Now this is a story all about how we dabbled in making a chatbot for Facebook.
While we mainly focus on showing personalised event recommendations on our website, we also have a Facebook page. There, people often turn to us for support and with questions about events.
Problem: We’re not on Facebook 24/7 (consumption of “piano cat” GIFs notwithstanding). Someone has to answer people while we’re away for the weekend or busy drinking wine — oops — coffee during work hours.
Why not make that “someone” a semi-smart bot?
That’s how we ended up trying out Chatfuel, which lets anyone set up a rather competent Facebook chatbot with no prior experience.
We had two goals:
- Build a support bot that could tackle the most common questions.
- Test out a recommender bot that suggested Valentine’s Day events based on people’s answers.
Here’s what we learned…
Support chatbots = Good
Chatbots work quite well for a limited set of possible questions with straightforward answers.
Guess what qualifies? The infamous “Frequently Asked Questions” section that every single company should be familiar with.
Most of the messages we get on Facebook are support requests. So we programmed our bot to present what’s essentially an FAQ menu in easily digestible, chat-sized chunks…like so:
Did it work? Yes, mostly.
While you can argue that the above is no more than a glorified navigation menu disguised as a chat, having an instant Facebook response helped people quickly get to the FAQ article they needed.
When all else failed, our chatbot was at least able to get people to submit a support request and ask them to wait for a human to answer.
What about bots that recommend stuff? Well, that’s another story.
Recommender chatbots = Ungood
Right out the gate, we had an issue: There are simply too many upcoming events to manually program into a bot.
At any given time, we have hundreds of events listed on Billetto. To make a chatbot work, each of these would have to be evaluated, placed into a distinct category, and added to the bot along with descriptions, links, and images. We’d need a dozen full-time employees to feed a single chatbot (just the way the robots would want it).
That’s why we limited our test bot to a specific scenario: Valentine’s Day. Valentine’s Day was just around the corner, we had a nice mix of Valentine’s events, and there weren’t too many of them.
Our Valentine’s bot walked people through a rather simple decision tree to recommend an event that matched their answers:
How’d it go?
Not…great. When people simply used the options our bot gave them, things worked out rather well:
But as soon as people went “off script” and tried to actually chat to the bot, it promptly fell into “broken Siri” mode:
Only a small portion of those who chatted to the bot ended up clicking on any event links, and none of them bought any tickets. Sad!
To handle more open-ended conversations, the bot would have to be programmed to recognise and respond to a large library of possible words.
Our main conclusion? Chatbots are quite useful at dealing with a specific set of support questions, but getting them to properly recommend events is a lot more tricky.
That’s not to say there isn’t a way forward. It’d probably call for our data science team to find a way to bridge the gap between event recommendations on our site and the Facebook chatbot. Time will tell if that’ll happen.
What about you? Have you had experience either creating or interacting with a chatbot? How would you address the challenges we described above? Do tell.
Watch this space.
Every Thursday, we’ll be posting about the promise and challenges of personalised event recommendations, along with Billetto’s current efforts and future plans.
Have some thoughts on or experience with event recommendations? We’d love to hear them. You can leave a comment or send an email with your thoughts to email@example.com. We’ll read it. Promise.