Artificial intelligence (AI) or the intelligence of machines, is quietly becoming a part of our everyday life. AI is increasingly used to automate the business process and make them smarter. However, AI system designers have a challenge. How to put AI in in ways where it can be best leveraged without creating fear of uncertain outcome? We have made significant progress in the design of the control space and establishing boundaries between human and AI forms. Most of prominent of them is Human in the loop (HITL). The most ubiquitous example of HITL is when each time we look at the google search results and select the most relevant ones, we become the human part of the learning loop for google search. This collaboration between AI forms and humans to achieve an optimum result is what makes HITL so important.
AI is a superset of many concepts like machine learning, deep learning, neural network, etc. In machine learning, machines learn to take decisions based on a trained model. The model is trained on large data sets. However, the quality of the trained model and the accuracy of the prediction depends on the range and diversity of the data set and the choice of appropriate algorithm. Confidence score to the prediction tells the probability of the accuracy. While it is fairly common to achieve confidence scores of 80%, achieving a higher confidence score of 98% to 99% is difficult because there are always outliers and hard cases a machine just can’t figure out.
Depending on the context, right threshold value for the confidence score is chosen when to bring humans in the loop and collaborate with machines. Human-in-the-loop drives better results in by taking advantages of AI strengths to provide a fast judgment on a complex data set repeatedly and human strengths like intuition, empathy, and strategic understanding. AI’s two significant weaknesses, lack of strategic judgment and lack of intuition is being complemented by the experience of humans in the business process to achieve useful outcomes.
In my organisation we have developed an AI-based platform for software testing optimisation which optimises timelines and reduces effort for the testing cycle of software. This platform consists of several AI enabled Bots. An automated cognitive defect bot performs defect similarity search using contextual algorithms to identify similar errors and classify them. Then human-in-the-loop workflow is kicked off, and a human expert reviews the classified similar defects. Human expert either accepts or rejects the defect classified as duplicate defects based on different factors like proper identification of similar errors and historical percentages of eliminated defects. If the defect classification values are below the acceptable limit, the human expert adjusts any necessary input parameter for the Bot to improvise the next execution. This is a great example of humans and machines working at their best to amplify the effectiveness of AI.
Human in the loop has wide range of applicability across business processes and across industries. For example, it is used in customer service, where AI powered chatbot converse with users, but service agents takes over when chatbot fails to address the conversation effectively. Another example is in the insurance industry, where insurance claim processing is handled through AI powered smart processes, however human touch points are introduced appropriately to verify and improve the accuracy. Another example comes from Stitch fix , a subscription clothing company. Users of stitch fix, visit its online site and fill out style surveys, provide measurements, offer up Pinterest boards, and send in personal notes. Machine learning algorithms then come up with recommendations. These recommendations along with personal notes and other contextual information is sent to the company’s fashion stylists, who then select the items to be sent to customers, rejecting others there by improving the recommendations. Customers keep what they like and return anything that doesn’t suit them, which further extends the HITL in an interesting way.
Successful HITL really lies in identifying the acceptable accuracy thresholds for AI powered activities within a business process. Real effectiveness however comes from designing the right interaction model between the AI powered activities and humans, such that process is most efficient and outcome is most accurate. Such interaction model not only creates smart processes, but also processes that continue to learn from humans.