A guide on how to assess whether an AI-based system is appropriate for your use case.

The first step to building a practical AI-based system is to define the problem that you want to solve in a clear way and make sure that it is a problem that is worth solving with AI. This is a critical step, as it will determine all the design decisions that you make throughout the rest of the system building process.

If you have any questions, comments, or would like Inspired Cognition to help in deciding the best use cases for AI in your organization, please feel free to get in touch!

Three Scenarios for AI-based Systems

First, let’s think of three broad scenarios in which AI-based systems can be used, categorized based on how much human effort is involved in solving them.

  • AI Substitutes: In this scenario, AI systems serve as a substitute for human effort, doing the same task human workers are already doing. This application of AI is appropriate when some degree of error is acceptable, so it is not life-threatening or extremely reputation-damaging when the AI system makes mistakes.
  • AI Assistants: In this scenario, AI systems serve as an assistant to a human worker, and will aid the worker in their job. These situations are most appropriate in the case where the system outputs must be very high-quality, or in situations where even humans struggle with the difficulty of the task and could benefit from assistance.
  • AI-enable Applications: In some cases, there are tasks that would simply be impossible for humans due to limits on human abilities to process large amounts of information quickly, or extreme danger involved in performing these tasks.

The table below gives an example of some real-life applications that fall in each category over several application areas of AI.

  Language Vision Robotics
Substitute Grammarly grammar checking uses AI to detect grammar errors (instead of a human proofreader). DALL-E 2 image-to-text generator creates illustrations based on natural language commands (instead of a human illustrator). Aethon robots deliver food or medicine to hospital rooms (instead of a human nurse).
Assistant Jasper copywriting assistants help copywriters write faster by providing them with AI-predicted suggestions. aidoc medical image analysis provides doctors additional information regarding the images that they are analyzing. Waymo driving systems assist drivers by keeping them in the same lane and maintaining appropriate speed on the highway.
Enabled App. Siri virtual assistants can answer factual questions about a wide variety of topics nearly instantaneously. Google image search finds similar images from a large database, for purposes such as detecting plagiarism. HEBI robots autonomously enter areas that are too small or dangerous for humans to enter.

What is the Value Proposition?

When developing an AI-based system, you will need to next need to decide the value that an automated system would provide.

There are rare cases where an AI-based product provides direct value to a user or to a company deploying the AI system. For instance:

  1. An AI-based system that automatically finds applicable tax rebates could be evaluated directly on the monetary value of the rebates it discovers.
  2. An AI-based system that automatically buys and sells stocks could be evaluated directly on the investment gains that it provides.

However, in most cases the system provides indirect value, by making an existing process faster or better, or opening up new possibilities.

In the case of an AI Substitute or AI Assistant, where a human worker is already performing the task at hand, there are basically three things to think about in determining the value of the AI-based solution:

  1. How much value do human workers currently provide by doing the task? For example, how much are they paid to do a certain amount of work, or at least how much time does it take to do a good job?
  2. How much better or worse will the outcome of an AI-based solution be, and how would that affect the perceived value? For instance, if an AI-assisted writing system helps a copywriter write more fluently and lucidly, how much more would their services be worth?
  3. How much more efficient will the work be when using an AI-based system? Similarly, when using an AI-assisted copywriter, could the writer produce 50% more articles of similar quality?

Answering these questions is an extremely important step in deciding which problems to tackle and how. But it also requires an understanding of market research beyond the scope of this AI guide, but you can find a more detailed example in the tutorial below:

Is The Use of Automation/AI Valuable?

Once you have an idea of the value provided by the system you are trying to create, it is a good idea to think about whether it would be worth creating a system for automation at all. There are several situations where the answer to this question is “no!”

In some cases, there is little added value in solving the problem. For example, building an AI substitute robot that plays golf instead of human players may be an interesting academic endeavor, there aren’t many people in the world who are thinking “wow, I really hate playing golf and wish there was a robot that did it for me!”. This may be an extreme example, but there are other examples of tasks that are quick, enjoyable, low-paid, or sufficiently niche that there wouldn’t be a huge amount of value in providing a substitute or assistant. In particular, you need to consider this carefully when building an AI-enabled application where humans don’t do the task currently. It’s possible that no-one has created such an app before because it was too challenging, but it’s also possible that no-one has created it before because no-one needs it.

In other cases, a non-AI solution is very good or sufficient. Because there is a lot of effort and uncertainty in developing an AI-based solution to a problem, if you can create a system that relies on a simpler process then that may be enough. Let’s take an example of summarizing news or Wikipedia articles so that users of a website can view a list of them on mobile devices. In this case, just displaying the first three sentences of the article (a method called “lead-3”) is almost as good as many AI-based solutions, and it takes a fraction of the time and cost to maintain! Before creating an AI-based solution, it is worth considering other solutions.

Finally, in many cases, even with state-of-the-art methods, an AI-based solution cannot be expected to achieve reasonable results. Some red flags with respect to this include when:

  1. It is not possible to gather the information necessary to solve the task, such as predicting crop yields without access to a database of weather information.
  2. The underlying answers to the questions are subjective, and humans cannot agree.
  3. Existing state-of-the-art models are reported to struggle with similar problems.

For “1.” and “2.” issues with respect to these can appear when formalizing your problem or starting to consider evaluation methods. For “3.”, this will require a survey of existing related tasks or methods, which is traditionally done as part of the system building process. It is worth noting that in particular, poor performance of state-of-the-art methods may be a hurdle, but given the recent phenomenal advances in AI, it is more of a hurdle rather than a major barrier.

Next Steps

If you have decided that building an AI solution is right for you, the next step is to formalize your problem.