The other side of the textbook: who built these ideas and whether they still hold, OR vs AI, the modern stack (digital twins, deep-RL, causal & agentic AI), the real movers & shakers, the failure rate, and resilience vs geopolitics.
August 2026 · Case 6 of 6
As you read — hold this question
The textbook stops at the SCOR model and the bullwhip effect. But who built those ideas, do they still hold — and if the big shipping lines already run world-class optimisation, what does AI actually add?
~1961
the year Jay Forrester first modelled the bullwhip effect at MIT — the “new” supply-chain ideas in your set reading are mostly 50–60 years old.
~95%
of enterprise generative-AI pilots delivered no measurable P&L impact (MIT, 2025). The honest frontier is as much about what fails as what works.
Provenance
The rest of the story
Who built these ideas — and do they still hold?
The set reading presents the SCOR model, the bullwhip effect and the economic order quantity as timeless furniture. They are not timeless — they were invented, by named people, to solve specific problems, and each has a shelf-life worth knowing.
Bullwhip — Forrester, 1961
Jay Forrester modelled demand amplification at MIT (Industrial Dynamics); Hau Lee formalised the four causes in 1997. Still holds — and AI-driven demand sensing is the modern attempt to dampen it, not abolish it.
SCOR — 1996, now digital
Built by the Supply-Chain Council (now ASCM). The 2022 SCOR Digital Standard added a seventh process, Orchestrate — evidence the framework itself moved on from the version in the slides.
EOQ — Ford Harris, 1913
The economic-order-quantity formula is over a century old and assumes steady, known demand — exactly the assumption modern volatile, multi-echelon chains break. It survives as a teaching tool, not a planning engine.
Takeaway: these are durable ideas with dated mechanics. AI doesn’t replace them — it re-implements their intent (smooth the swing, plan the order) with data the original authors never had.
The honest comparison
The other sideQuiz: OR vs AI
The big lines already optimise. So what does AI add?
Fair challenge: Maersk, UPS and DHL have run operations research — mathematical optimisation — for decades. UPS’s ORION routing system was saving a reported $300–400M a year by 2016, long before the current AI wave. So the honest answer isn’t “AI replaces that.” It’s that AI and OR do different jobs, and the frontier is stitching them together.
What OR does well
Given clean inputs and hard constraints, it finds a provably good plan — routes, schedules, order quantities. Deterministic, auditable, mature.
What AI adds
It supplies the inputs OR assumes: better demand forecasts, disruption signals, image and text understanding. It handles uncertainty and messy data OR can’t ingest.
The frontier framing is “predict, then optimise” (Elmachtoub & Grigas, Management Science, 2022): let machine learning predict the uncertain future, then hand a clean forecast to the optimiser to make the decision. The two are complements, not rivals — and a firm already good at OR is better positioned to benefit, not obsolete.
The other side
This means the sophisticated incumbents may gain least in percentage terms — they already picked the optimisation fruit. The dramatic AI wins are often at firms starting from spreadsheets. “AI transformed logistics” is truer for the laggards than the leaders.
The modern stack
The rest of the story
The toolkit the textbook skips
Beyond “machine learning” and “generative AI,” here is the real vocabulary practitioners use in 2026 — flagged for what’s proven versus genuinely frontier.
Method
What it does — and how real
Digital twins
A live simulation of the physical chain you can stress-test before acting. Proven — used by DHL, Maersk; NVIDIA Omniverse is a common substrate.
Deep reinforcement learning
Learns inventory / routing policies by trial in simulation. Proven at scale — Amazon’s SCOT reported ~10–12% inventory reduction (arXiv, 2022).
Agentic AI
LLM-driven agents that plan and act across systems (re-order, re-route, negotiate). Frontier — heavily marketed in 2026, thinly proven.
Causal AI
Models cause, not just correlation — “will this promo cause a stockout?” Emerging; strong in theory, early in practice.
Graph neural networks
Reasons over the supplier network as a graph to find hidden dependencies and risk. Emerging in risk and multi-tier visibility.
Foundation models for time series
Pre-trained forecasters (e.g. lineage of Amazon Chronos) applied to demand. Frontier — promising, not yet a default.
Read the tags: digital twins and deep-RL have receipts. Agentic and foundation-model claims are where the hype outruns the evidence — ask for the P&L number.
