When and how to use AI in your RPA processes?
Robotic Process Automation helps organizations automate mindless repetitive tasks, providing great savings, freeing up human capacity to do more added-value work and improving customer experience. However, repetitive tasks are not the end of the line. Hyperautomation proposes the use of Artificial Intelligence within RPA processes in order to increase the level of automation and include also tasks that are not easily definable as simple “if this, then that” rules. Such tasks would include identifying the topic of an e-mail, understanding data from a document or even assessing the validity of an insurance claim.
To bring some clarity to this field of hyperautomation, we have compiled an overview of the options available to help you plan your own AI journey in RPA.
Custom vs. Pre-built models
Building a custom AI model is not an easy endeavor. It requires specialized work and a combination of research and trial & error to find the best algorithm that efficiently solves your problem. If what you are trying to solve is a very specific problem to your business alone, this may very well be the only way.
However, there are many scenarios that apply to almost all industries, such as invoice document extraction or e-mail language identification. There are many pre-built models available on the market that you can use, each optimized for a specific problem.
Here are just a few of the AI solutions providers that you can use in your RPA projects:
Hosting your AI service: SaaS vs. PaaS vs. IaaS vs. On-premises
It is well known that AI models are resource intensive. They require large amounts of data for training & retraining and their best performance is achieved when they run on GPUs. This, along with the standard benefits of the cloud (uptime, reduced maintenance costs, etc.), makes the on-premises version very rarely the best option – we would only recommend it if you have strict data policies that do not allow data processing outside your premises.
Having said that, let’s now talk about the three cloud service categories:
SaaS
Software-as-a-service is the most straight-forward option. It allows you to just call a cloud-hosted pre-configured function with your parameters and get a response back that will be interpreted by your RPA flow. There are thousands of AI services available online, not limited to the big players mentioned above. The most important thing is to find the right model for your problem, but also pay attention to the training & retraining capabilities.
PaaS
With Platform-as-a-service, things start to get complicated. Take Microsoft’s Azure AI platform, for example: you can use the Azure Machine Learning service to manage your end-to-end ML lifecycle, as well as work with Visual Studio Code and popular tools such as PyTorch Enterprise or TensorFlow. You can even use low-code and no-code tools such as the Azure Machine Learning Designer, pictured below.
While going into further details is beyond the scope of this article, it is important to understand the variety and complexity of the tools that are available on the market for your data science team to leverage.
3. IaaS
Infrastructure-as-a-service will help you with one of the problems mentioned above: access to state-of-the-art hardware tailored to your ML projects. Everything that runs on top of this hardware will be your responsibility.
Training & Retraining
In the AI field, the saying goes that “a machine learning model is only as good as the data it is fed”. Some algorithms are already pre-trained and work very well without any custom training or retraining (e.g., language identification models will surely be trained by their SaaS providers, and the effort to retrain them would not really make sense). But there may be cases in which you have the right algorithm, but need to retrain it with your own data. An example that comes to mind would be a demand prediction model for an e-commerce business: the problem is the same, but no two stores would have the same order patterns, so it is important that the model you choose to use is open for retraining.
For example, UiPath offers the Validation Station, a product that both ensures that the result of the Document Understanding algorithm is reviewed by a human if the confidence provided by the algorithm is not high enough, and enables model retraining without the need of any data science knowledge:
Using Validation Station, a user would manually review the results of the document extraction, and any corrections found would be fed again to the ML algorithm for retraining.
The same type of functionality is achievable within the Microsoft universe with Power Apps – see the Invoice Processing Starter Kit for inspiration.
RPA Integration
Finally, to top it all off, let’s not forget where we started: integrating an AI solution into an RPA flow in order to increase the level of automation within a process/organization. Well, the basic & commonly available integration strategy is through web service calls: publish an ML model on a web server, make it available via a web service API, and then consume it from your RPA flow by authenticating and then calling the respective endpoint.
However, it is even easier to integrate UiPath AI Center within a UiPath Robot – you just drag & drop the ML Skill activity, select the model from your AI Center, and enter its API key.
Similarly, when working with Microsoft Power Automate, you can simply use the steps from the AI Builder category. For example, extracting standard entities such as dates & places is as easy as shown below:
Conclusion
So, what is the best approach for you? What this article intends to show is that the answer is “it depends” and, more importantly, to show what it depends on. It all starts with the problem you are trying to solve, as well as the constraints that you have. Your context will determine whether you need to hire a data science team or use one of the over-the-counter specialized ML models available for purchase, whether you acquire just the service in a pay-as-you-go model or operate your own infrastructure.
Need help figuring all this out? Get in touch with us.
