Updated: Dec 22, 2019
The chief supply chain officer (CSCO) of a major retailer was sitting in his weekly operations review meeting. Inventories for certain categories of goods were running dry at some stores while at the same time being overstocked at others. These goods were replenished using some of the most sophisticated software money could buy. The CSCO wanted to know why and how this happened, and he wanted to know now. Various lieutenants weighed in with their diagnoses but it was soon clear that while everyone had an opinion, no one had the data-driven answer. The CSCO concluded the meeting by tapping one of his lieutenants to root out the causes, fix the problem, and report back.
What happened next is all too typical in the world of enterprise software. The software in use at this and other large companies is very sophisticated – it has been configured to handle a complex web of scenarios, policies, and situations. Smart consultants worked with the retailer to configure the software to exactly match the way the business is run. But is the software smart? Is it intelligent? The short answer is no and here’s why.
The leader tapped to fix the problem immediately called the consulting company (a partner to the software company) that implemented the software. The consulting company came in and took a look at the problem and quickly pointed a finger at the software – “the software is screwing up; the replenishment logic doesn’t make any sense.” Soon the CSCO was lobbing calls into the senior executives of the software provider. The software provider now had a mini-crisis on its hands and immediately deployed its best consultants and product development employees onsite to diagnose and fix the problem. After a few hours onsite, the software vendor had the answer – “the software is doing exactly what it’s supposed to do. The problem is you are no longer running your business the way the software thinks you are running your business.”
Why Is This Happening?
This is pretty much the state of enterprise software today, whether it’s ERP, SCM, CRM, or one of the many other acronyms that describe the inner plumbing of enterprises. The software is sophisticated – it can be configured for myriad situations, policies, constraints, scenarios, and multi-variate and cross-relational scenarios. However, it cannot discover new scenarios, it cannot feel what is happening in the real world; it cannot sense that something seems wrong; it cannot intuit that something has changed; simply put, it cannot learn. This problem is particularly acute in the world of supply chains – supply chains are complex and changing all the time, making it difficult to maintain the integrity of a digital twin (the digital model against which decisions are made). There is no internal learning ability to maintain such integrity.
Today, this learning occurs offline by humans. Newly implemented SCM software typically has a reasonable representation of the real world – its digital twin is not an identical twin, but if the data model is robust and it has been properly configured, it’s a good match. However, the minute the software starts to be used, its configuration starts to drift apart from the real world to the point where the digital twin is no longer an identical twin, but more like a distant cousin. In the real world, lead times, production and storage capacities, process capabilities, routes, and even the underlying supply chain structure change (for example, ecommerce customers are asking for much more dynamic source/destination combinations).
Lead times in particular are repeat offenders – they are typically configured as static values and often updated infrequently. In reality, they are dynamic values with changes that can cause significant swings in operating plans. In an increasingly precise omni-channel world where companies are asking suppliers for high adherence to tightening delivery windows, this may cause highly inconsistent results.
Thus, we have the common problem of drift, in which model representations (digital twins) often do not match the world they are supposed to be representing. These same models are being used every day to make critical decisions on production, distribution, and inventory. Software generally does not provide a diagnostic that gauges the health of the digital twin in its representation of the real world. Drift is a concept that has been common in the world of process control for decades. In supply chain software, a common problem is that model representations are deterministic, not stochastic. Deterministic means that most, if not all key input variables are given a single, average value; stochastic accounts for the fact that in the real world, the variables may vary over a certain range of probabilities.
The software company fixed the CSCO’s problem. The CSCO was upset but ultimately thanked the software company’s senior executives and praised the abilities of the people they sent onsite. Internally, the software company associates sent onsite were praised for their dedication and skills. However, in the meantime, everyone in the ecosystem lost – the CSCO’s costs went up, and his operation also negatively impacted revenue at a time when his CEO had been hammering on him to partner with the chief digital officer to figure out how the supply chain can help drive revenue. Not only was cost increased and revenue lost, but the most precious of all supply chain resources was lost – time.
The software company also lost – it had to expend executive time and some of its key subject matter experts had to be diverted from driving revenue and product innovation to fire fighting and fixing problems. At a time when its executives and customers were clamoring for more innovation, some of its best product people were out fixing problems with the existing functionality.
In a world that is exploding with new technologies, a natural question to ask is “why can’t the software figure this stuff out for itself?” Why can’t it detect the drift and head off problems at the pass? The experts at Kinaxis have examined this problem and have employed machine learning algorithms to create a learning loop in their software. The algorithms monitor values from the real world and recognize drift and then provide the ability for the model to “self-heal,” leading to what they call the “self-healing supply chain.” Today the healing occurs through expensive human intervention, as shown in the real-life retailer scenario discussed earlier. Worse yet, the lag time between when the supply chain starts to take on the problem and when it is healed can be significant – it could be weeks, months, or in some cases years. Like a chronic illness, it then gets solved by throwing out the existing software solution and starting over, in the hope the illness will go away. Unfortunately, all too often, this just results in the same process starting over again.
The self-healing concept is powerful and has implications for all supply chains. If software can detect drift and then do something about it, all members of the ecosystem benefit, most all end-user companies and their customers. Even if the software does not automatically make changes to correct the drift, it can minimally raise it as an alarm to humans. This alarm can include recommended changes to the configuration of the software. Humans can then review the system-recommended changes and take them through an approval process. In the meantime, there will be no lost sales, increased costs, and all manner of related waste for all ecosystem partners.
Subsequent to the original post of this article, top analyst and longtime colleague Steve Banker profiled how Manhattan Associates is using similar techniques to keep its WMS better synchronized with actual labor and task times. This offers a second validation point for the Kinaxis self-healing concept.
The historical consulting-partner and software-company state-of-the art solution to this problem has been to offer ongoing paid-for “managed services,” typically long-term contracts that provide services intended to help you continuously derive “entitlement value” from your software asset. This is over and above the maintenance, support, or subscription contract you also pay. In self-healing lingo, this is tantamount to having a team of doctors on retainer full-time. This is expensive; furthermore, these costs are typically not comprehended in the original fully-loaded cost analysis for the original investment and use of the software.
While much of AI investment in supply chain management is going towards the digitization of labor, demand analysis, and procurement analysis, this is a rich value area of opportunity for AI, and one which has the potential to revolutionize enterprise software.