Building an e-commerce bot powered by craft ai

by Perrine Brunel | Dec 16, 2016 | Use Case | AI  Bot  Customer Experience  e-commerce  Hackathon  Learning 

Building an e-commerce bot powered by craft ai

“If you want to talk to me, send me a fax!” - George Abtibol La classe américaine

The story of how we won the Air Liquide “#SparkCX” Hackathon by building an e-commerce bot that knows customer habits: George, the Dentalist

When the Be My App team contacted us to participate in a Air Liquide hackathon we didn’t know what to expect but we trusted their advice and as usual we weren’t disappointed. From November 4 to 6 a craft ai elite team worked on building the best customer experience for dentists.

Our project, George, the Dentalist, was a “mentalist” for dentists: a bot able to take purchase orders of dentistry product and, most importantly, be proactive and recommend new orders when it’s the right time.

Why did we name it “George” ? Because the actual procedure for dentists to order products from Air Liquide is to send a fax. One of us was a big fan of French-language comedy La classe américaine and the fax reminded him one of a line where the hero, George Abitbol (a.k.a. John Wayne), says “If you want to talk to me, send me a fax!”.

It is a prime example of how craft ai can be instrumental in a bot. Let’s take a closer look!

Features and UX

We can’t explain in details how it works as we promised we wouldn’t, but the main idea was to create the most easy-to-use purchase app for dentists. Easy to use, but also easy to deploy.

To achieve this, we decided to create an SMS bot. SMS are simple, and everyone uses them everyday. The conversational format is the easiest way to proceed to avoid the “how to use my app” learning curve, and it’s easy to deploy. The “only” UI to do was to create a back office to register the dentists with the bot and start the process.

This is how George, the Dentalist was born. George is able to take purchase orders from a dentist via SMS, to learn those orders and to propose proactively to repeat the order when the dentist would usually make it.

Technical Architecture

George the dentalist - Software Architecture

The workflow is simple. The first thing to do is to register the dentist with George to initiate a conversation. In a real world scenario this would be done when the dentist is added to the vendor’s CRM, in our “hackathon” world we used a simple HTML form.

Then, the dentist can initiate an order by SMS by sending a list of items and quantities. George will answer by listing what he understands and notifies the dentist when he does not.

To do that, we used three distinct technologies for the three main parts which compose George:

  • Callr - for the SMS part
  • Recast.ai - for the NLP (Natural Language Processing) part
  • craft ai (that’s us!) - for the learning part

For the first part, Callr allows us to create a phone number for George in order to send and receive SMSes. When a message is received, its content is sent to Recast.ai to make sense of it. When the content is an order or part of an order, it is pushed to craft ai.

Thus, we can learn the content of the order and the date at which it was made, to automate the reordering process.

At the end of the month, using what it learned, i.e. the dentist’s decision model, George assembles an order that reflects the dentist’s ordering pattern.

Since our goal is to learn the purchase habits of the dentist, we decided to made a craft ai agent per product and per dentist. By doing this, we learn the frequency of purchase of each product per dentist and also the quantity for each purchase. Plus, since each product is independent from the others, their consumption can vary from a month to another, so with this system we can easily match the requirement of the dentist.

Based on the decision models learned from this agents we are able to suggest a purchase order for a dentist.

After receiving it, the dentist can alter the order based on what he needs, cancel it or confirm it. This information will reinforce the learning of the agent.

Finally when the order has been confirmed, it is transferred to Air Liquide to be treated.

During the hackathon we used data from the current Air Liquide backoffice to learn decisions models from actual dentists. In practice we could tap into the same source to take into account all sales channels.

Conclusion

This was our first public and “finished” experimentation with craft ai inside a bot. This was a great experience and gave us a lot of new ideas on how to use craft ai inside bots and on e-commerce use cases.

We could go further and enhance the decision by providing more context information to craft ai. Using a platform like Doctolib or Mon Docteur we could retrieve the dentist’s appointments; we’d then be able to learn, for example, after how many appointments a new order should be made.

The benefit of using craft ai in a bot is that we learn the habits of the user. So we can fit the expectations of the user and avoid spamming the user with informations that is not relevant.

In this use case, we used George to automate the process of purchase order, in more general terms we made the bot proactively trigger requests. But this is only one of the ways you can use craft ai in bots. For example we can also do some content recommendation based on what the user asks or we can automatically parametrize the requests to the bot so that the user doesn’t need to provide the same information every time; “Which ‘Mary’ do you want to call?”.

Try it yourself: sign up with craft ai at https://beta.craft.ai/signup!

Share