The role artificial intelligence (AI) can play in automating all kinds of processes, including business ones, has been much hyped in recent years. The booming interest in AI stems largely from the fact that the practical business applications of this technology have only emerged relatively recently. Cost-effective computing power and rapid growth in the number of digital data sets have fueled increased use of AI, especially the subfields of machine learning and deep learning. The technical potential available to us today and the speed of progress in the field of AI are hugely impressive.
Critical questions about AI
But AI does not have the answers to all our questions yet — and it is unlikely to anytime soon. Right now, strong AI — a form of intelligence that can behave flexibly, act on its own initiative, and think in a comprehensive way — does not exist. All we have so far is weak AI. In other words, AI applications are only able to perform the tasks they were specifically developed for. Therefore, as an element of automation, we must regard AI as no more than a tool.
Weak AI is not able to think for itself. Certain points should therefore be kept in mind when deploying it:
Legal and ethical considerations: Who is responsible if AI makes a bad decision, for example when a self-driving car causes an accident?
Risk of bias: AI learns from the data it gets, but it never questions this data. If the data sets are skewed in a certain direction, AI will adopt this bias. This can lead to discrimination when AI is used in recruiting processes and data sets are not weighted, for example.
Trust and transparency: AI infers its own decision-making criteria from data sets. This means that AI applications are a black box: it is not clear to users how applications made their decisions. One solution for this trust issue would be to ensure that AI applications just provide suggestions, leaving the actual decisions with users.
Complexity vs. simplicity: In inferring its own decision-making criteria, AI is designed to be used for complex problems — in other words, in situations where decisions cannot be made on the basis of basic if-then rules. For simple problems, however, a predefined set of rules might be a far more efficient solution. This should be borne in mind.
AI in the automation of business processes: good and not-so-good applications
AI can be a tool for automation. But if AI is a hammer, not every business process is a nail. Incoming invoice processing provides a good illustration of how some applications of AI make more sense than others:
Determining the invoice type: The task of deciding whether a document is an FI or MM invoice can be broken down into a single yes/no question: Is there an order number on the invoice? A predefined set of rules is perfect for this task — there would not be much point in deploying AI.
Agent determination: In companies with structures that have developed over years and poorly documented processes and responsibilities, the job of entering into the system which employee is responsible for which documents and tasks can be pretty laborious. Though it is possible to enter the decision-making logic manually, AI could save time and money. What is clear, though, is that when it comes to approving invoices, agent determination involves aspects of compliance. These can be clearly specified in predefined sets of rules. As AI cannot provide 100% certainty, it is important to determine whether full automation based on AI algorithms is even possible.
Document capture: While in most countries the content of incoming invoices has to follow legal specifications, there are no rules on how invoices must look. As a result, they come with an almost infinite range of layouts, though of course some of the differences may be minor. In terms of the time and efficiency required for processing invoices, it is not feasible to train the system with layouts of hundreds or even thousands of supplier invoices. Training documents is an area where AI definitely can help out.
Account assignment: Specifying all rules for account assignment in advance can be extremely laborious — in fact, it might even be impossible to incorporate all conceivable scenarios. This is where a list of suggestions generated by the system comes in useful. This list can be based on AI, or it can simply show previous account assignments — both are helpful. If you are thinking about deploying AI, you should consider what information is needed for correct account assignment. Recurring elements such as G/L accounts or cost centers are better suited to automation than one-off or temporary objects such as SD orders and projects.
Conclusion: Rigorous targeting makes best use of AI
In summary, AI has drawbacks and limitations as well as advantages. If you are considering using AI to automate processes, you should first examine what information is needed, and from what systems. Often the historical data in the ERP systems will not suffice. A great deal of additional information may be required, and not all of this information may be in the requisite digital form.
However, if these limitations are taken into account, AI can be a very useful tool for simplifying and automating workflows. It has the potential to take your automation of document-based processes to the next level.