10 tips for successful adoption of your machine learning products
Artificial intelligence and internet of things expert Rudradeb Mitra explains how companies can meet the needs of customers with machine learning.
Let me start with a conjecture: the biggest difficulty for products based on machine learning (ML) will be user or customer adoption.
How did I come to this conclusion? While having a conversation with a top executive of one of the biggest European insurance companies, he told me: “We have the money and technical talent to build sophisticated ML-products, but we do not know how to make users adopt those products. We spent a millions of dollars on an ML-based app, but only got around 300 users. We do not understand why people do not want to use our app.”
Similar views were echoed by one of the top executives of an Indian software house with multi-billion dollar revenues, who told me: “We are having problems with getting our internet of things (IoT)/ML products adopted by users and customers.”
Why are these companies not successfully meeting their goals in this space? Over the past few years, ML/artificial intelligence (AI) has received plenty of hype. Figure 1 shows the Gartner hype cycle.
Many startups built products and were able to launch to early the adopter market, as early adopters often adopt hype products to try out. Now though, ML/AI is moving from the hype phase to the correction phase. In this phase, the market moves from early adopter to majority. The majority adopt the technology if they see a clear tangible value. This is also the group where customers and users of corporations belong.
Figure 2 shows the Technology adoption life-cycle and the innovation hype cycle.
So how can ML products be adopted by the majority? Typically, corporations think top down. They build products and expect their customers to use them. Unfortunately, disruptive technology adoption cannot work that way. And today, with heightened competition and tech giants like Google and Amazon entering every kind of business segment they can, corporations need to rethink how they build and launch disruptive products.
In this article, I will share tips on how corporations can increase chances of successful adoption of their ML-based products.
1. Companies must provide incentives for users to share their data. In the past, many corporations used fine prints in the user agreements to collect user data. The new General Data Protection Regulation law will not let corporations do that anymore, at least not in Europe. You need to create incentives for people to share their data by providing them with a better experience or a way to save their money.
For example, when I was part of the team that built the Road Skippers, we gamified the system with incentives like free movie tickets and coffee. This meant people were very happy to share their data. This was in contrast to the approach taken by many insurance companies, who asked users to put black boxes in their cars. That did not work. Figure 3 shows some of the screens from the app.
2. You cannot force users to adopt a new disruptive technology – you must also look to give back value to the users for their data. Once you collect data, you need to show the value of the data you have collected back to the users.
In 2010, I was working on a project for CEZ around smart meter adoption.
We were trying to answer a key question: how to make users adopt smart meters? We studied various smart metering projects around the world and spoke with different stakeholders.
Our findings pointed us clearly towards one thing – provide an interface showing feedback. Figure 4 - our recommendation to CEZ (University of Cambridge Final group project).
3. When it comes to product design, it’s important to let the users be in control. The machine should be seen as an aid to help humans and not to replace them. As human beings, we want to feel that we are always in control. I do not know what will be in place in ten years time, but in the current phase of AI products, users want to feel that they are in control. Therefore, build tools that help users to make decisions or make things easier for them.
As an example, I am working with an AI bot startup (Figure 5) and we are trying to make sure that the bot asks questions and helps the user to make decisions. In this way, the user always feels they are the decision maker.
4. Do not try to change behaviors. If you are building a tool that requires users to fundamentally change their behavior to use it, that’s no good.
I have learned this the hard way. I was building a startup that needed people to change the way they do certain things, specifically around travel. But it did not work! That’s because people want to do things the way they are used to doing them. Changing behavior is a long and difficult process, which requires persistence.
Corporations do not have that kind of mentality, but startups do. Do not go into the business of changing behavior – leave that for startups!
5. With product engineering, it’s vital that you do not overcomplicate the architecture. So many times have I seen corporations and startups overcomplicate their product architecture. There is a tendency to choose the ‘fancy’ technology, even if a simpler technology might do the job.
A few weeks ago, I was talking with an IoT company in the UK that is building an ML-based prediction engine. They were thinking of using a data warehouse. A data warehouse is system for sophisticated data analytics and business intelligence. But in most cases, tools like Mixplanel or Tableau can be enough to meet a company’s data analytics needs. I could see that they did not need a data warehouse, but the person in charge was eager to use his superior skills in the data warehouse space, so he was pushing for that.
Below (Figure 6) is a simplified version of the architecture I devised for an IoT predictive analytics company:
Complicated architecture makes the product development and iteration cycles way longer and will affect the adoption cycle, as often during the adoption phase, you need quicker turnaround.
6. Also, be sure to choose the right database or databases. The database has a big effect on every aspect of your product, including the user experience. ML-based products use ‘big data’ and the natural tendency is to select a No-SQL database. However, there are problems with No-SQL databases. Most of our non-machine generated data, as well as legacy data, is relational in nature and best suited for a SQL database.
When we were building Road Skippers, we ended up using two databases – one SQL and another No SQL. Figure 7 illustrates the two databases.
A non-efficient database structure can make query time long and thus create a bad user experience. Because of ML-products (and IoT), there has been a sudden increase in time-series databases (Figure 8). A time-series database is a database optimally suited to store data that is generated over a period of time. Many ML-products use such time-series data. Last year, I was co-operating with a SQL-based time-series database, and I found this solved the problem of multiple databases in a very neat way.
7. If possible, try to use the simplest best-fit ML-algorithm. Machine learning models can be quite complicated. The training can take a lot of time and if the data is not clean or good, efforts will go waste. It is better to start with the simplest best-fit model. In most cases, I have seen a simpler Neural network like word2vec often does the job quite well. Or Neural networks doing classification or clustering can be good enough to solve many problems. There are a few cases I have come across where more sophisticated networks like long short-term memory (LTSM) will be needed. Figure 9 shows a LSTM cell. One of the cases for LSTM is here.
Below (Figure 10) is a simple classification neural network which can be used to solve many real-world problems:
8. There are some important things to remember when it comes to your product launch too, and the first is to educate your customer or user. Recently I started reading a book named named The Challenger Sale. The authors studied over 6,000 salespeople around the world and classified them into five categories. They saw that the salespeople they referred as ‘challenger salespeople’ outperformed every other group. Who are these challenger salespeople? These are the people who challenge the norm, are more knowledgeable and educate their customers. The customers trust them and ultimate buy from them.
9. Also look to ensure consistent metrics across marketing, sales and products. I have seen companies using a different metric for user acquisition channels, marketing and sales. Sometimes those metrics can conflict and your product development team may be optimizing a metric which does not reflect an increase in adoption. An article on this topic with more detailed explanation can be found here.
10. Be sure to identify enthusiasts as well. Within your majority users, some of them will be the early majority. Identify them well in advance, communicate with them, and create a Facebook group to engage with them if possible.
Last year, during a panel discussion one of the bank executives claimed: “We are not worried about AI. We will be able to easily adapt.” My reply was: “That is the exact reason why big corporations will be disrupted by tech giants.”
Perhaps, then, the most important lesson to learn in this space is to not be complacent or overconfident.
Rudradeb Mitra will speak at the CCX Forum: Channels and Customer Experience. Save your place at the event and join him and many other top level speakers!