A year ago, at the F8 conference in April 2016, Facebook opened Messenger to bots. From that moment on, everybody wanted one. Agencies, online stores, banks, beer brands, all proudly announced that they “had a chatbot.” The articles spoke of the end of mobile apps.
I have been working with online stores for a few years through Retargeting, and I look at every new technology with a single question: does it bring money or just applause? So after a year of hype, I tried to take an honest stock. As a customer, I tested quite a few bots, abandoned many of those conversations before the end, and found a few that actually helped me. Here is what I think about chatbots in May 2017.
Where the fever came from
The idea makes sense. People spend more and more time in messaging apps: Messenger, WhatsApp, Viber. In Asia, WeChat has shown for years that you can order food, pay bills and book a taxi without ever leaving the chat. On the other hand, few people still want to install yet another app for every store. If the customer already lives in Messenger, why not serve them there?
Add voice assistants to the mix. Siri, Google Assistant, Amazon’s Alexa and Microsoft’s Cortana got us used to the idea that we can talk to a computer. Messaging plus “artificial intelligence” sounded irresistible. The problem is that, in practice, the two are nowhere near equally mature.
What most bots really are
Let’s be honest: the vast majority of bots launched over the past year are not intelligent. They are decision trees with buttons. “Want to see our products? Yes / No.” “Pick a category: Dresses / Shoes / Accessories.” It’s a website menu stuffed into a chat window. But it isn’t a conversation and it isn’t AI.
Where bots try to understand free text, things get messy fast. Natural language platforms such as wit.ai (bought by Facebook), api.ai (bought by Google) or Microsoft’s LUIS handle short, predictable sentences well. “Where is my order?” works. “I ordered a dress yesterday for my sister’s wedding, but I think I got the size wrong and it won’t arrive by Saturday anyway” doesn’t really work. And in Romanian, typed with or without diacritics and full of typos, it works even worse.
Why users get frustrated
The frustration almost always comes from the same place: the promise outruns reality. If a bot introduces itself as an assistant that “understands you,” people write to it normally, like they would to a person. After the third “Sorry, I didn’t get that,” they close the window and, most likely, never come back.
- There’s no way out. The bot keeps you in a loop and won’t let you talk to a human.
- It has no memory. It asks for your order number a third time, though you already gave it.
- It talks too much. Three welcome messages, emoticons and forced jokes before the answer you actually want.
- It sends unsolicited messages. Some brands treated Messenger like an email list and started bombarding people with promotions.
Facebook itself implicitly admitted the problem: at F8 in April 2017 it put the emphasis on buttons, menus and bringing bots into conversations between people, not just on “free conversation” with a program. That’s a good sign. It means the industry has started to get realistic.
Where they already work
There are situations where a simple, well-built bot beats a form or a phone queue. The rule I’ve noticed: bots do well where questions are repetitive, answers are clear and the customer wants speed, not conversation.
- Order status. “Where is my parcel?” is probably the most frequent question any online store gets. A bot connected to the order system and to the courier can answer instantly, at any hour.
- FAQs. Shipping costs, return policy, payment methods, opening hours. Things you already have on your site, but customers can’t find.
- Bookings and appointments. A restaurant table, a service appointment, a delivery slot. Few options, clear structure.
- Simple recommendations. Three or four questions (“Who is the gift for? What’s your budget?”) and a handful of suggested products. It’s a shopping guide, not a brilliant salesperson, but it helps.
- Handing over to a human. Perhaps the most important feature. The bot takes the request, collects the basics and, when it runs out of answers, passes it to a colleague with the full context.
Look at that list and you’ll notice that none of these cases requires advanced artificial intelligence. They require good integration with the store’s data and a well-thought-out script. It’s the same conclusion I reached about marketing automation for small online stores: the technology matters less than the logic behind it.
How an online store can test a bot without burning money
Say you run an online home goods store and you’re considering a Messenger bot. Don’t start with an agency and a budget of several thousand euros. Start small:
- Read your inbox first. Take the last few hundred messages you received by email, live chat and on your Facebook page. Group them. You’ll probably find that five or six types of questions cover most of the volume.
- Automate only those questions. Use a do-it-yourself bot builder with buttons and predefined answers. There are several, with free or cheap plans.
- Add a “Talk to a person” button from day one. Visible, not hidden. A bot that won’t let you out is worse than no bot.
- Say clearly that it’s a bot. Low expectations are the easiest to exceed.
- Measure. How many conversations get solved without a human, how many are abandoned, how many lead to an order. After a month you’ll know whether it’s worth going further.
One detail that matters: don’t use the bot to push unsolicited ads. Messenger is a personal space. A useful message, like “your order has shipped,” or a well-timed reminder about an abandoned cart, is welcome. A promotion a day gets you kicked out of the customer’s life.
What will change as machine learning improves
I need to be careful here, because prophecies are cheap. What we can already see: speech recognition and machine translation have improved visibly over the past two or three years, thanks to neural networks and the huge amounts of data held by Google, Facebook, Amazon and Microsoft. It’s reasonable to expect that understanding written text will move in the same direction.
What does that mean for an online store? Bots will probably learn, step by step, to understand longer and more natural sentences, remember the context of a conversation and make better recommendations based on each customer’s history. It’s the same path that e-commerce personalization has taken: from simple rules to models that learn from behavior.
How fast? I don’t know, and I don’t trust anyone who gives you an exact year. Human language is much harder than it looks, and smaller languages like Romanian will get these improvements later than English. Until then, a good bot is a modest bot: it does a few things well and knows when to call a human.
So, fad or future?
My answer: both. The wave of bots launched just to appear in a press release is a fad and it will pass. The idea of serving customers where they already are, quickly and without forcing them to dig through your website, is the future. Between the two lies the boring work: clean data, well-written scripts, people who take over the hard cases.
What has your experience with chatbots been, as a seller or as a customer? Did one ever solve a problem for you, or did you immediately look for the exit button?