By: Jan Scholz and Aneesh Datta
Introduction
Going by how companies market themselves and what the media reports, it seems Artificial Intelligence is everywhere these days. However, despite the hype, there is very little actual use of AI in production today. Sure, there are companies such as Google and Amazon that use AI in their products to assist you in writing your email replies or recommending products you might be interested in. However, most people still go through their entire day without getting help from or delegating work to an AI. Why is that? Where are all the AIs we keeping hearing about?
The short answer? They haven’t been built. In this post, we’ll describe the obstacles to creating AI-powered products and how to build them the right way.
How to identify opportunities
Ironically, the industry is struggling to find applications for AI in spite the extraordinary advances in the field. Take for example Open AI’s impressive GPT-2 language model. Based on a few sentences it can write an entire essay incorporating the topic, theme, and preferred style. While this is truly mind-boggling, what are the real applications of it? If the AI was good enough, it has the capability to replace authors and even script or news writers. But it’s not quite there yet. So how can we leverage its current powers?
The example of GPT-2 illustrates how the industry has it backwards. We are blinded by the tremendous advances in this field and are trying desperately to find problems that can be solved by AI. In reality, we should be thinking the other way around: taking a human-centered approach that starts with understanding user context and needs. With this approach, AI is another tool at our disposal — a very powerful tool that allows us to imagine further than we ever thought possible.
The Double Diamond framework
We approach product development in a way that puts the observation and discovery of the user needs at the start of the problem-solving process. This is part of our end to end process captured using something called, the ‘double diamond framework’.
The goal of the first diamond is to ensure that we are solving the right problem. We do this by uncovering the core challenges faced by users and translating these insights into ideas. Once these are understood, and ideas have been defined, we move to the second diamond. Here the goal is to solve the problem the right way. We do this by designing and building the experience by implementing a build-measure-learn process that allows hard data to guide our priorities and validate our hypothesis.
In the table below, the first three columns highlight the activities and outcomes. The last column describes how to think about AI along the stages of the double diamond framework.
Although we shouldn’t conclude if ‘AI’ should be part of the solution in the first diamond (steps 1 and 2), being aware of the tremendous capabilities that this toolset provides can lead to better ideas. For example, we can consider AI whenever we come across users who are frustrated by having to engage in tedious or repetitive tasks. AI is well-suited for automating these kinds of tasks.
Note the sequence of our thinking process: we started with the user and their frustration, devised ideas to solve the problem, and then considered AI as a tool to engineer the solution.
AI Product Execution:
Once you’ve defined how to solve the problem and how to address the needs of the user, the next step is to build the solution. Here are a few guiding principles to consider:
1. Start Small: Structure a solution that increases your odds of success. This could mean, at the onset, having the discipline to not take on big problems that require complex AI solutions. Instead, a more effective strategy is to prioritize solutions to problems that are less complex because this increases the odds of success. This approach also provides the added benefit of building AI momentum within the organization.
2. Augmentation versus Automation: The term AI is synonymous with automation because AI allows users to delegate repetitive or undesirable tasks that were previously done manually to computers. However, there are times when users prefer to work with AI instead of fully automating the task. Here are two reasons: :
- The implications of a false positive or false negative are high: For example, use cases where an error may compromise human safety or increase financial risk, the benefit of fully automating a task may not outweigh the cost.
- Augmentation can be an effective way to incrementally work towards automation. The value of partially automating a solution can reduce the time and bolster the ability for the developer and the model to learn faster.
Data is a Distraction
There comes a time during the design of an AI solution where the fear of having too little data is raised. At this point, the wheels grind to a halt and the whole project can get derailed by shifting the focus away from the user problem. Instead, efforts are shifted to mindless exercises of accumulating huge amounts of data, creating data infrastructure, and data quality and cleaning exercises.
This is not to say that data isn’t important, but it shouldn’t come at the cost of compromising the user experience. Data, model, and user experience are interdependent and cannot be solved independently. Spending time on data alone risks acquiring the wrong data and creating ineffective infrastructure. Building models and exposing them to the user can help to inform the data acquisition process. Many of the questions around volume, state, and predictiveness of the data can only be answered by building the machine learning model. The question of whether the model’s insights are needed, useful, and actionable can only be answered once the targeted user is exposed to them.
Finally, AI applications should be designed to collect their own data to enable scaling. Autonomous vehicles are a good example of this. The “autopilot” is essentially an application on four wheels driving around gathering data. Each time the driver intervenes and corrects the AI, it learns valuable feedback. This improves the autopilot’s predictions and thus increases the number of times humans choose to engage this feature. As a result, more autonomous driving data is gathered. Instead of aiming to collect data upfront, AI application designers should build this virtuous circle.
Conclusion
What we’ve laid out here might seem eerily familiar for everyone with a background in product development. It seems that the software industry is yet again going through another cycle of “reinventing-the-wheel” with regards to the development of AI solutions. Taking a human-centered approach by starting with user needs and then thinking about how AI can address those needs is a powerful recipe for innovating and building experiences users love.
