Achieving Seamless Service Provisioning with Intelligent Bots
Digital Service Providers (DSPs) have begun leveraging RPA to automate various processes in telecom industry especially in operations and business support systems. The core of Order-to-Activate(O2A) journey processes are highly automated using standard RPA bots such as service provisioning and activation. However, DSPs are not able to achieve complete automation as it requires considerable manual assistance in processing unstructured data and in decision making.
Orders flow into the order management portal through various channels such as e-commerce, call center agents, retail and sales agents etc… The RPA bot fetches all the required details from the order management portal and only processes the orders with structured data and to execute the subsequent steps to complete the automated service provisioning. As significant numbers of orders are unstructured, human intervention is involved in the validation process and to initiate the next best action for processing unstructured orders.
Handling unstructured data is the key challenge
Standard RPA does not help DSPs in end to end automation, which results in more processing time and impacts overall order to activate (O2A) cycle. The major challenge arises while handling the unstructured data with typical rule-based bot
- Traditional RPA bot accept only structured input data and work on rule-based approach
- Standard bot gets struck with some exception while performing transactional activities and requires manual intervention for subsequent action execution
- Manual intervention increases overall process time and the chance of human error – high risk factor
- Response data in order fallout process are in free text format and it needs multiple scripts to extract the required strings from the unstructured input data
- Decision making on certain tasks require human intervention which delays the process completion
According to a Forbes report, 80% of enterprise data existing today is unstructured. Processing of unstructured data requires innovative approach using NLP (Natural Language Processing)/ NLU (Natural Language Understanding) technologies.
Developing Machine Learning (ML) based solution strategy to achieve seamless service provisioning
NLP/NLU-based decision engine can play a critical role in processing the unstructured data, derive insights and provide the next best action. The major four steps involved in implementing NLP/NLU decision engine for end to end automation of service provisioning processes are
Building the decision engine using NLP/NLU technique
The core decision engine comprises of two major functional components namely pre-processing module and prediction module.
Pre-processing unstructured data with natural language toolkit
Python based natural language (NLTK) toolkit is used in unstructured data pre-processing and developing a model to dynamically interpret text data by applying various control methods such as
- Vectorizing data: Bag-of-words
- Word cleansing
- Stop words removal
- Word lemmatization and categorization
Predicting next best action with classification techniques
This stage involves building classification models to predict the next best action for the bot to execute by leveraging SKLearn library. The machine learning algorithms used in decision making are
- Support vector machine
- Random Forest
- Gradient Boosting
- K-Means and DBSCAN
Based on the input text data, the NLP/NLU based decision engine predicts the next best action to be initiated by RPA bot. The decision engine can be optimized to improve the performance and efficiency by fine tuning the hyper parameters (to improve the accuracy) and retrain the model with latest data set (to achieve better prediction results).
Deploying on-premise or cloud-based decision engine
This step involves deploying the NLP/NLU-based decision engine in the form of Containers on cloud-based platform or on-premise infrastructure and then exposing the service in the form of API/SOAP.
Integrating NLP/NLU based decision engine with standard RPA bot
This step primarily explores the various aspects of the standard RPA bot integration with the order management and decision engine in the existing system. Bot consumes API/SOAP service from the NLP/NLU decision engine to carry out next best action and to complete the remaining service provisioning process steps.
Active self-learning framework to improve prediction accuracy
Completely replacing the human intervention with a NLP-based decision engine in one go may not be a wiser approach. The model must be trained for a significant amount of time to increase the confidence level. This requires DSPs to build an active learning framework, where both NLP/NLU decision engines and manual Subject Matter Expert (SME) validation co-exist to increase the prediction accuracy.
Process flow involved in active learning framework includes
- Standard RPA bot provides the unstructured data from portal to the decision engine
- NLP/NLU decision engine predicts primary and secondary action steps to be taken and gives input string for SME annotation
- After SME performs the annotation, the final response is fed back to the decision engine for self-learning and appropriate prediction
- Prediction results with next best action are given to standard RPA bot to complete the remaining service provisioning steps
Active learning framework improves the prediction accuracy level over a period of 3-4 months. Once the NLP powered intelligent RPA bot gets trained enough to handle different scenarios, the manual validation step can be avoided eventually, to maximise automation potential.
Key benefits realized after implementing intelligent bot-based solution
- Three-fold (3X) improvement in order processing efficiency
- Complete removal of SME dependencies after intelligent bot implementation
- Prediction accuracy can be improved by 85-90% using active learning framework
- Reduce manual efforts by 80% using NLP/NLU decision engine with standard RPA capabilities
- Faster order fulfillment resulting in superior customer experience (CX)
By Avaiarasi S
