When designing an AI system, the most important questions aren’t whether to build on Tensorflow or AWS, the most important questions are about your users.
- who are your users?
- what do they care about? what are their values, hopes and fears?
- what problem can you solve for them?
- how will you do that?
- what does “success” look like?
AI isn’t a fit for everything. One of the most important steps to take early on is to determine whether there is an intersection between your user need and what AI is good at.
One of the most important things to assess is whether the AI can create an unhealthy level of automation bias. Although we are advocates of the idea that human+AI is likely better than human or AI alone, poor design or poor management of the system can create a human that is too trusting in the machine. As AI is used for more and more social decision making, automation bias can potentially have a larger negative effect. Consider upfront how automation bias carries an acceptable level of risk and may degrade the human experience. This could rule out an AI solution, at least for now.
The potential for value destruction is highest at the start, so the time you spend understanding the problem is some of the most important. Spending time talking to people, investigating the data and watching how people currently solve the problem stops you falling into the trap of thinking of AI as a solution looking for a problem. One vital step in this phase is to include many diverse views – both from inside your company and from different user groups.
AI can make solvable problems that were previously impossible to solve, so talking to people who have been around for a while can be a gold mine. As can talking to people in the field; knowledge is created on the frontlines, where problems are readily seen and opportunities more apparent. People who already use automated systems with customers or in the physical world usually have a unique perspective on opportunities.
Mapping work flows is one of the best ways to find opportunities for AI. It will help you understand whether to automate or augment and also how to use a machine-human learning cycle to improve an experience. Once you’ve completed the mapping and you’ve identified what part you want to improve, the trick is to get a handle on what things are enhanced by AI and when not to use it. With an early view on this, you can confidently point to a solution at the intersection of AI and user need.
When AI is probably better and worth it:
- you want to provide personalized recommendations – where different content for different users or at different times and in different contexts is an important part of the user experience.
- users want predictions
- when users want to chat, talk or interact using natural language, or use a bot
- when you can’t program every outcome and need something to classify or identify entire groups or classes
- when you are looking for a “needle in a haystack in a wind storm” (low occurrence events, especially when they change over time) and when there is a lot value in looking for emergent patterns across a lot of data
- when content that changes is more valuable than having a predictable interface
When AI is probably not better and likely not worth it:
- predictability, especially in the interface, is important and this stays the same no matter what the context
- when simplicity and standardization are important
- when you can’t afford to get it wrong – where set rules are vital
- when people need to understand everything about what happens in the system – it needs to be explainable and transparent
- when people say they don’t want AI
Although AI is powerful, it’s also expensive, complex and requires a lot of skill to build, explain, debug and maintain. Early on on the process, use this assessment worksheet with your users and your team to determine whether AI is going to improve your use experience.