In our most recent newsletter, I mentioned how our customers are telling us that their management is shifting their view of ICT. Instead of being seen merely as a cost, it is seen as a business enabler that is necessary to support businesses through their digital transformation.

We share that view. ICT is an enabler in our own business, and we outlined six projects we’re working on to help with our own business transformation. You’ll see details on our Website and LinkedIn about our system health and ticketing tool (we call it the CM View).

One of our most fascinating transformation explorations, though, has been with Artificial Intelligence (AI) or Machine Learning. Whilst this project was something of an experiment, I am pleased to report that early in the project we are seeing promising results.

The AI project has been easier and less expensive than anticipated, and our staff are enjoying being part of something that sounds futuristic enough to come from a sci-fi novel. Far from the stuff of a writer’s imagination, after a little back and forth we managed to identify a couple of worthwhile projects with a good chance of success.

We have two projects underway.

Project 1 – Interpreting Emails with AI

Business Benefit: Improve Customer Satisfaction and Increase Revenue
Budget $40,000
Estimated Revenue $1,000,000 plus

Aim. We wanted to gain insight from customer emails to understand how they feel about us; what they need help with; and who we need to do better with. We wanted to use AI to forecast trends, and highlight where action is required. We will also gain strategic insight about the technologies we should prioritise for future investment.

Criteria:

  • Real-time email collection
  • Email content sentiment analysis
  • Long and short term customer satisfaction analysis
  • Individual email search, based on:
    • Customer
    • Date
    • Industry
    • Recipient
    • Customer interest and trending topics

Process:

The system will process about 1,000 inbound emails every day that are linked between our email server and CRM. This is a critical piece of the process, as it enables us to identify customer emails and sort into the categories we want, such as industry verticals.

The tone of emails will be analysed using Natural Language Processing (NLP) tools from the Watson APIs suite. Negative emails will be flagged by the program for the marketing team to review and action. The application also collects the content of every email, and counts occurrence of key words to gauge interests based on industry or individual customers.

The application is written in Java and implements the JavaFX library for GUI development. 24 hours of programming have been invested, with another 40 hours to come. A dedicated server is running the application and making the Watson calls. We managed to keep the approximate cost, including hardware, well under budget, making it an inexpensive way to improve our service and customer insights.

In the likely event we choose to put this into production, there will be ongoing charges for the Watson calls, and we will need some additional code to CMLive, so that we can give our customers real time insight to the trends we are seeing.

Project 2 – Simplifying Choice with Machine Learning

Business Benefit: Improve customer satisfaction and reduce cost
Budget $50,000
Potential Return $30,000 – $80,000 per annum
Aim. We wanted to simplify the way we process tickets from our Remote Monitoring and Management (RMM) tool. Our CM View automatically generates tickets on system issues it uncovers, with some automatically closed when the issue resolves itself without intervention. With machine learning, we will be able to predict which tickets are unlikely to need assistance, so we can focus our energy on the tickets that do. We can improve our customer service by focusing resources in the right places.

Criteria:

  • Accurate predictions
  • Live predictions
  • Updating model

Process:

A large volume of tickets are automatically generated by the CM View monitoring system each day. Over half are automatically closed by the CM View when the issue resolves itself. The goal is to identify in advance which of the generated tickets will be automatically closed, so engineers can manage their time more efficiently.

To perform this classification, a Decision Tree Classifier (DTC) has been used (see below diagram with sample results). The DTC is trained on existing ticketing data, and past outcomes are used to predict behavior of future tickets. The system can provide a level of certainty about whether the CM View will close the ticket as it is generated, with results shown to the engineering team.

The model is providing almost 90% accuracy so far – and the more data that is analysed, the more accurate the predictions will become.

Our ticketing project uses the Python Scikit-learn library to perform classification tasks. Identifying the key sections of the ticketing data has taken the most time, along with tweaking parameters of the model to achieve the greatest accuracy.

We have invested 60 hours of programming and research so far, but as we become more familiar with the system, we can work far more quickly. From here, the remaining work is to integrate the model into existing workflow, and enable use of data that is constantly updating. Another 40 hours of programming time will be needed to complete the project, with the cost including hardware staying well under budget.

Lessons learnt
As the options become more sophisticated, they can be readily matched to business needs. What is important is to have a clear direction from the start – as with any new technology, it is only worth spending your budget on AI if there is a business case. When you have a clear outcome in mind, AI is no longer futuristic… it has a place in improving your business efficiency today.