventories during the seasons in response.
Basically, speed of learning and cumulative learning can both rise to the level of measurable competitive advantage in some cases.
So here’s my question.
- If rate of learning can be both a competitive marathon and, sometimes, a competitive advantage, can prediction also be a competitive advantage? Can you get so much better at cheap prediction, that you have a measurable demand or cost advantage?
That’s my question. I think machine learning can definitely be a digital marathon. But I’m not sure about competitive advantage yet. I’ve watching for this.
One last rant first.
Data is Mostly an Input. Mostly Ignore Data Network Effects and Data Advantages.
One idea that has been floating around is the “data competitive advantage”. Or the “data flywheel”. The argument goes like this.
- The more people that watch Netflix, the more data it gathers about what viewers like.
- Netflix then tailors the content to the individual and to the general audience, which improves the service.
- The improved service attracts more customers from competitors, which generates even more data and further increases the service. And so on.
This often called a data network effect. Or a data flywheel. I think this is mostly wrong.
I agree with points 1-2. You can use data to improve services. This can be used, in particular, for personalization, which is becoming widespread. My Amazon and Netflix pages are personalized to me. And that’s great and makes the service better. Good.
But Point 3 is a big assumption. And I think it rarely happens. Just because Netflix is improving my service does not mean that YouTube and Amazon Prime are not. Data-drive improvements in service are rarely located in only one company such that everyone shifts. Most businesses are personalizing and using data to make their services better. It’s just increased table stakes for being in this business.
Data is super important and there are cases when it can an advantage. And there are a few cases where you can get a flywheel (i.e., a feedback loop). But I don’t buy the data network effect idea. And I usually don’t buy the data advantage. I think this is mostly just personalization and using data to improve services. Which everyone is doing.
I think the right way to think about this is not “data”. But instead to use terms like learning and prediction. We can see how a learning advantage (which requires data) might make one company cheaper and give it a real competitive advantage. I think we can also see how a prediction advantage could be created (which also requires data).
That’s my current approach. I find once you ask the right question with the right language, things become clearer.
- An Intro to Discount Rates and Cost of Capital for Digital Valuation (Asia Tech Strategy – Daily Lesson / Update)
- Why DCF Sucks for Digital Valuation. (Asia Tech Strategy – Podcast 101)
- An Intro to Digital Valuation (Asia Tech Strategy – Daily Lesson / Update)
From the Concept Library, concepts for this article are:
- Network Effects
From the Company Library, companies for this article are:
I write, speak and consult about how to win (and not lose) in digital strategy and transformation.
I am the founder of TechMoat Consulting, a boutique consulting firm that helps retailers, brands, and technology companies exploit digital change to grow faster, innovate better and build digital moats. Get in touch here.
I host Tech Strategy, a podcast and subscription newsletter detailing the strategies of the best digital businesses in the US, China and Asia.
My book series Moats and Marathons is one-of-a-kind framework for building and measuring competitive advantages in digital businesses.