Artificial Intelligence & Robotic Process Automation - FAQs

Context

I posted a few articles earlier attempting to answer general questions about Artificial Intelligence & Machine Learning. I received several follow-up questions related to the application of Artificial Intelligence in Robotic Process Automation.

I answered a couple of questions in an earlier article, but due to the volume of follow-up questions, I am posting a separate article below.

What is the difference between Robotic Process Automation & Intelligent Process Automation? 

Robotic Process Automation

Robotic process automation (RPA) replicates repeatable & everyday human actions associated with business processes in an organization. The software ‘robots’ automate predictable and repeatable tasks. These ‘robots’ sometimes referred as ‘bots’ have the potential to reduce costs, improve process quality and consistency, and enable greater scalability. RPA uses a combination of process mapping & workflow to record the functions you are trying to automate. Behind this are a set of & improved ‘screen scraping,’ technology and macros that sits on top of your enterprise applications and automates the keystrokes of an employee.

In other words, RPA solutions are considered an improvement over traditional ‘screen scraping’ technologies and it ‘records’ / ‘replays’ the workflows of a human operator. RPA is not artificially intelligent and is primarily used to automate repeatable tasks. RPA uses a set of macros, scripts and screen-scraping (record and replay) technologies, records what a human operator does by capturing keystrokes and repeats those workflows whenever a new ‘transaction’ comes in. The system can work without a human operator or can assist a human operator to complete tasks quicker.

Intelligent Process Automation (aka Cognitive Process Automation)

Intelligent Process Automation (IPA) is the use of Artificial Intelligence in businesses processes that are routine & repeatable but requires human intervention, intelligence and inputs.

IPA uses a combination of Computer Vision, Natural Language Processing, Advanced Pattern recognition through machine learning, ‘Next best path/action’ using advanced deep learning algorithms. It uses human-like knowledge representation, perception, learning, reasoning, problem-solving to the business problems. Also, it creates a cognitive network (similar to the human brain) from different areas of business and the combined intelligence from different process areas can solve complex, time-consuming problems for organizations.

There are several levels that IPA can operate. It can operate at the lowest process level (4 or 5) similar to processing inbound documents (invoices) or it can operate at a higher level optimizing & automating an entire L2 or L3 business process (managing to a higher level KPI). The organization needs to consider the change impacts before deciding on the level of change they want to introduce through IPA. 

Can you provide an example of Intelligent Process Automation?

Account Payable (AP) is one of the critical sub-processes within the broader context of Finance & Accounting in an organization. An Intelligent Automation system can be implemented at AP process level, or a lower sub-process level (e.g., processing inbound invoices & bills), or cross-functional (managed at a cross-functional KPI level) or at a higher level. An IPA system learns from past experiences before the implementation and will continue learning with future experiences. The system is not coded explicitly with ‘rules.’ Instead, it is given a set of goals and it will find the optimum path to get there.

Typically in an organization, the mission of AP department is to “pay only the company's bills and invoices that are legitimate and accurate.” The AP team need to compare what company ordered, what was received, whether the proper unit costs, calculations, totals, terms, etc. Also, they should prevent paying a fraudulent, inaccurate, duplicate invoice. They have to be careful that they pay invoices on time, and the unpaid invoices are appropriately accounted for, etc.

A Level 3 IPA system can be built, and it should be able to do the following functions (Level 4 Processes) at a minimum

• Intelligently process the extraction of data from the invoices, which are often in different formats (e-invoices and non-electronic). The system can locate and extract dozens of data elements that often span multiple pages, including invoice number, vendor name, invoice date, P.O. number, addresses, amounts, currency type and line items details. The system will be able to process invoice formats that it has never seen before, and it will be able to ‘auto-correct’ common clerical errors.
• The system can then compare the orders, ‘shipment received/services completed ’ information, units, costs, etc. (similar to humans)., auto correct cross-functional issues, etc. without hard-coding ‘rules,’ using learning experience.
• The system can Identify risks (Supplier, Financial, etc.) and act on ‘risks’ with decidedly less human intervention
• The system will be able to self-audit and also assist in Internal Audit.

The real value of Intelligent Automation is not just to manage a Level 3 business process, but to manage a higher-level KPI. For example, the system can be built to manage the KPIs like ‘Material Cost Variance to Target’ or optimize ‘Free Cash Flow’ to a certain predefined metric. The system can find the best path to get to the target metric through a combination of machine learning algorithms. The most commonly used method is called as 'Deep Reinforcement Learning'.

