Original article was published on Artificial Intelligence on Medium
There is a misconception that Artificial Intelligence (AI) can be used in every context. However, the truth is that you can not always expect AI to solve problems in your business even though it might seem to be a good fit. When it comes to using AI, you definitely must be cautious since the business has a lot of elements that do not exist in academia where AI is heavily used.
In business, there are many considerations such as resource limitation and customer satisfaction that define the development process and enforce acceptance criteria associated with technology development. In this article, I want to introduce an analytical framework that helps you determine whether AI can suit your business needs.
How can I define AI in a more practical way?
With some exceptions, AI is a technology to design computer-assisted tools to enhance the quality or the number of decisions when there are uncertainties to make decisions. Below, I explain two different use cases of AI where there is a need to increase the quantity and improve the quality of decisions.
Increase the number of decisions
When we need to explore a large volume of data, we can not do that manually due to the number of decisions that are to be made.
For example, an individual can not effectively search in a large dataset of text, image, or voice without using an AI-based search engine.
Improve the quality of decisions
When we need quality decisions, we may not simply have access to the field experts due to the lack of financial or human resources.
For example, we may not have access to specialist physicians wherever and whenever needed.
What are the common applications of AI?
In general, AI can be used effectively in the series of tasks listed below:
- Predict the future using historical data such as weather forecast
- Identify outliers that exist in the data such as fraud detection
- Perceive complex unstructured data such as document understanding
- Extract underlying patterns in the data such as recommender systems
- Optimize processes to save time and energy such as personalized ads
In all the above use cases, we use AI to enhance the quality and quantity of our decisions in different contexts using data-driven methods.
For example, if we want to show an ad to a group of users, we may not simply decide who should see the ad. AI helps us optimize this process.
Note that these are some common applications but you can not simply generalize them to every use case.
Can you use AI in your business?
When you want to consider using AI in a business use case you must always use the “expectation vs complexity” analysis framework. If this framework does not suggest viability in the use case, you must think twice before spending much time and money.
Here, I want to describe what I mean by expectation and complexity in this framework.
“Expectation” means how much we expect an AI-based solution to provide advantages for the business. It can be measured by a function based on 3 variables: quality, quantity, and value.
Expectation = f(quality, quantity, value)
- Quality is the expected performance associated with the use case. Quality is rooted in false positives and false negatives as well as other parameters that can be used to measure the experience. There are many standard metrics to measure quality but you must design your own metric.
- Quantity is simply the number of times that the AI-based solution is used to make a decision in the process.
- Value is the monetary value that you collect or lose by each output of the AI-based solution. Value is rooted in the business use case. For example, failure prevention in mining industries can create tremendous value even though failure may not happen frequently.
The explicit definition of expectation function is different in each use case. You do not need to think of this function as a complex algebraic function; I used this format to convey the message!
“Complexity” means how difficult it is to build an AI-based solution to effectively solve a problem. It can be measured by a function based on 3 variables: data, technology, and problem.
Complexity = f(data, technology, problem)
- Data- To solve any problem with AI, you have to access a large volume of clean data, labeled or unlabeled. Having a large volume of clean data, especially labeled data, is very expensive but you can not solve the problem without it.
You always must allocate a considerable part of your time and budget to collect clean data.
- Technology- A problem that was not solvable 5 years ago can be solved today due to the fast advancement of technology. For example, computer vision did not find its place in the industry before the deep learning era. Now, there are a lot of specialized deep network architectures with pre-trained weights that can be employed to solve a significant number of industry problems. You can read more about the technology readiness level here.
The required technology to solve the problem must currently exist, at least to some degree.
- Problem- Sometimes, problems are not just solvable by AI. For example, the future values of company stock on a long horizon do not significantly depend on its historical values. So, you simply can not use AI to predict that. AI is used in algorithmic trading for short-term predictions, but that is it.
The problem must be solvable by AI in nature; otherwise results may not make sense.
The explicit definition of Complexity function differs in each use case. It is difficult to quantify complexity for a certain use case. That said, the use case can be benchmarked against existing cases to provide you with an estimate of complexity. If the complexity of developing an AI-based solution is high, you may need to revise your development plan.
To build an AI solution, you confront a lot of challenges that consume time and money. I suggest you analyze your business case within the “Expectation vs Complexity” framework. If this framework suggests viability in your use case, you can start spending time and budget. Otherwise, there might be other methods for your problem that you should explore.
I hope you enjoyed this article and can use this framework in your decision-making process. I will present more practical examples of this framework to show how it works in the future.