Building intelligent Slack workflows

Intwixt helps customers design business workflows that can be deployed to messaging apps like Slack or Messenger. They’re simpler to deploy than traditional Web and mobile apps, but new challenges like usability are still being solved.

When creating a conversational UI, it is important to accurately interpret the user’s intent.

Extract behavior patterns from user conversations

Machine Learning (ML) significantly changes the design of intelligent conversational bots. In order to provide the same capabilities without using ML, an application would have to ask the user to answer a long set of questions, which would negatively impact the user experience. Take for example the use case of a food-ordering conversational agent. It is very easy to envision that the service would ask questions like "are you a vegetarian?", "are you vegan?" and tailor the user experience based on those answers. A user of the service would expect to be asked to answer this kind of question. Also, when it comes to the implementation, the service developer can easily model the user profile with properties like "vegetarian" or "vegan". Unlike these facts, which encompass the profile of the user, behavioral patterns are much more difficult to deal with both in terms of user experience and implementation. For example, the fact that Johnny is not vegetarian but he is ordering vegetarian on Tuesdays because he does not eat meat on Tuesdays, or the fact that Mary is not vegan but she orders vegan food every other Saturday when her vegan friend Susan comes for dinner. It is difficult to model these behavior patterns as part of the service. And even though that can be done, it really doesn't make sense for the food ordering service to ask a user to outline and configure all these behavior patterns. It would significantly impact, in a negative way, the user experience.

Machine Learning is the key to dealing with these behavioral patterns. An intelligent service that uses ML can learn from the conversations with Johnny and infer that there is a very high probability that if he orders food on Tuesday, that food is vegetarian and can learn from the conversations with Marry that on Saturday she is likely to order vegan food (the service won't really know that Susan is coming for dinner but that doesn't really matter 😃).

Now, the technical challenge for dealing with these behavioral patterns is that traditionally, ML has been very difficult to use. ML is often referred to as a team sport and certainly, when it comes to very complex ML models, it takes a great effort and a team of people (e.g. data scientists, AI engineers, developers, etc.) to integrate ML into applications. However, does it have to be that hard to build ML models from conversational data and build intelligent, ML enabled bots?

To answer these questions and address the challenges, Intwixt has been searching for a lightweight ML component to add to the interactive workflows.

Earlier this spring, Intwixt requested a demo through our website to see if Aito’s predictive database could be a step in the right direction. Intwixt got access to a free Aito environment and started the experimentation independently. Quickly they noticed that this was the kind of solution they were looking for. They were able to integrate Aito to the Intwixt platform in one day.

Using Aito to make the interactive workflow more intelligent

Contextual awareness is key to solving usability challenges and this is where Aito shines. If, for example, you need to collect multiple fields of information from a user, you can use Aito to predict what the user will most likely say. Instead of prompting the user to enter successive inputs, the app can prompt them once to confirm what they’ll most likely do.

Intwixt users can now choose Aito as part of their interactive workflows. Possible use cases are predictions, smart search or recommendations. When creating a workflow, the UI allows you to create a Data Model, which can be used to create tables in Aito or other databases. This makes it more fluent to manage data, whether it comes from one database or another.

Each time the user confirms the prediction (and even when they don’t), Aito’s APIs make it easy to train the model and increase predictive accuracy. It’s a great enhancement to usability without having to develop and manage complex ML models yourself.

“We value the broad applicability of Aito - a general purpose tool for automation tasks. Aito can organically grow with new content, and new queries and workflows can be iteratively developed,” says Sabin Ielceanu – Co-Founder at Intwixt.

A simple example

The flow begins when a user enters a message in Slack (e.g, "play colors game"). The Aito service is then called to predict which color the user will likely choose. No matter the path taken, the final choice is fed back into Aito for a full feedback loop to improve prediction accuracy over time. It's a simple use case, but it reveals how approachable ML concepts can be when wrapped with familiar data interaction semantics like read and write.

Creating an AI-enabled Slack workflow in Intwixt's UI

More detailed information on how to set up the predictive Slack workflows , including a demo video can be found from Intwixt’s tutorial.

Would you like to try out setting up an AI-enabled, interactive workflow yourself?

Try it out today!