Information loss in Supply Chains is crippling. Can AI, Gen AI help?
If Supply Chain systems and processes are re-written grounds up today with first principles using latest tech, like AI, how do they look like? Incremental AI lipstick on a legacy pig or more?
Summary: Prevalent processes and metrics cannot see the rich information in supply chains - hence the blindness of everything that relies on them. including AI. If Supply chain systems and processes are re-written grounds up today with First Principles making use of latest tech like AI, how do they look like? Will that be just adding incremental AI lipstick on a legacy systems pig or completely different? A fresh view of what needs to be solved and why.
This is a 3-part series. Part 1: Deep dive into what needs to be solved for, so that we know ‘what is the business need?’. Part 2: First principles view of supply chain solutions. Part 3: Solution approach that leads to AI/ Gen AI more than lipstick.
Part 1: What needs to be solved, and why?
Supply chains, the real-life, physical supply chains, are a web of several aspects with a series of causes and effects working together to drive business outcomes. They are complex systems. Simplifying this complex causality web into singular, linear aspects – inventory optimization, delivery performance – has two devastating effects: loss of information (rich contextual data) and being disconnected from business objectives (e.g., cash flow).
Supply Chain decisions and actions are results of situations that lead up to the moment of decision and action. Current systems, processes, KPIs are designed for point-in-time decisions and actions, disconnected from leading scenarios. (Elaboration with examples of Causality, Context in Business and Supply Chain here).
These processes and KPIs have existed since 1970s, defined in line with the-then available tech and systems, as very simplistic representations of eventuality, let alone causality. They are comically inadequate, even in simply representing the real-world supply chains, let alone solving for business challenges. Here are the proofs!
1. KPIs, Metrics, processes have not changed for 5 decades. If anything that uses data and tech has not changed for 30 years, something is wrong. One look at the formulae shows how bare minimum these metrics are, and how hopelessly reductive they are in representing reality. Min-Max, ABC, OTIF, MoQ etc, which are extremely simplistic representations of reality. Good folks at Lokad, Joannes, Alexey have been doing excellent, painstaking service of explaining these with nice comics here.
2. If that’s not convincing enough, this should: the founder of Operations Research in supply chain, R.L. Ackoff, lamented how (and importantly, why) the application of OR descended into caricatures of real-life supply chain, long ago. The Future of OR is Past, F/Laws).
3. Making expensive business and operational decisions based on these simplifcations leads to these challenges:
Caricaturing reality: Supply chains are complex in real life and represent intricate system and business relationships. Modelling this rich supply chain context (background information) with very simple metrics misses out capturing valuable information that will be useful when plans change and things go wrong, which is not that infrequent. If this context is not captured well and in time, any talks of resilience, Generative AI cannot be realized.
Misalignment with business: Supply chain is not the end, business outcomes enabled by supply chain are. High inventory accuracy at a factory warehouse while production suffers from part shortages; solving for OTIF but sales lost due to stock-outs are a few but very common examples of this expensive misalignment. This is called ‘solving for proxies’ (optimizing for narrow supply chain metrics, the proxies, instead of solving for business outcomes). Isolated inventory optimization, even using the best of AI, is not of use if business gaps are not closed.
These two aspects 1/ grossly inadequate modelling with stick figures and pretending it is a reasonable representation of reality, and 2/using narrow, siloed metrics, measures, processes disconnected with business objectives, form the focal challenge. This goes against the common supply chain rhetoric - end to end visibility, control tower, resilience etc, so let’s spend a minute to understand why this is the challenge to focus on and solve for.
Why is this the focal challenge? Do we really need to solve this?
An important question to ask is: Are we capturing the reality as it exists, even with some resemblance of reality, in the current practices? Are we capturing data and relationships sufficiently to enable decision making?
Supply Chains are complex and beautiful (yes! talk to me for examples), in the sense that the million parts miraculously work together and deliver business outcomes, despite the disjunct mess they operate in. In the current world of SC this complexity is reduced to a cartoon version, which then is used to make business decisions. Imagine Mona Lisa (or Pieta for me, any day) represented by a cartoon, how much ever great the cartoonist is! It’s not just that a lot of detail will be missing, but the essence of what is being represented will be missed out forever. Effectively modelling the reality and capturing related data and relationships is of utmost importance.
Since current processes and metrics capture a small % of the reality, data and decision alternatives required to act when something changes is limited by what is captured / measured. There goes your promise of resilience. In real life, plans always change, ports get congested, suppliers don’t speak gospel all the time, vehicles are not placed as planned, delays, forklift issues, congestion, workers are not available and many more, while a day of delay can result in $ millions in cost of production delays for customers. We cannot pretend with oversimplification and not capturing all data and relationships in that process.
That is why this is the focal challenge to solve for. Here is a real-life example of modelling realistically vs over-simplification with the resulting loss of data, information forever:
Scenario: A B2B manufacturer of industrial components, Tier -1/N.
Business Objective: Improve / respect Customer Service Level while balancing impact on production schedules and impact on sales (revenue).
Note that business objective here is not a singular measure (service level) but balances business with operational – as this is how the real life is. If one wants to take singular, simplistic view to ‘optimise inventory’ without asking ‘for what business outcome?’, this paper is not relevant.
Supply Chain Task to achieve the above: Pick -> Pack -> Load -> Dispatch to align with the stated business objective.
In real world, above business objective is a function of pickers, forklifts, layout, inventory, picking policies, vehicles, vendors, orders, docks, yard, plans (vehicle load, dock door etc), available staff, dispatch, customer details, service levels, tolerance as per contractual service levels, margin, and cost (to serve), with related systems like WMS, TMS etc. even in a very simplistic world.
