Writing
On autonomous decisions, the AI·IO·ML operating model, the shared world model, and the inversion of the planning function.
In September I called the likelihood of realization my measure of fragility. That was one word doing two jobs. The probability deserves its own name, and fragility belongs to the thing Taleb used it for: how fast a plan gets worse once the error grows.
A solver's guarantee is that the plan conforms to the specification. It holds with full force where the specification has departed from the world. That gap is handed to the modeller, who is given no instrument to measure it.
Every tail product was waved through on marginal cost. Two hundred of them, and the plant was losing money with nothing to point at. A new AI governance paper argues for buying the ability to decide later rather than the decision now; supply chain has had that argument since 1996, and this is why it did not help.
Mathematical optimization cannot invent an answer, but it shares System 1's worst habit: it states one answer with no likelihood of realization, which is how fragile it is. And by the time it arrives, most of the options it was choosing between have gone.
Every large language model is System 1, and Jev is the fastest. Kahneman's whole argument was that the checking layer is the hard part.
Adam DeJans and John Elam's novel puts it plainly: real operations are not solved once, they are decided again and again. Underneath that sentence sit four different planning jobs, four different ways of deciding, why the detailed plan is not required to reconcile to the aggregate one, and one uncomfortable question about your own data. Here is each of them in plain terms, and what we had to build for each.
Almost nothing we call a black swan is one. Kay and King's radical uncertainty names the real category: events you can imagine, may even expect, and still cannot attach a probability to. But being unable to price the event is not the same as being unable to prepare the response, and the second is where the work is.
Enterprise AI's real blocker is not accountability. It is that most organisations have never written down what they optimise for, and the per-decision approval queue is where that unresolved argument gets settled one order at a time. The fix is to govern the yardstick instead of the decision.
Lora Cecere is right that end-to-end supply chain planning is a myth: the only connector between tactical planning and transportation was always order management, and thirty-five years later it still is. The fix is not another integration. It is to stop connecting applications and start connecting decisions, with one vocabulary, one context layer, and a cascade that passes context, guardrails and targets rather than plans.
Markets moved from discretionary to systematic decades ago: every position sized against a calibrated edge, every outcome attributed. Supply chain planning never made that transition, because its unit of work is a cell in a plan and a cell has nowhere to put a probability.
The World Economic Forum just published the blueprint for the AI-first enterprise. It reads like a spec for what we built. Block by block, here is the WEF diagram beside its Autonomy equivalent, in the version where a wrong answer is a stockout.
Google Research recently published evidence that an AI taught to reason under uncertainty generalises better than one taught from perfect-hindsight answers. It is a clean statement of a principle we built our decision agents on: teach the process under the calibrated band, not the answer you only know in retrospect.
A widely-shared technical retrospective makes the honest case for what today's AI cannot do: out-of-distribution reliability, self-explanation, causal reasoning, and long-horizon decisions. In supply chain, those four limits are the spec we built to. Here is the mechanism for each, and the operating model that keeps a human in control.
PwC's 'From Process to Pulse' gets the diagnosis exactly right: the bullwhip is a latency problem. But the cure is not one fast loop. It is the right loops, at the right horizons, each acting on a calibrated band.
Ram M catalogues five ways good employees accidentally weaken their organizations, each with a hidden cost. We hold AI agents to a higher standard than we hold ourselves, and there are four reasons that is exactly right.
Ari Joury's argument is right. Agents collapse the deployment cost. Here is what that looks like inside an industrial supply-chain substrate.
LLMs are still just Prediction Machines. Supply Chain Planning needs Decision Machines.
A new McKinsey piece maps the contours of agentic enterprise transformation with genuine insight. But the destination it describes requires a mathematical foundation the consulting narrative does not provide.
Every general-purpose technology since the printing press has restructured work. The firms that accepted the restructuring early captured the productivity gain. The firms that adopted the new technology in the form factor of the old one became the case studies the next generation reads. Agentic AI is running the same play right now.
Why Berkeley's "next competitive moat" isn't a metaphor, and what it takes to actually fill it.
Most AI-native risk products are solving the wrong problem, extracting a supply-chain graph from text. Autonomy starts from the other direction. The technique stack inverts.
Jordi Visser's thesis on how digital economic activity transitions from human-constrained labor to machine-driven execution, and what it means for planners.
Why your supply chain planners spend 95% of their time not planning, and how BCG's Rules of Response apply to information flow.
A supplier with a 10-day average lead time that varies from 5 to 25 days is not the same as one that consistently delivers in 9-11 days. Distribution fitting changes everything.
The supply chain planning industry is stuck. Here's our diagnosis, our guiding policy, and the coherent actions we're taking to fix it.
Why the most critical layer of supply chain knowledge has never been captured, and what changes now.
