Leveraging Artificial Intelligence for Increasing Business Efficiency — III
by Gaurav Marwaha and Ritika Dusad, PhD
About this series
In this series, we try to stay away from the hype and focus on how to apply Artificial Intelligence (AI) to specific business challenges and derive meaningful business benefits from this technology. As Andrew Ng, co-founder and former head of Google Brain and pioneer of online education through companies like Coursera and deeplearning.ai has recently mentioned in an interview- focusing on being ‘AI first’ may not be the best approach . In fact, ‘‘in terms of how I execute the business, I tend to be customer-led or mission-led, almost never technology-led.” We concur with his point, and through this series of blogs, we try to point out different methods to be utilize the power of AI while being focused on solving business needs.
In previous blogs of this series, we have introduced the business problem of optimizing throughput of a workflow that is already set in motion. To attempt solving such a problem with AI, understanding the data being generated at various stages of the workflow and how the data flows. In fact, in many businesses like loan processing, assessment of an employee’s performance is based on number of cases they process.
Now that we have understood various aspects that go into identification of a business problem that can be solved through AI and understanding of data elements, we have finally arrived at the stage of discussing a tangible example.
Before we start discussing various AI elements and algorithms, let us do a brief recap of workflow optimization that is done by humans themselves.
A knowledgeable business analyst sits with a technologist to create a workflow design. This design is optimized to achieve maximum productivity in the shortest period of time while keeping overall risk under control. Whenever human tasks are involved in this workflow process, a simple queuing theory is employed to optimize work allocation at that stage, i.e. for each new task in the queue the assignment is based on
→ Current user load
→ Round robin
→ First come first Served
→ User efficiency
→ Pull based pool allocation model
In addition to the above-mentioned points, the following important considerations are also included when a task is assigned-
→ Region of operation
→ Access rights/ role
→ Approval levels
There are facets of human nature that are unknown to workflow and task assignment designers when the workflow is being designed. In any organization that has more than a hundred employees that are geographically spread out and are using a workflow system, the following factors cannot be predicted beforehand-
It is interesting to note that a digital system that accommodates this workflow would typically store historical data for all of the points mentioned in the table above as well as that of cases assigned to agents in field:
It must now be clear why it is impossible to design a workflow that can optimize a workflow when a lot of critical factors are unknown beforehand. Various patterns can emerge out of a system when these factors are studied closely in run-time. These patterns change with time, as workflow users change and geographies vary. The beauty of an Artificial Intelligence system is that when fed with data, it can determine these patterns with far greater speed than that of human intelligence, even though AI was inspired by the human mind.
Before diving into the sea of AI, let us quickly look at a simplified view of task allocation in a workflow. A user action triggers an event that creates a new task to be assigned to another user, notifications are sent to this other user, audit entries are created and assignment is done based on a subset of algorithms that were mentioned in the first half of this blog.
The AI solution proposed to increase business efficiency kicks in right after a task has been allocated by the basic queuing algorithms. Initially the AI program remains dormant when the system is setup; deterministic algorithms are relied upon for task assignment during this early phase. As time passes, audit trails of users’ actions are created and the AI model learns from these. Once ‘enough’ data is gathered to ensure a certain level of accuracy of the AI model, this AI model is put to use to improve allocation strategy.
It is important to keep in mind that the goal of an AI model that is to increase business efficiency is to provide business benefit. In this case, the business benefit can be taken as speeding up of cases flowing through the system.
For this particular case, the AI Algorithm we recommend is Reinforcement Learning (RL), which happens to be one of the three major type of Machine Learning Algorithms that are in practice. The famous DeepMind programs Alpha-Go and Alpha-Star that beat human players also employ RL. To give a brief definition to RL, it is an algorithm that tries to perform a task by maximizing the rewards that it received. Such an algorithm is typically employed when a lot of training data is not available and the only way to learn about an environment is to interact with it. Here is an outline of the key terms from RL that are useful to solve our problem-
- Agent
The program we train, with the aim of optimizing the time required to complete a task. - Environment:
The world in which the agent performs actions, for us the task assignment space is the environment. - Action:
A move made by the agent, that causes a status change of the environment. In our case the agent makes a move to change the assignee of a task from user A to user B, based on past learning. - Reward:
If the action taken by the user results in a faster completion of the task, it leads to a positive reward.
It is important to note that the Agent in our problem can learn from multiple Environments: the Environment of user activity or that of user preference for a task. It is our choice to decide on whether to build a single Agent that learns from all observations it makes across various environments or to build multiple low-level Agents that feed their learning into a presiding Agent; RL allows for both.
Reinforcement Learning can work with zero initial training and as users use the system, the Agents learn and start to take Actions which can result in Positive (faster completion) or Negative (slower completion) Rewards. The algorithm is self-correcting, that is when an Action slows the completion of a task, thereby resulting in a Negative Reward, the model tries to reduce the task completion time through its next Action thus trying to achieve a Positive Reward. This is quite unlike the irrational human mind as we know it.
In this blog series, we defined the business problem of workflow optimization, discussed various data elements associated with the problem and recommended an AI Algorithm to solve this problem for a particular case of assignment of loan cases to datacentre employees. We have kept ourselves limited to discussing the concepts of leveraging AI to increase business efficiency rather than going into details of how AI models can be implemented. We hope that you will now be able to put these concepts to use in your own business environments. Do let us know which AI Algorithm you chose to employ for your particular business problem in the comments section below.
Nucleus Software’s FinnOne Neo Customer Acquisition System comes with an in-built robust workflow management system that allows users to setup complicated task assignment strategies and implement real-time optimizations. The workflow can be managed using a user-friendly interface and its metrics can be extracted from the system. At Nucleus, we are always eager to partner with those financial institutions that want to explore the use of AI in loan origination, as AI unravels the next frontiers of the Financial Services Industry.