In a world which aspires to be a bit more digital, it will also have RTLS, Computer Vision or some way of sensing and interacting with physicality, picking robots / AGV / SGV, more evolved (optimized) picking policies for hybrid environments, forklift routing, safety systems, docking sequence, and a few more systems marrying the old with the new.
This simple, everyday scenario, even without advanced tech, has 100s of parameters and much larger number of relationships. And these relationships and their statuses change frequently. Plus, it gets more complex with time, with new systems, data sources, more integrations, errors, corrections etc. We are not considering the scenarios leading up to the present moment of decision / action. Still, relevant factors at this moment run up to 100s.
But prevalent approaches today, even the best of the systems, hardly use 5-6 parameters from the above, to optimise, only the operational aspects (i.e, no business objective alignment) and even for this, will require good time to integrate and manage data, only to break them when new systems / data sources are introduced after a few months.
Because so many parameters and relationships are missed out, there is no path to recovery if something doesn’t go as planned, as the information related to how they are connected, mapped, correlated for the correct context and causality is lost. This is because of the practice of linear flow (i.e. plan, then execute, mostly by two different systems, without a closed, timely feedback loop, not even a possibility of a path to closing the loop, as all relationships are not captured).
So here is the summary of the challenge: being able to capture the supply chains in a realistic way, at least partially where it looks closer to reality than simple stick figures. This ensures that relevant context – data, relationships, and related changes are captured.
Note that the focus is only on modelling the reality, to ensure relevant context – data, relationships, changes, and pace of changes are captured. This by itself will not solve supply chain challenges but this is the first step that enables more effective solutions, as the base substrate, the context, is retained. Without solving this, any AI/Gen AI, resilience etc talks are futile.
So, what is needed to solve this focal challenge?
Without jumping to a technology, or buzz word or a vendor solution (first principles, remember?) let us describe what the solution must do to solve the focal challenge.
Solution Objectives:
Supply Chains are rich in contextual information, as supply chains are amidst a lot of moving things – physical and notional, internal and external to an organization. Context is everything in being meaningfully responsive. But the context is also dynamic, volatile and keeps changing (in real, physical world).
A true representation of supply chain reality is possible when at least a part of this dynamic context (existing relationships) is captured, and at least a part of the dynamism (changing relationships) itself is captured. Modelling such system requires capturing and amending the context dynamically and in sync with the pace of change.
Formal statement of solution objectives:
Model a semblance of reality, which has, in the least, below characteristics:
Functional:
Model and capture relationships: Model relationships across 100s of converging aspects of that represent physical and business reality. Manage many integrations, with accurate lineage and provenance in rapid, flexible, convergent context. Rapidity is important not to spend months running integration projects.
Model and capture causal factors: a way to record casual factors, weightage to factors based on intensity of influence, and chain of causation (series of causes and effects – a system for capturing this). It is not required to know all causes and effects, but leave space/mechanism to capture causal factors yet to be discovered.
Model and capture evolving environments: Ability to rapidly capture dynamic context - the pace of change of business activity, new causal factors, change in weightage etc., especially where discontinuity of context exists (silos - in systems, processes, measures etc). Hard integrations cannot handle new additions every few months, which will be the reality in industrial environments – Lidar, CV, Edge IoT. Even after integration, causal factors change dynamically, which needs to be accommodated.
Technical:
Relate data across relational data models (a large portion of enterprise data resides here) and other data model types seamlessly.
Dynamic data model to add and relate any data source ad-hoc and continuously and composable with existing data sources.
Naturally aligns with ML – normalized, denormalized data frames, data sanctity and lineage reliability.
The focus is on relationships and the change in relationships across data, not just data itself. This model (structure) holds the context - space in terms of ‘relevant data’, to rapidly merge and relate 100s of data sources, in the context of business parameters and allow this structure to change frequently in-line with the changes in physical reality. Let’s call this Context Fabric.
It is fair to say that, if this is solved, in the way described above, we will have a base substrate to hold relevant context that enables more effective solutions, at least we will have captured the context, relevant information that is handy that contains a problem space and potentially a solution space (when things go wrong) in terms of data and relationships. This does not sound achievable with existing data-tech but Math based solutions exist.
Contrast this with how the context is captured today - simple metrics, simple ‘business process’ definitions, rigid that cannot be changed once defined without massive integration projects etc.
Once this is in place, the Context Fabric contains the data required to train and deploy AI models or simply look for good alternatives in the decision space, go back or forward in time, besides of course the simplistic KPIs/ Metrics reports, covered in Part 2.
Conclusion: Prevalent approaches to model supply chains, even the most evolved (e.g., knowledge graph) are poor, rigid stick-figures of reality. It was fine 30 years ago, but continuing with the same now, when enough compute and data tools to represent complex reality exist, doesn’t help.
Grossly inadequate stick figures to represent the reality result in simple metrics and processes that miss out 99% of context. No matter how hard a vendor sells a ‘resilient supply chain solution’, unless this, reliable modelling, and data capture challenge, is solved for, supply chains cannot be responsive.
A true model must capture sufficient data, in the context of business relationships mentioned above (as defined by the organization). Modelling such system requires capturing and amending the context dynamically and in sync with the pace of change.
In the traditional approaches, one can achieve the above objectives partially, without all the flexibility and dynamism, with a series of lengthy integration projects.
PS: All views are personal. Protected by Intellectual Property Rights, owned by the author. Patent applied for.