Can you provide some technical details on how an IPA system work?

Examples of certain parts of an IPA system used in Account Payable area are given below.

Reading and ingesting invoices (a Level 4 process in AP)

The invoices that an organization receives on a daily basis in various shapes and forms, and sometimes with information handwritten on it. Many companies used OCR (Optical Character Recognition) Systems to process invoices, but there are several constraints on using it. It just read the characters and word but does not classify them, interpret them, learn from them and use it appropriately like a human reads a document.

Deep Learning architecture using CNN (convolutional network) or the enhanced version of CNNs like real-time object detection architecture like YOLO can be used to identify objects and within an invoice and also read items within an object like vendor name, invoice date, P.O. number, addresses, amounts, currency type and line items details, signature, information scribbled on the invoice by humans. An example of such a YOLO architecture network is provided below. In this example the system identify different objects from a single image. In case of invoices the system identify the different objects (sections) of the invoices and then reads what is written/printed.

A sample architecture of a YOLO network is provided below

The ‘YOLO network’ is designed to read real-time data from cameras and interpret what the objects are and act on it. A production quality AP system deployed by a company can recognize/ understand invoice objects and read the actual data from an invoice can be extended easily to process any documents for the company. This same system can use to detect frauds, anomalies, duplicates and classify them appropriately. Such a system can perform as good or even better than humans can.

If there are handwritten notes or scribbling in the invoices, those are processed through an Natural Language Processing (NLP) subsystem. This also can be combined with voice inputs /instructions from humans as well. For example a CFO or a controller can call into the system and say 'Please delay the processing of invoices from XYZ company for next 24 hours' and the system will be able to react to that without human intervention.

Once the information about an invoice has been gathered & classified, the details are passed to another AI subsystem that is used to do three-way matching, making adjustment entries, etc. This is done through a Deep Reinforcement Network architecture. This approach will allow the system to decide what best to do best with the invoice. The organization provide a higher level KPIs and let the system finds it ways on what is the 'best path' to take.

Reinforcement learning can be used in successfully in situations approaching real-world complexity. The system will derive efficient representations of the environment from high-dimensional sensory inputs, and use these to generalize experience to new situations.

Reinforcement learning (RL) is about an agent interacting with the environment, learning an optimal policy, by trial and error, for sequential decision-making problems. An example of such a learning system is provided below

A RL agent interacts with an environment over time. At each time step t, the agent receives a state in a state space and selects an action at from an action space, following a policy, which is the agent’s behavior, i.e., a mapping from state to actions, receives a scalar reward, and transitions to the next state, according to the environment dynamics, or model, for reward function and state transition probability respectively. A value function is a prediction of the expected, accumulative, discounted, future reward, measuring how good each state, or state-action pair, is. When a RL problem satisfies the Markov property, i.e., the future depends only on the current state and action, but not on the past, it is formulated as a Markov Decision Process (MDP), defined by the 5-tuple.

If the above description is difficult to understand please do not worry. You can look up any literature about Deep Reinforcement Learning. In addition I will write a separate article on applying Deep Reinforcement Learning to manage higher level Financial KPIs & Risks for an organization.

Can you provide a few sector-specific examples of Intelligent Process Automation?

Generally speaking several back-office processes like Account receivables, Purchase Order Generation, Costing functions, Expense reimbursement, etc. that can be intelligently automated. This can be done with build in risk management functions and internal audit functions so that the system takes an optimal path. Also several customer-facing functions like customer service, order taking, etc. can be intelligently automated.

I will write sector specific articles on Intelligent Process Automation in Retail, Financial Services & Manufacturing, etc. since there were several questions related to them.

Are there off-the-shelf software products that are available to support Intelligent Process Automation?

Several products are available in the market to support Robotic Process Automation for organizations, but most of them are not intelligent. Intelligent Process Automation can be build using commercial and open source AI platforms & frameworks.

Also, IPA systems will provide organizations a competitive edge, and it is recommended that they are build using the domain knowledge/intelligence and experience acquired by your organization over the years. You can also define the path the organization would like to take in the future, and you can specify the KPIs that are important for the organization. A generic off-the-shelf tool will not provide you those capabilities.

Conclusion

If there are additional questions or comments please send to [email protected]