Everyone from governments, to universities, to companies, large and small, seems to develop an AI strategy. AI appears as the most desirable addition to any organization, one simply must have it. But before you rush to joining the AI craze, it is important to think of the reasons and the degree of fit between the characteristics of your company and the benefits and costs of AI implementation. AI is not a good in itself. AI is a set of tools that needs to have a clear place and purpose in your organization to be truly beneficial. For example, AI can delegate decision making to algorithms and thus free up more human time to dedicate to other more complex and visionary tasks. In addition, AI can help companies meet customer needs and expectations. AI can solve problems and improve processes, but these must be adapted to the specific context of your organization. The first step to consider introducing AI technology in your workplace is to evaluate the existing processes and structures and see if AI can help solve problems or improve flows in your specific arena. Not all problems need an AI solution. For example, many government authorities or companies want to introduce chatbots as a way to improve user interaction with websites. However, looking at the type of questions that a regular customer service gets, may reveal that most people are concerned with the same set of issues and that he answers to these issues. Bow down to either filling out one or two forms or getting in touch with a human expert, in this case introducing a chatbot would be overkill. Instead, optimizing their website navigation and providing a FAQ with better links is a faster and simpler solution. So no AI for a AI's sake but contextualized it. Building an AI strategy relies on two pillars. One, evaluate and decide if AI is applicable to your organization and would help solve problems and improve processes. And two, if the answer to 0.1 is yes, then ensure that your organization is well placed to maximize the use of AI systems. To address the first pillow, you must start by mapping and evaluating the existing processes in your context. Implementing AI can start even with a single process that is to become automatized. That's the first question you need to answer, inspired by the work of Ronald archery is, looking at the status quo, how does the problem gets solved at my workplace? What steps do we go through to deal with the problem, and are the clear rules that define the problem-solving process? Let us use a fictitious example to guide our thinking through this exercise. Let us say that the problem you are trying to track is how to reimburse expenses. The goal is to delegate the decision to reimburse or not to an algorithm. First, you need to chart the existing ways of solving the issue. To do so you interview the people in the organization who are currently involved in reimbursement approval. To keep it simple let's assume there are only two, Alex and Robin. When talking to Alex, he walks you through every step of the process. How he first categorizes every expense into per-existing categories, travel, hotel, books, participation, fees, etc. He then looks at the sums involved and compares them to the budget allocated to the person and to the activity. Finally, he checks if there are internal restrictions as two categories of activities that can be reimbursed and if there are limits to these sums. If the reimbursement request has met all the conditions then it is approved. This process is very precise but also quite time consuming and there's poorly with expenses connected to new activities. The other person whom you interview, Robin, describes her process in very different terms. She does not consult the internal handbook nor does she categorizes all expenses. Instead, she looks at the history of the expenses for the person requesting the reimbursement and matches the size and type of expense to their previous standards. She also consults her record of yearly complaints. If she finds the present reinvestment request to fit with the past behavior of the person demanding it. And if she also finds there were few or no complaints about the past decisions, she decides to approve the request. This process is more efficient than Alex's as Robin spends much less time considering each individual request. It also has the risk of doing double work or misclassifying some expenses. The two persons in the example illustrates to decision making processes that can be connected to two types of digital technologies. Alex's approach describes a model driven problem solving path, whereas Robbins is a data driven approach. Identifying and classifying the problem solving processes in your workplace will help you choose the right digital system for your context. A model driven AI system depends on the existence of explicit knowledge about an issue. The description of the relationships between entities belonging under that issue, and in particular on the existence of rules that govern the overall behavior in that area. These rules constitute knowledge that is mastered by experts on that issue. The experts must explicate the rules for an AI to construct a model of that segment of the reality. A real world example of a model driven system comes from life sciences. Their medical experts make diagnosis on the basis of patient information. Just looking at the patient information even lots and lots of it would not be helpful in making the diagnostic. Experts need to classify the information and tell the computer what signs to look for to identify a disease. As the medical diagnostic example illustrates the model driven AI has the advantage of transparency. It is very clear why a diagnostic was given by looking at the rules followed. One of its disadvantages on the other hand is that it works well only when the rules are explicit and clear. One of the most famous exercises in AI image classification is to ask an algorithm to identify a dog in a series of pictures. Basically the algorithm is asked to classify images in dog and not dog. A model driven AI would require writing some explicit rules on what defines a dog? So take a moment to think about which rules would that be. [MUSIC] A dog is a four-legged animal. But hey, what is an animal? With two eyes, but what are eyes? Facing forward. Where is forward? And two years on both sides of the head. What is a head? As you can tell from this mental exercise, the fact that defining a dog must be cross referenced with defining other categories, species, body parts, directions and others. Makes creating explicit and clear rules very difficult. This is where data driven AI may be more suitable. This approach does not strive to define rules of formulas to model reality. It starts with the reality and trains the algorithms to find the rules that govern this reality. This is the purest form of AI in which algorithms find out and teach themselves the rules behind the decision. The process is deceptively simple. It starts with very large data volumes about the process or issue we are trying to understand. In our case, it would start with millions and millions of images of dogs and not dogs. Human coders annotated data so that the percentage will be labeled dog or not dog. This training data is fed into the classification algorithm that produces some results. These results are evaluated as correct or incorrect, either by humans or by the algorithm itself and a new round of tests is performed. At every round the corrections are taken into account so that the results are continuously improved until there is a high percentage of reliability. The computer was able to correctly identify dogs in a majority of the pictures it was exposed to. Data-driven AI is suitable in context where there is ample data that documents the process we want to automatize. And whether our existing algorithms to help classify this data, be text, image, audio and so on. Going back to our reimbursement example. Robbins approach illustrated the data driven solution by relying on historical data about a person's behavior and on corrections given in the past when making the decisions. One of the downsides of the data driven approach is its lack of transparency and explicability. Even when the decision is correct, the algorithms are not always able to report why the decision was taken. This may have unwanted or unethical implications which we'll discuss more in week three. The area of explicable and transparent AI is fast developing right now. We have now described the existing processes in the organization that deal with specific problems where we think AI can help. The next step to consider is what information exists on hand to help train and customize the AI system. [MUSIC]
