AI in an ERP world: The search for game-changing industry use cases
The AI bandwagon seems overloaded with tech companies, consultants and advisors espousing the great business opportunities that AI will deliver. But while the number of large language models and AI services grows steadily, genuine game changing, value-add use cases remain oddly elusive.
Given the strategic importance of AI, why is this the case? Our team of SAP and AI experts got together to explore this problem with a view to understanding the root cause and to develop a framework to help us – and our customers – harness AI’s revolutionary potential. In the process, we uncovered some surprising (and humbling) truths about the role human intelligence plays in applying AI effectively.
As a leading SAP consultancy, we regularly run customer workshops to identify and develop business use cases in our technology centers of excellence and AI labs. Over time, we’ve observed several common themes that may in part explain why so many of us struggle to define strong AI use cases. We discovered that it is our thinking, and not AI, that’s the problem. Specifically, the traditional “systems thinking” that structures how many of us were trained to approach problem solving tends to fail us when it comes to identifying good AI use cases.
To adopt AI effectively, we need to think differently.
Understanding the limitations of traditional systems thinking
Most business software developed and implemented over the last 30 years is based on structured data, rules and algorithms to process the data, triggering further steps in a logical process flow and storing business data in structured databases. These processes are often very complex, but are ultimately defined by a set of business rules and behaviors coded in the application software.
The generation of software engineers and consultants working with these applications thus inherently think in terms of rules, processes and data structures. I must admit that when trying to apply AI to business problems, I, too, often struggle to shift away from defining a set of rules and data structures. The problem is that AI does not work in this way. AI learns from large data sets of source data and sample outputs, recognizes patterns and develops its own rules to create responses to prompts. The training of what you expect the AI model to produce and a feedback mechanism to modify the result is critical. In true AI, as the AI model learns, it modifies the way it responds and may not always produce the same result.
To better understand how systems thinking affects the application of AI, it helps to revisit some early AI scenarios and reflect on how these have influenced our view of AI in the context of ERP. Early AI use cases largely revolved around chat bots using the AI’s ability to interpret and respond in natural language, with many of us first experiencing through the likes of Siri or Alexa. And guess what? Beyond understanding and processing a request, the actions triggered were often coded actions based on our dear old friends, rules and processes.
I’d now like to discuss some real-life examples from our customer workshops and how they challenged us to develop our framework for developing AI use cases.
Example use case 1: Using AI to “fix” a poorly designed process
The explosive arrival of ChatGPT showed the creative power of AI and its ability to create natural language responses based on learning verges on unnerving. I recall my first experiment with ChatGPT was asking ChatGPT to write lyrics in a specific musical genre. The quality of the output was truly surprising (and should cause many in creative industries to feel distinctly uncomfortable!)
The business world has also tried to harness this creative power. We have seen several use cases emerge which use AI to write human-like responses related to business events. At first glance these appear to be good opportunities for AI. However, if we ask if the responses to these business events need to be read by humans, or whether we would prefer documents to be exchanged using structured messages, the answer generally is that documents exchanged in business processes are far more effective if structured using common, agreed standards or business networks.
For example, we recently had a customer request a POC for automating a returns process where the trigger was an unstructured email from a customer complaining about a product and requesting a return. We developed a solution using ChatGPT to interpret the email and create an appropriate response based on the sentiment of the customer’s email.
On the surface this seems like a good use case for sentiment analysis, using AI to create a sentiment-appropriate response. Look more critically and ask the question “would you design a best practice return process this way?” and the answer is probably “no,” as this approach is more of a tactical sticking plaster for a poorly implemented returns process.
The first question should have been “What would we expect a good returns process to look like?” In a b2b scenario, a return would be triggered through an ERP system linked to the original purchase order and a goods receipt quality inspection. Communication of the response would be via an interface or business network, creating the returns order in the vendors system. Neither party would wish their process to be driven by email. In a b2c scenario, web retailers would expect you to initiate the return via their web shop referencing the original sales order and would expect to issue a returns number for the return shipping and receipt inspection, followed by a credit or replacement. Do any of these need AI to implement? No – these are all easily delivered using traditional approaches.
Our conclusion? Avoid use cases that follow a set of rules or heuristics or are just fixing a flaw in the underlying process.
Example use case 2: Expecting AI to do what it’s not designed to do
A number of ideas generated during our use case ideation workshops gravitated towards predictive capabilities. Why not use AI to look at patterns and predict what will happen next? However, that’s also not really AI. It’s machine learning and statistical forecasting, which has been available in most business applications, including SAP, for some time. Again, the limitation is our ability to conceptualize the possibilities of AI, rather than just trying to find complex problems for it to solve.
This is a good segway into the next theme our workshops uncovered: that many discussions on AI use cases tend to gravitate towards using AI to solve complex problems. Many people assume you can just put AI in front of something that is complex – thinking, surely the AI can learn to fix the complex problem for us – but without any idea how this could be achieved. Again, this is not how AI works.
Fundamentally, the issue is that many people in technology do not really understand AI and how it can be applied to solve business problems or create new business solutions. Likewise, many AI experts don’t understand the world of structured ERP data to which we need to connect AI solutions in order to create a value adding use case.
So how do you define a good use case for AI?
Our guidelines for AI use case development
After reviewing these experiences our team refined our AI ideation process and developed guidelines to help them avoid the aforementioned pitfalls and focus instead on ideas more suited to the application of AI solutions. The list of guidelines is actually quite simple, but has proved useful for eliminating inappropriate use cases early in the ideation process.
A good AI use case must:
- Use genuine AI capabilities to add significant business value
- Have a clearly defined business problem it is trying to solve
- Take unstructured data, interpret it and do something useful with it, or extract information and help create something unstructured from data more efficiently than a human operator could
- Have access to data it can learn from
- Have a feedback mechanism so that the quality of the AI responses can be validated and improved
Following these guiding principles, our teams have since identified and prioritized an inventory of use cases. We then formed an AI innovation squad to develop them by bringing together SAP functional domain experts, technical experts and AI engineers from our Cloud Native AI lab to develop AI extensions to our SAP industry cloud solutions. The result is that the teams can now more easily identify good value-adding use cases, several of which we have committed investment and are currently developing.
Reflecting on our AI journey, I’ve noticed that many people are looking for use cases and problems to solve with AI. A better approach is to recognize that AI is a tool among many other tools that can be applied to solve a business problem, improve business processes or deliver something new. This is why, rather than develop a specific framework for applying AI in business, we have added AI as a capability to our Intelligent Process Engineering framework (IPE).
IPE is a way of approaching process design, to challenge the designers thinking to maximize process integration, automation and autonomy by applying intelligent technology solutions. It is not a prescriptive methodology, rather it is a set of challenges and questions to get the process designers to consider a number of different perspectives and to analyze the problem from a different point of view. This is supported by HCLTech’s IdeaX virtual innovation lab to enable rapid prototyping and learning using the latest SAP technologies for integration, automation and AI solutions. The overall objective is to shift the team’s way of thinking about solving business problems and delivering solutions. This shift in perspective is perhaps the most important lesson to take away from our experience.
In conclusion, embracing the benefits of AI in business applications and processes requires us to leave behind the rules-based thinking we have learned and refined over the years and learn to think in new ways. We shouldn’t underestimate how hard it is to look through fresh eyes, embrace a new approach and see possibilities which are probably beyond our current frame of reference. But for me, the most important first step was to realize the limitations of my thinking based on rules and algorithms, and instead to learn to visualize solutions based on AI.