The movers & shakers
The rest of the story
Who is actually leading this
The set reading names almost no companies. Here is who practitioners actually watch — the platforms, the operators, and the picks-and-shovels.
Planning platforms
Kinaxis, Blue Yonder and o9 Solutions — the “concurrent planning” and demand-sensing engines large enterprises actually buy.
Operators at scale
Amazon SCOT (deep-RL inventory) and Walmart (Element, Pactum) — the firms proving methods in production, not slideware.
Data & visibility
Palantir Foundry for the data backbone; project44 for real-time transportation visibility — the unglamorous layer everything else needs.
Physical & picks-and-shovels
Symbotic (warehouse robotics) and NVIDIA (the simulation and compute substrate) — the “physical AI” end of the stack.
The reality check
The other side
The experts — and the failure rate
The strongest voices in this field are notably un-breathless. The honest headline is that most AI initiatives struggle before they succeed, and the reason is almost never the algorithm.
~30%
of AI projects abandoned after proof-of-concept (Gartner, 2024–25).
#1
blocker is data quality & integration — not the model (McKinsey, 2024–26).
“AI is an augmenting force — not a replacement force.”
— Prof. Yossi Sheffi, MIT Center for Transportation & Logistics (2023)
The uncomfortable pattern
Vendors quote the wins; the peer-reviewed and analyst numbers quote the failures. Both are true. A mature reader holds them together: AI in supply chains is real and useful and most deployments underdeliver first — usually because the data plumbing wasn’t ready.
Where it’s heading
The other side
Resilience, geopolitics, and the rules catching up
The last decade rewrote the goal. Ever Given wedged in the Suez Canal, COVID, and the chip shortage taught every board that a chain tuned only for efficiency is brittle. The new mandate is resilience — and the two pull against each other.
Efficiency
Lean, just-in-time, single lowest-cost supplier. Cheapest when nothing goes wrong — and everything went wrong.
Resilience
Buffers, backup suppliers, near-shoring, visibility. AI can sense disruption earlier and re-plan faster — but it can’t manufacture slack that management chose not to fund.
And the rules are catching up: the EU AI Act now classes some automated decision systems as high-risk, adding transparency and oversight duties to the very negotiation and pricing bots the set case celebrates. Governance is no longer optional.
Where we land: AI doesn’t choose between efficiency and resilience — leaders do. AI just makes the trade-off visible, and executes whichever one you were brave enough to pay for.
Quick recall — without looking back
Test yourself on this case
Question 1 of 3
The set reading’s core ideas aren’t new — give provenance for the bullwhip effect and say what AI does about it.
Jay Forrester modelled it at MIT in 1961; Hau Lee formalised the four causes in 1997. AI-driven demand sensing tries to dampen the amplification — it doesn’t abolish it.
Question 2 of 3
If the big carriers already run world-class optimisation (OR), what does AI actually add — and what’s the frontier framing?
AI supplies the inputs OR assumes — forecasts, disruption signals, messy data — and handles uncertainty. The framing is “predict, then optimise” (Elmachtoub & Grigas, 2022): ML predicts, OR decides. Complements, not rivals.
Question 3 of 3
State the honest reality on AI success rates and the single biggest blocker.
Most initiatives underdeliver first — ~30% abandoned after PoC (Gartner), ~95% of gen-AI pilots with no P&L impact (MIT, 2025). The #1 blocker is data quality and integration, not the model (McKinsey). As Sheffi puts it, AI is an augmenting force, not a replacement one.
Module 8 · Long Form · What AI Actually Does in Supply Chains
Module 8 · Short · What's Real, What's Hype
Sources
Provenance
Forrester, J. (1961) Industrial Dynamics, MIT Press; Lee, Padmanabhan & Whang (1997) “The Bullwhip Effect in Supply Chains,” Sloan Management Review. SCOR Digital Standard, ASCM (2022). Harris, F. W. (1913), EOQ.
OR vs AI
UPS ORION savings — INFORMS (2016). Elmachtoub, A. & Grigas, P. (2022) “Smart Predict, then Optimize,” Management Science.
Gartner (2024–25) AI project abandonment; McKinsey State of AI (2024–26); MIT / NANDA gen-AI pilot study (2025). Prof. Yossi Sheffi, MIT CTL (2023). EU AI Act (2024).
Innovare Study
Long-form video: What AI Actually Does in Supply Chains. 2026. youtu.be/HlD8PWRNDiM