6 Ways To Avoid a Chatbot Apocalypse in 2017

Jason Brenier | Conversational AI

I’m hoping that 2017 doesn’t become the year of the chatbot apocalypse. Not an out-of-control-AI-taking-over-the-world variety of apocalypse, but rather an end-of-species event caused by bot abandonment. This year, thanks to an ever-expanding catalog of mediocre bots rushed to market prematurely, bot fatigue is a true threat from the very public that we’d like to engage.

Since nobody wants a world filled with mediocre bots or bot corpses, I thought that I’d share a few ideas about how to make your next bot project more effective and engaging (before it’s too late). After all, bots have feelings too (at least most of us seem to think so) and deserve better from us. Moreover, because I’m a linguist, you’ll notice that a lot of these suggestions relate to what we call conversational intelligence, or how to make your bot appear to communicate somewhat naturally.

Here are some ideas to help you make your next bot better:

1. Start by introducing yourself. In a chat window, it’s not always clear that you’re even there, so first announce yourself. Give the user some clues that you’re present and ready to start a conversation. This may seem obvious, but as developers, we sometimes forget the basics of human interaction. Say something cordial or helpful. Invite your partner into the conversation and provide some grounding. Ask how she’s doing.

The Poncho weatherbot fails to identify itself, even when asked to do so directly.


Kalani Hilliker’s chatbot, built on Automat’s AI platform, starts the
conversation with a bubbly intro and invites the user to participate.


Kalani Hilliker’s chatbot is an excellent example of how this should work. She starts the conversation by introducing herself, explains who she is and where she came from, and then immediately draws the user in by presenting two tangible incentives: a CoverGirl coupon and a chance to engage with the real Kalani. What could be more enticing?

2. Define the conversational setting and participants. Make it clear how you will be interacting with the user. Is it as a direct conversational participant in a one-on-one chat? Perhaps this chatbot is merely in the background proactively offering timely recommendations or is there to be called upon when needed? Whatever the case, it’s important to inform the user who else can be invited into the conversation, whether it’s friends, other bots or human agents.

Google Assistant lets users know how it can be invoked, when it’s listening in and
when it’s part of the conversation.


In this case, we are extending conversational intelligence to define the context for all interactions.

3. Tell the user what you can do. Are you a chatterbot, a virtual Rogerian psychotherapist, or are you there to help order a pizza? Provide some navigational grounding for your interaction so that the user has a sense of how to interact with you.

Digit’s automated personal savings assistant provides clear
guidance about what it can do.


You don’t have to be a veritable Shakespearean communicator. This menu-driven function serves the purpose that the user needs. There’s not much of a chance of abandonment here with this kind of targeted simplicity.

4. Get to know your conversation partner. Build a model of each user so that you can be a more helpful and empathetic communicator. Although data (and more data) will help, learning about your conversation partner can be done implicitly by paying close attention to what’s been said:

User: I want to transfer $100 from my savings account to my checking account.

(The chatbot learns that the user has at least two types of accounts: a savings account and a checking account.)

Sometimes it needs to be done explicitly by asking direct questions:

Chatbot: I don’t seem to have your preferred branch location on file. Where do you usually bank?

Since learning is such a critical contributor to a chatbot’s value, let’s go a little bit deeper and focus on natural language processing (NLP). NLP can provide an excellent set of tools for learning who your user is, sensing your user’s state, determining what your user would like to do and predicting how best to respond. Here are a few examples of what NLP can do for chatbots:

Sentiment analysis: How is your user feeling?

User: I have a question about my bill, and I’ve been trying to get through to an agent for the last 45 minutes!!! ?

(The chatbot learns that its user is currently frustrated.)

Intent extraction: What does your user want to do?

User: I’d like to travel from SFO to JFK on Tuesday.

(The chatbot learns that the user would like to book a flight.)

Dialogue act classification: What’s the nature of your user’s response? Is it a question, a statement, a confirmation or something else?

User: Yes, that’s the correct destination.

(The chatbot learns that the user has confirmed the travel destination.)

User: What do you mean?

(The chatbot learns that the user is signalling a lack of understanding.)

5. Don’t be dumb. Don’t ask about things that you’ve already been told or that you should know anyway (it’s really annoying when humans do it, and even more annoying when bots do it). If you don’t understand a question or don’t know the answer, it’s okay to admit it, but try to be helpful by pointing the user in the right direction or asking a clarifying question.

The Wonder bot pays attention and remembers important things
about its users for later retrieval.

Granted, forgetting things doesn’t just happen in the world of chatbots. I personally hate entering my credit card number when reaching a call center and then being asked by a human agent to repeat the process. Still, most of us forgive call center employees for not being in sync with their company’s interactive voice systems. But chatbots? I’m not so sure they’ll get a break.

6. Keep a human in the loop. Develop a strategy for escalating tasks to humans when things get too hard. Hand off work to a human when you are asked to do something that is beyond your current intellectual means. Listen to interactions that occur between users and human experts to learn optimal responses for future conversations.

Meya and Front have joined forces to create an environment for
bots and humans to collaboratively solve customer issues.


Although this might seem simple, the end result is significantly better. Leaving humans out of the loop is a risky proposition. Meya’s system addresses this by letting bots rely on humans when they need help and giving humans the ability to pass off tasks to bots when they have been reliably automated.

Some Final Thoughts

At the end of the day, be someone worth talking to. Think about building up a rich inventory of responses that add novelty to the conversation, and keep users engaged. Better yet, use natural language generation (NLG) to create customized responses that are perfectly tuned to specific users and conversational contexts.

Developing engaging products is always a challenge, and the conversational paradigm brings with it a new set of requirements for designing conversational experiences. It’s more important than ever to include UX and design in the development process. All of these examples are simple reminders to chatbot developers that they need to pay special attention to how user interactions take place and develop new techniques for crafting successful user experiences. These techniques are likely to be informed by disciplines other than computer science, including linguistics, cognitive science, psychology, sociology and design.

And I would be very remiss if I didn’t remind you that users feel strongly about controlling the privacy of their data. It’s important to set expectations with users about what parts of your conversation will be remembered and how that information will be used. Otherwise you’ll just come across as creepy. Once expectations are set, follow through with what you’ve promised, and your users will begin to trust you.

Not to put too much pressure on you, but avoiding the chatbot apocalypse is in the hands of a broad product team that is thinking about a far wider set of considerations than when building a web page or an app. The rewards are there for those who get it right. Meanwhile the chatbot graveyard awaits anyone who doesn’t.