Chatbot Best Practices — Learnings and Insights from companies like ABN AMRO, Booking.com and Heineken

WeAreBrain attended the Chatbot Conference in Utrecht, both to learn more about what’s happening globally in AI technology development as well as present our product Tur.ai. We took the opportunity to learn as much as we could from companies like ABN, Reaal, booking.com and Heineken on their AI experience and chatbots so far.

It is more complex than I thought. You need to do it well or not do it at all.

A key sentiment at the conference

People first. Technology second

Start with real use cases and real conversations. It is important to approach AI and chatbot development with your user in mind. You need to consider how conversations are constructed. How would a conversation between two humans flow? How many ways could a user ask a specific question and how would you respond naturally. We have found that building a conversation ‘tree’ is incredibly helpful when planning our bots. This allows you to build a full conversation with any type of deviation taken into consideration. When you do this you are also able to take better control of the conversation with your bot steering the conversation to the best possible outcome.

The missing social construct for communicating with a bot

ABN AMRO UX designer, Anja de Castro shared some great actionable insights from her experience. The most interesting, at least for us was around “who” the user believes they are talking. For example, a user generally prefers a picture of human being as an avatar rather than an illustration of a robot or a speech bubble. You may be wondering why and it comes down to psychology and human socialisation. Our theory goes as follows; as humans we interact with humans (who speak), animals (while we’re not able to speak to animals we can pick up “communication’ cues through body language and the noises they make) and then finally things (we grouped both living things like plants and non-living like, well, everything else). As such humans don’t have the mental models in place to create a substantial difference between bots and people. This means there is evidently more comfort in seeing a human face on the other side of a chat. However, we as designers need to manage a level of expectation upfront by telling the user what the bot is able to help with.

Aspirations versus daily reality

The majority of the Dutch case studies presented at the conference were about customer service bots. Added to that a number of discussions amongst the audience centred around bots replacing the standard FAQ list most companies have on their websites. While various presentations looked at other bot applications for e-commerce and sales or alternatively the more generic personal assistants like Alexa and Google Home. It was a curios contrast. It is clear the industry at large are thirsty to develop bots than can deal with more complex human interactions but it appears companies still feel more comfortable with easier applications of chatbot technology, which brings me to the next point! Starting small…

Start small — humility or fear?

The conference gave us great insight into how others are managing and adopting AI. At the end of the day we were, as group, asked to vote on the top lessons learnt throughout the day. Universally the audience decided that ‘start small’ was the key outtake. While we agree that “start” is absolutely important, we’re not sold on small, because sometimes small is too small to matter. And if it’s too small to make a difference it will never be taken seriously enough for it to grow. So I would say “start smart”.

How to train your robot

Repeatedly presenters were keen to dispel the myth that your bot will learn by itself. This is certainly the long term aim, but for now, realistically only the most advanced machine learning algorithms have this kind of capability and even then it still requires substantial training by their human counterparts. This means you will need to monitor your bots’ behaviour and adapt to new questions continuously, which means a user-friendly bot-trainer interface is crucial.

Content is king. And a bigger cost

While there are almost instant solutions that are able to import or crawl your existing FAQ’s, turning the content into a bot (not that content written for a website would ever be appropriate or effective in a conversational dialogue). However, content managers with a bit of experience (like Bol.com with more than 10 years of experience running bots) are very concerned with taking the knowledge acquired through interactions with your bot and putting it in other knowledge databases. So information ‘IN’ seems fine. ‘OUT’ is an issue that remains to be solved.

Assisted human-bot collaboration

While it is something we’ve been going on about for a while, I’d like to mention it once more. Stop using your humans on robotic tasks! The key idea behind AI and automation is to create more space and freedom for humanity. Allow your teams to build, innovate, get creative and truly bring compassion and empathy into your consumer interactions.

Instant trumps human

This one is pretty simple. Convenience. Users generally choose instant help from a chatbot ahead of help from a human later.

Bot creation & management squad

Reaal, ABNAMRO and Proximus all spoke about putting together a multi-disciplined team to launch and support your chatbot. This ensures that you cover all your bases and, added to that, as you refine your bot you have multiple perspectives to lend from. Their suggested dream team: Domain expert, content agents, service designer, conversational strategist, conversational architect, bot developer, conversational designer and a copywriter. And of course, the product owner to drive the product vision, the roadmap and organisational change.

Bot creation and management tools — a young and growing marketplace

At WeAreBrain we believe that user-friendly tools for bot management are crucial. A few presenters also highlighted this factor as a key determiner of success. Sogeti demonstrated their model for a variations in providers, showing three tools live on stage. The difference in user experience was substantial, proving at the same time that conceptual models for bot management tools still leave a lot to be desired. Having said that, improvements are being made consistently. Our key takeout from this particular insight is that your management tool vendor should become your partner, going beyond just delivering technology.

Ideas, insights and coffee breaks

In between presentations, we had a number of really interesting conversations that sparked a number of ideas and thought-starters for us:

• Why not build a bot that is able to guide a user through privacy settings? We’re thinking informed GDPR.
• When choosing the communication style of your chatbot you really need to make it into a personification of your brand. Go deep! Consider looking at various personality tests like Myers Briggs that could really inform the character of your bot.
• If you’re looking for the most authentic dialogue flow don’t tell your test subjects they’re talking to a bot. Knowing it’s a bot has frequently changed the way people ask questions and how they respond to your bot.
• Once you’ve tested and you’re ready to go live DO tell people they’re talking to a bot. It will help serve your users better, especially with current technology capabilities.
• Calling your bot by the same name as a person in your team seems to be a no-go. It just creates unnecessary confusion
• If you are building the ‘look and feel’ of a dialogue yourself, then the design and timing of your bot typing should be taken seriously. These small things may seem inconsequential but it really does make a difference to the experience for a user.

And finally here are some…

Golden Rules

• When building your conversation set don’t talk to colleagues only as a way to develop your bot dialogues. They have too much product knowledge and you’re likely to miss out on the most basic elements needed for a conversation the provides value for your users.
• If you’re building a customer service bot, don’t only use your FAQ list. Think about the ‘strange’ questions you’ve had in the past and use those as well.
• Go live with a subset of your customers. If your bot lives on your website place it only on the pages that are relevant.
• Don’t spend hours and hours building a robust conversation flow attempting to break your bot. You’ll be surprised how quickly you find the holes when you chat to your first version of your bot.
• If you place a bot on “my” dashboard, I expect the bot to be personalised. You’re in my space best you know all about me.
• There are some questions your bot shouldn’t answer for ethical reasons. Work those out and make sure you have built something into your bot conversation to alert you to those kinds of queries.
• If you plan on promoting your bot to the press, it would be wise to create a few media-only scenarios. Anything a little more personalised generates better value for your user.

Until next time, happy bot building!

Anastasia Gritsenko

Previous story

Automation, Bots and Virtual Assistants — AI’s bigger picture

Next story

Meet Yuriy, balancing the studying professional life like a boss