Even if Jim Carrey in Bruce Almighty had a squad of angels to handle the prayers rushing in, he would have needed more than God powers to categorize, triage, and assign those ‘tickets’. Now, God can choose to ignore many of those prayers and yet keep his flock intact. For less lofty entities such as customer support helpdesks, that’s a no-option.
At the core of any company, you will find a helpdesk working hard to resolve user complaints and requests, or what the industry calls tickets. But there also will be times when you find that helpdesk less than efficient. They could be using automation to handle a few critical but repetitive tasks and yet the outcome may not be perfect. Probably because their automation is rudimentary, and not as advanced as the intelligent Ticket Field Suggester that we will be elaborating on in this blog.
This singular artificial intelligence-powered feature has significantly reduced ticket assignment and resolution times for customers of Freshdesk, our flagship customer engagement offering. It has also contributed to improving the productivity of our customers’ support agents, allowing them to focus on more complex tasks requiring their specialist attention.
Rule-based automation: an imperfect middle
Much of the inefficiency in customer support stems from the frustrations mounting on the agents handling the helpdesks. Support teams receive hundreds of tickets every day—or thousands, depending on the size and/or the nature of the business. These flood in from phone calls, emails, Twitter posts, or the company’s website. The complaints and requests are diverse and in language that’s often vague. All these make even categorizing the tickets a huge challenge.
Automation is the obvious solution to this. Typically, rules are set for the system to classify, prioritize, and route tickets to specialist agents based on the nature of the tickets. This works well so long as the tickets meet the conditions in the automation rules. But what if they don’t? That’s when scaling the automation rules becomes challenging. Automation rules have proven to be stubborn, tough to scale, and too cast in those rules to be able to organically evolve.
Over time, helpdesks gather a large number of rules covering very specific situations. This makes it difficult to modify, update, or remove a rule because that could result in unpredictable consequences. We have seen accounts that have over a thousand rules that operate on every ticket.
To be sure, while rule-based automation is important for routing and categorizing tickets, it doesn’t solve the problem entirely. In spite of having automation rules for updating certain ticket properties, agents still have to manually update the rest of the ticket fields. If we apply the Pareto principle here, 80% of the frustration that agents face while handling tickets is because of 20% of the work—the manual updates.
Peaking efficiency with a smart, scalable system
That’s why we built the Ticket Field Suggester. To intelligently handle that 20% of the work and save your agents’ time to focus on what they are best at—delighting your customers. The AI-powered feature we have developed represents significant advances over traditional rule-based dispatch systems by easing operational use, ensuring data privacy and security, and enabling personalization.
In designing the Ticket Field Suggester, we extended automation with Artificial Intelligence-based categorization, prioritization, and routing of tickets. Using historical ticket data, we now are able to automatically identify the nature of the issue raised, what priority it should receive, and which team it should be assigned to.
We chose to use a Natural Language Processing approach since our data is primarily in the form of unstructured emails sent by customers. Inherent in language are strong signals about the nature of an issue and its severity, essential for categorization and prioritization of tickets. We leveraged that for Freshdesk.
A key challenge we faced was scale. We found there were nearly 12,000 Freshdesk customers for whom we could build predictive models to automatically fill fields necessary to route, prioritize, and categorize tickets. In all, we are now managing nearly 35,000 models, keeping them updated and fresh at all times.
This approach is efficient and also speedy. To automate model building, we used Amazon’s cloud machine-learning platform, SageMaker, which reduced the build time to 1 hour from 24 hours.
What’s wrong with traditional systems
To partially automate the process of ticket assignment, helpdesks configure rules that determine how a ticket should be categorized, what priority should be assigned to it, and which agent (or group of agents) should handle it.
These rules are primarily based on keywords and keyphrases detected from the text of a ticket. For example, if the subject and the body of a ticket contains the word ‘refund’, then the ticket would be assigned to the Billing team. This requires strong domain knowledge of the vertical or the industry that the helpdesk operates in to be able to identify overarching patterns across thousands of tickets so that the rules broadly capture the diversity of the tickets. While this may be true in some cases, it is not likely to be true universally.
Problems with this approach:
- Rules based on specific keywords may fail to identify tickets with synonyms or misspelt words that have the same intent. For example, if a customer asks for their “money back” instead of seeking a “refund”, the rule will not route that ticket to a Billing agent.
- Rules, once created, are frozen. They are derived from experience but not flexible enough to accommodate changing patterns after they are created. These rules will do a good job on past patterns but fail to adapt to new ones.
- Over time, rules added by different administrators and agents will accumulate into a complicated, opaque system that makes it hard to modify, update, or remove them for fear of unpredictable consequences. For instance, removing a certain rule may lead to a customer’s urgent ticket remaining unanswered for weeks, resulting in a poor experience and user churn.
The TFS solution
Let’s conclude with understanding how Ticket Field Suggester overcomes these issues by enabling flexible routing as it learns from historical patterns observed in tickets.
- AI models reflect the underlying distribution of support issues and tickets rather than the administrators’ assumptions.
- Classifiers use natural language processing and can accommodate synonyms, misspellings, slang, colloquialisms, and jargon, which rules cannot easily do.
- Classifiers are periodically rebuilt to reflect changes in ticketing patterns, ensuring that they remain current and do not ossify.
- By removing the onus of creating rules from administrators, rules do not accumulate over time and result in a complicated system that precludes change.
In addition, classifiers are built on a per-customer basis with no interaction or data shared between customers, enabling three key advantages:
- Customer data and end-user data remain private and secure.
- Models are specific to the vocabulary used in a customer’s industry or vertical. For instance, the word ‘bank’ has very different meanings in finance and environmental tourism.
- The system is language-agnostic and does not require customers to either use a pre-specified language (such as English) or even specify to the system the language they are using.
Optimizing the future
Using our AI-powered technology for suggesting ticket fields, Freshdesk customers are experiencing significant reductions in ticket assignment times. They are able to get tickets with field predictions to agents faster.
We believe there is much more to be done. We want to offer predictions for more tickets and more customers—and help not only with ticket assignments but also with enabling admins to route tickets to agents in an optimized fashion.
Co-author: Akkshaya Gururajan
Writing by: Feroze Jamal