Scaling customer support with AI and chatbots

A deluge of support requests often accompanies business growth. Firms might hire more agents to scale their service organizations, but that approach does not guarantee efficiency – organizations at scale also need to ensure faster service. And to do that, they need automation. Andrew Navin, a Customer Service veteran, explains how he scaled his support team’s capabilities to match his organization's mammoth 6x growth.

Andrew Navin

Customer Service Leader, Freshworks

What's in this discussion

  • How Freshworks scaled customer service using AI
  • The need for customer service automation
  • How AI helps with deflection
  • Ensuring accuracy of data with bots in play
  • Bot use-cases for customers and agents



Part 1: Scaling customer support to accommodate 6x business growth

Andrew Navin: As of 2020, Freshworks had experienced a 6x growth spurt over six years, in terms of ARR and customer base. The increase in the customer base had a direct impact on the number of support tickets we received. What is the best way to handle such a rapid increase in the ticket count? One option was to hire aggressively. In the last seven years, we grew from a one-member support team to a 130-member support team. We hit our target metrics (such as CSAT) month-on-month, and felt the need to push those metrics even further to thoroughly delight our customers. This led us to experiment with technology to bolster our customer service operations. Once we began considering the addition of AI to our support repertoire, we had to answer two questions:

  • What kind of tickets would AI help us deflect? We analyzed the tickets we were receiving and categorized them as L1 (simple questions) and L2 (troubleshooting). We found that nearly 45% were simple L1 queries that had been asked several times in the past and could be resolved within minutes – a good candidate for AI-based deflection.

  • Which ticket source should we focus on first? On analyzing the sources we found that the increase in ticket volume was partially because of our omnichannel operations. We focused more on live channels like chat and phone to build better relationships with customers. Chat had even overtaken the traditional email channel in terms of popularity. This became an obvious choice for the AI experiment.

Part 2: Automate the simplest queries first

We aimed to automate the deflection of L1 tickets so our agents could spend less time on menial tasks, and more time on complex ones. This way, customers would get quicker responses on simple queries because they wouldn’t have to wait for the agent to search for a basic help article and send it to them.

We started small so that we didn’t hamper our existing support experience. We replaced the Raise New Support Ticket option in our support portal with an Ask Freddy button. The idea was to have Freddy (our AI-powered bot) help customers by suggesting knowledge base articles that aligned with their query. And if their question was a complex one that Freddy couldn’t answer, only then would they be prompted to create a ticket.

This was a good idea on paper, but reality confronted us. AI-powered deflection is only as good as the content it can pull from. We found that our knowledge base did contain answers to most L1 queries – however, they were buried deep inside solution articles and were not easy to find or read. This would have been disastrous for a customer hunting for a precise answer. We soon gave every article a clear title, made it short and easy to consume, and eliminated redundant articles and duplicates.

At the end of this effort, we reduced the number of knowledge-base solution articles from 426 to 200. We also reduced the number of categories, and gave the knowledge base much-needed structure. We also published 800 bite-sized FAQs. Now that we had the content ready, Freddy would be a great medium to channelize and deliver it to customers.

Part 3: Measuring the success of bots and deflection

Freshdesk’s reporting capabilities were essential to helping us understand if our efforts had been successful.

We found our ticket deflection rate to be 12%, with net enquiry deflection reaching 30%. Remember the ticket source split we looked at earlier? The support portal contributed to 12% of ticket volume. But now, only 5% of our total enquiries were email-based support tickets. The key takeaway from this exercise was that content creation and bot training is critical to success. 

Another takeaway: Measuring these outcomes is important, because they not only tell you whether you’re putting effort in the right place, but also allow management to realize the impact of the support function.

Part 4: Types of agent assistance bots

Using Freddy to deflect customer inquiries was just one of the experiments we carried out. The other experiments were focused on agent assistance.

  • Agent Guidance Bot: At Freshworks, the agent onboarding process is quite elaborate – it’s six weeks long! However, it’s only normal that even after training, there are times when agents require hands-on help during customer conversations.  The agent guidance bot provides it by prompting them on the next best action, regardless of which channel the agent is operating on. By guiding agents to standard resolutions, a well-configured bot can introduce consistency to support interactions.

  • Email Bot: It is primarily focused on deflecting email-based L1 questions. Why? Because assigning a ticket to an available agent takes up to 2 hours – why should the customer have to wait that long if our knowledge base already has an answer to a simple query? We simply have the email bot analyze the email and auto-reply to the customer with relevant knowledge base articles and FAQs. The time and effort this bot saves us are tremendous.

  • Thank-you-detector Bot: Out-of-office emails and thank-you emails can accidentally reopen tickets and wreak havoc on support metrics. Our agents were particularly worried by this issue, so we instituted a bot that analyzes the content of every incoming email and decides whether a particular reply should trigger a ticket-reopening workflow or not. 

  • Sentiment Journey Detector: Freddy tries to understand the customer sentiment in every conversation, regardless of channel. It then maps the customers’ perceived emotions, allowing supervisors and managers to sense frustration across support interactions and take corrective actions. Using sentiment analysis helps us be proactive and preemptively fix actions that might irk our customers.