In our last post regarding AI & RPA, we delved into the necessary theoretical part, now we’ll be having a look at some more concrete applications for AI in the realm of RPA. We’ll be looking at 3 examples that use chatbots, machine learning, and NLP to supplement RPA and allow it to go past some of the limits that we talked about previously.

One point we would like to get across before we begin is to state explicitly that the aim of RPA shouldn’t be to replace the human user entirely, instead the goal for us is to help users in their daily tasks, making them more effective than they were before. We’re not trying to replace Bob the trader on the trading floor, in fact we endeavor to give him a whole new kit of tools to make him that much more productive.
Even with automation, the employees are key in processes where fuzzy decision points come into play, where variables are just too varied and unstructured to be fully automated. Some things just can’t be decided by robots. That said, let’s look at what is already possible today.

RPA and Chatbots

A simple use case to start. Chatbots are a proven use case, being used most often as customer-facing help agents on websites/apps, or as internal assistants for companies to help their employees. How does that relate to RPA?
In practice, most of the processes that we automate and maintain, are run based on a schedule. This schedule can vary: daily, weekly, monthly, or whenever an input file is placed in a folder. While this variation gives us some degree of flexibility, life would be even easier if the automated process could run whenever the employee needs it.
One way of achieving this is constructing a chatbot that the employee can interact with. By building a predefined exchange scenario, the chatbot can ensure that the user has provided all the necessary information to be able to run the process successfully. This can be done through a combination of trigger words and predefined rules.
Once all the data has been provided, a trigger can be sent to start the robot. The data is passed along, often through mail (with data structured in the text, or as a .csv/.xml/.xlsx -file).
This solution can be easily integrated with current RPA tools and is already in use in various industries and sectors.
Some examples include: verifying a client’s account information when a specific request is made, or creating a new user profile across various applications.

Hacking Humans

One aspect of RPA that increases development time is the time required to analyse all the decision points in a process. It requires thorough preparation from the process owner and the analyst to ensure that all decisions are correctly mapped. This is often made even more difficult by processes that have no definite set of rules, where employees take action based on what they think is the correct action.
With machine learning we can now have AI “monitor” the many click actions and decisions made during these processes. After a sufficiently large dataset has been analysed, this AI will have learned what the normal human decision behavior looks like and is able to replicate it with high accuracy. The process looks like this:
• Monitor human behaviors and human decision making to train a machine learning model (based on Tensorflow).
• The RPA workflow is activated and reaches the decision point, the machine learning system can “predict” the correct decision based on previous choices made by the user performing the same process in the past.
The benefit here goes in two directions;
Firstly, this Machine Learning-driven solution allows a previously not fully-automatable process to be fully automated, only requiring intervention for very specific and rare exceptions.
A second use could be to notify users new to the process when/if they are making a decision that goes against the norm, like a virtual assistant that catches your mistakes.
The catch here is the same catch that stops many good AI use cases from being realised: it requires a dataset that reaches a very high standard in both quality and quantity to be able to train the model.

Support Tickets, NLP and RPA

Here is an all too common scenario: You need to improve your IT user support but unfortunately the ticketing tool is managed by another company which is not necessarily incentivized to help you out. They don’t want to give you access to the API and the backend, so you only have access to the actual text of the support tickets.
RPA can be used to automate applications without access to the API, simulating a user making a lot of copy and paste actions. In this case, we intercept the support requests sent via email from the ticketing tool with a RPA bot, even scraping the required data from the support tool directly if required.
If you do that, you are left with an enormous amount of unstructured, digital data: all the text from your support tickets.
To prioritize incoming tickets, we can use emotional analysis to determine the emotional content of the message, and all the emails/messages sent by the user, to define a psychological profile for that user:
• A scoring system is applied to rank the priority of the request in relation to the operational and hierarchical function of the user. e.g. CEO + Angry = High Priority.
• We write a response adapted to the emotional content of the request with variations in form, so that the user thinks it’s a human who wrote the message.
• We automatically classify the request type based on trigger words in the e-mail, with a subsequent validation by a support expert.
• The support issue is connected to the necessary RPA support process. e.g. Archiving, KDB updates, communications, approvals, ect.
This allows us to prioritize support tickets and reroute them to the relevant RPA processes.
The catch with this use case is the same: the need for a sufficiently large historical dataset of support tickets. If you get 6 support tickets a week, it will take a very long time to train the process to correctly identify the type of support ticket (based on specific trigger words).
That concludes our 3 A.I. use cases.
Next week we’ll be concluding this series of articles with part 3: prerequisites for success.
Thanks for reading!
Authors: Jerome Fortias, Daniel Fastenau