By Ana Celeste Almeida Vieria, Player Support Manager, Nordeus
Ana Celeste Almeida Vieria, Player Support Manager at Nordeus, embarked on an automation journey just over one year ago. Here’s how she went from fully manual customer support to automating 80% of all incoming tickets — and improving CSAT in the process.
For additional insights from Ana, watch the Player Support Webinar
I’ve been a member of the Nordeus customer support team since 2010, when automated customer support felt about as realistic as colonizing Mars. By 2018, though, customer support technology had morphed into a whole different ball game — and we decided to start playing.
We decided to start with our most popular game, Top Eleven. Top Eleven is a football management simulation game that launched in 2010, and has been incredibly popular ever since — today, we have over 220 million registered users. Of course, with this popularity comes challenges for support teams — we support thousands of players on a monthly basis, helping them on their football management journey. My goal was to support every single one of them in the most efficient, scalable, and cost-effective way possible.
Beginning the automation journey: experiment, learn, and experiment again
We decided to begin our automation journey with Top Eleven because, with a strong and reliable user base, we could use this to our advantage in trying out ways to give our best support. While we do have regular releases every other week, our player confidence enabled us to explore without high risk.
Below are the steps we took to successfully automate. Keep in mind that the approach we took is not prescriptive, but hopefully can shed some light on what your own automation journey might look like.
Step 1: We built a foundation of issue labeling
The journey started in December, 2018. The first thing we did was to implement Helpshift’s SensAI Predict and train it to label incoming issues. After two months, SensAI labeled issues with high accuracy (it’s currently at 94.9%!), but we continue, even to this day, to give the model constant feedback. This regular iteration ensures that the AI maintains its high accuracy through every new release and in the face of unforeseen new issues. Today, 42% of incoming issues are automatically categorized into one of 23 labels created through AI training.
Step 2: The fun part – we deployed over 20 custom bots
Once you have a foundation of predictive labels, you can deploy bots based on them. Currently, we have over 28 bots in production, an iterative process that took around nine months. Building these bots was a learning process, but one that was extremely rewarding. Today, 80% of all incoming tickets interact with a custom bot, and 62.87% of all in-app issues never interact with a live agent. To drive that home: over half of our incoming issues are resolved exclusively via automation.
Step 3: An easy win with QuickSearch Bot
QuickSearch Bot was an incredibly quick and easy win for us as it complemented our existing FAQs. We saw a deflection rate of over 5% within the first month of implementation — and the implementation didn’t require any development or time investment. Especially if you have up to date FAQs and in-app deflection is high, QuickSearch Bot is by far the simplest way to get an automation win right off the bat.
To get to all of these wins, I learned a few lessons along the way:
- Make sure the user knows they’re talking with a bot
Set expectations and establish trust and honesty — there’s no reason to hide the ball here.
- Never give more than five options in the initial menu or more than five steps throughout the bot flow
I learned this one the hard way, through testing and experimentation. After five steps/options, you begin to jeopardize your player’s experience.
- Keep a log of changes to bot steps
Here’s another one I learned the hard way: keep a change log, especially if you’re regularly updating and iterating your bots in response to product updates.
- Set KPIs for your bots
Be clear from the beginning about what you expect your bots to give you. Is it end-to-end resolution? Improving efficiency through upfront information collection? Setting expectations is important both for customer service managers and for the players themselves: be clear about what the bot’s purpose is.
- Never stop experimenting
Building and deploying bots is a really fun journey, but it requires attention, creativity, and a few failures. The results may not come immediately, but believe me, they will come. Once you start seeing results, though, don’t rest on your laurels. We’ve already automated well over half of our customer support, but our roadmap is still long: this year, our goals include improving CSAT, deploying localized bots, enhancing custom bots with rich media, enhancing QuickSearch Bot with updated FAQs, and automating in four other languages besides English.
The “Cost” of Playing the Automation Game
Lowering operational costs and improving efficiency was my goal at the beginning of this journey. I knew that bots and automation infrastructure would add some costs, so I needed to be sure that it paid for itself and more.
Over the past year, our operational costs have decreased by 40%, we decreased time-to-first-response by 21 hours, and decreased time-to-resolution by eight hours. One of my biggest concerns about the cost of automation was that it would lower CSAT — but last year CSAT stayed flat, and in the first month of this year, from January 2020 to February 2020, we saw a 10% improvement! Like I said above, you may need to spend time, effort, and creativity to get all the results you want, but you will get them eventually.
So is embarking on the automation journey worth it? I would give a resounding yes.