The nerve-racking process of putting your bot live.
As one of our chatbots is about to see the light, I decided to calm my nerves by writing about what I have learnt about the process.
Identify the best process for your business
So far, we have tried two different approaches to putting a bot live: massive rollout or segmented rollout.
The first approach takes balls, but is definitely the way to go if you are going for a big product launch and you are focusing on marketing. If you have built a well defined bot, that does just a few things right and is mostly guided by buttons and rules, then you are safe. But if you’ve built a customer service bot, you might want to rethink this approach, as it is very risky. One of our clients decided to go live with his Facebook customer service bot with all its 300k followers. It was a risky move and luckily it all went great, but if anything had gone wrong, it would’ve been really bad.
The second approach is safer, as you work incrementally once you know your solution is working as it should. It really doesn’t matter if you work with 5%, 10% or 20% of your customers, the important thing is that you try it with enough people to have a wide range of users and cases to learn from. You can also target certain type of customers or do it randomly, it all depends on how well you understand your user in the beginning and what you want to learn.
Manage expectations and control the narrative
Always, always, always communicate what your bot does and what it doesn’t. Clients always ask me if they should tell their customers that they are speaking to a bot and I always give the same answer: Yes.
Let’s say your customer is someone who has a negative predisposition towards bots, if you tell him first hand what to expect, you have the opportunity to under promise and over deliver. But if you don’t, any little mistake or any friction in the journey is going to matter, a lot.
For me, the worst scenario is having a user thinking that something the bot is not trained to understand, it’s an error. He is judging your solution by something it wasn’t meant to do, and that, somehow, invalidates everything that it does do.
That’s why controlling the narrative is so important. Chatbots are tools, and you need to teach your user what he can do with it. If you use a hammer to open a door, you are not going to have a good experience.
I have also noticed that people don’t read introductions and tooltips, so don’t be subtle with your on-boarding process. Be explicit and invite curiosity.
It’s going to have a life of its own
If your solution is a true chatbot -that means that it uses a NLP and is able to handle an open conversation-, it is not a deterministic solution. That means you are not able to control all posible outcomes from all user interactions. That comes with the territory and is important that your client and users understand that.
There are bound to be some mistakes that you will have no idea why they happened. Even worse, you are not going to be able to reproduce them in order to fix them. That’s a big part of handling expectations, knowing that there’s a percentage of error that goes beyond control and it improves with time and use. Just like a person, it’s gets better with experience, but you are not going to know exactly how or why.
The importance of the first 10 days
If you have selected the right process to putting your chatbot live, you should be able to see and fix all mayor issues within the first 10 days. On those days I treat our chatbots like babies and monitor every move.
There are different type of issues you are going to find on those first 10 days. There are going to be training issues related to the NLP -because people speak weird-. I have dealt with user inputs that I don’t even know what they meant. That’s part of the learning curve of your chatbot and that’s the most important job you can do on those 10 days to ease the learning process. Once you get a sense of how users really interact with it and how they speak to it, train it like hell so it can be better prepare for the next user.
There’s also going to be cases for which your bot has no answer and you are just realizing that. Skip the temptation of adding new conversation on the fly and prefer a more systematic approach. Register all cases, prioritize them, make sure they don’t overlap with any of the previous knowledge or with each other. Work on a development space and test it a lot, before you add any new knowledge or conversation.
You are always on prototype mode
Mechanic becomes boring after a while. You should always be looking for ways of spicing up your conversation experience. Change or add more responses to the same intent, work on your bot personality, be as creative with emoji and gif as you dare. There’s many things you can do, but it all comes to always work on prototype mode.