For about 3 months we've been working on getting our new chatbot ready for launch. The time is finally now.
Feedback from offline users
To ensure that the chatbot has a natural place in the user journey, we held a user-lab where different users were given the possibility to give feedback, ideas, and criticism. We had a setup of a phone with the web-app and a laptop with the desktop version. The feedback was added live to our Trello board, thereby giving the development team the possibility to catch and fix the critical errors right away.
The user-lab provided us with lots of information on how users communicate with a chatbot. We also used the session to test out a few different concepts and designs we wanted feedback on.
A point of improvement for us to keep in mind when arranging the next user-lab is to be able to recruit users matching our personas even better.
Launch in a controlled environment
When the code needed for version 1.0 was written, we decided to soft launch the web-app in Danish to become more aware of where the conversations would be heading. Predicting how a conversation goes when designing the initial dialogues is very difficult. That is why we decided to launch in Danish first and using the findings to improve the English version while training Watson.
We recruited users on a few different forums to help trying out ForkChoice and thereby generating data for Watson to use.
Testing never ceases
Because the dialogue keeps improving and the use-cases vary from user to user, the testing is a key part of release and post release. We have allocated resources to babysit the chatbot for the first few weeks since release, This has proven well spend, based on the number of scenarios we did encounter (and still do). Luckily the training in Watson and the changes to the dialogue is done in a fairly smooth environment in the Bluemix console. But the key takeaway from this process must be to spend time on documenting the whole dialogue when designing it. It will make the whole training/improvement process much smoother.
Are we there yet?
Not even close. We still need to train ForkChoice to be able to provide a more crisp dialogue and communication with the user.
Other than that, we still have a few different API's we want to integrate thereby offering, even more, encouragement for the users to integrate ForkChoice into their routine when going out for food or drinks.
Remember to visit ForkChoice and chat your way to new taste experiences.