AI is the next layer of value for your enterprise

Enterprises are still in the exploratory phase when it comes to adopting Cognitive, Machine Learning and AI capabilities to RPA (mostly due to pre-conceived notions doing rounds – I’ve explained it in detail here). The fact remains that having Cognitive/Intelligent capabilities in your RPA yields bigger and long-term benefits that makes RPA a strategic necessity.

Where do you feel the need?

Typically, standard RPA solutions are used to automated processes that are repetitive and time consuming. But in most cases when it is being done, the process itself may not be structured. What do we do then? The options are either to get the process structured or get the RPA bot to handle the unstructured information. This is probably the first instance where you will find the need for an AI bot. Let me elaborate with an example of invoice processing. If you’re handling semi-structured data like invoices, it comes from different suppliers in different formats. The invoice content/data to be processed remains the same. It will have the invoice date, the invoice number, and the amount. But how will the bot identify the content when the format keeps changing? Have you faced this situation and thought how you’d need a intelligent bot, or a bot based on AI?

Now this logic applies not just for invoice processing, but also pretty much everything that sends data in a semi-structured format – whether it is remittance receipts, BOQs, inventory lists, etc.

This is where the next level of automation – Intelligent RPA comes into play. Here, the bot needs to understand the intent of the email/incoming data and take action based on this. This is a standard use case that that we come across very frequently. Here, you apply an AI engine on top of the NLP engine to drive outcomes through the RPA bot. And yes, in most cases it may start with manual interventions where a person confirms/validates before the AI bot takes any action.

There would be other problem statements that you would have encountered – which you would not have initially realized as an AI problem. For example, if you have to compare two contracts, i.e., when you have to do a pre- and post-contract comparison. A regular text comparison would show you so many differences due to the OCR comparison. But what if you could put in a intelligent bot which could identify the key entities that had changed, and the staff only need to monitor the entities that are highlighted by the bot as relevant to the context?

Exceptions – how do you handle them?

Another major requirement for intelligent automation is when you need to handle exceptions, especially when you’re using a regular bot. Designing a rule-based bot which can handle a large number of exceptions is a difficult activity as every rule needs to be built into the system. This is an important area where intelligent bots could come in handy. Here, the rule-based bot could be designed to handle the easier exceptions, and the complex ones and learned by the bot by observing the human interventions in these exceptions. Of course, this may not handle every manual intervention.

One of the main challenges with implementing cognitive/intelligent bots – is the infrastructure and the gamut of applications required to do this. At bare minimum, you would need a Machine Learning/Deep Learning platform. RPA tools like JiffyRPA comes out of the box with these, but otherwise, you would need to put in effort to integrate RPA into the AI platform which by itself is significant work. A very tight integration will enable you to perform streaming analytics, enabling deep cognitive capabilities on the bot.

Consistency and accuracy

Another challenge is the duration it takes to test an intelligent automation solution. For example, take the aforementioned scenario of invoice processing – how do you ensure that an invoice which is processed today will be processed the same way if another invoice comes from the supplier tomorrow? In such cases you may have to pilot it for a much longer period.

No machine learning model comes with 100% accuracy, while in automation you cannot have even a single incorrect entry into the system. So how do you ensure that errors are minimised as much as possible? One way of ensuring that is to have every decision made by the bot reviewed by a human before an action is taken. But this would result in reduced efficiency. The benefits of automation and processing it straight through goes out the window. Now the next step that we can put in is a complete validation where every field is validated for accuracy. But, will it be possible in every scenario? Finally, what is the scope of testing? In the case of standard rule-based automation you just have to test if the designed rules are working properly or not. How will you test intelligent bots, and test for scenarios that are yet to happen?

Navigating through the challenges .

Hence, when we plan out an intelligent automation approach, it is pertinent that we plan for sufficient duration to run multiple iterations and simulate the production run. The overall effort to validate the results need to be accounted. The validations need to be sufficiently detailed to ensure that no incorrect data goes into the system. Alternately, a user confirmation module could be installed which just reviews every output and confirms before the data foes into the downstream applications. These actions will ensure sufficient accuracy and achieve better outcomes from intelligent automation.

I’ll be explaining this a bit more in my upcoming session during APAC Live, let’s connect there.