Past two weeks were full of interesting news and insights related to AI and technology relevant of physical operations, let me reflect on just a few to provoke thinking and reflections.
0️⃣ McKinsey: CEOs must build an AI Assembly Line - or risk being industrialised out of relevance.
McKinsey's latest report for industrial CEOs opens with a provocation: the companies that win the AI era will not be those that use AI as a tool - they will be those that treat AI agent deployment itself as a production system.
Implications for supply chains. The assembly line framing is deliberate and worth sitting with. McKinsey is not talking about AI as a productivity tool or a cost-reduction initiative - they are describing a structural shift in how industrial organisations create and deliver value, analogous to the introduction of the production line itself. For supply chain leaders, the practical implication is that the question is no longer whether to deploy AI agents in operations, but whether to build the organisational infrastructure to do it at scale and with discipline.
But it also comes with some risks when wrong KPIs and expectations are set across the organisation. Lets have a look at another insight, my favorite in this issue.
1️⃣ Amazon workers are gaming AI metrics by creating useless agents as reported by FastCompany.
Amazon is tracking how many AI tokens its employees consume each week, reportedly targeting 80% of developers using AI tools, with consumption visible on internal dashboards. The result is predictable: employees are creating pointless AI agents - not to get work done, but to drive up their token count.
Implications for supply chains. This story is the direct cautionary footnote to McKinsey's assembly line thesis. McKinsey calls for an agent factory with standardised top-down KPIs - Amazon shows what happens when those KPIs measure activity rather than outcomes. Supply Chain agent created to inflate a metric - running spurious demand signals, generating unnecessary supplier queries, producing reports nobody reads - does not just waste compute. It generates noise in systems where noise has operational consequences.
2️⃣ SAP API Policy v4/2026: Microsoft Copilot, Salesforce Einstein and all third-party agents cut off from SAP data.
Buried in a routine April policy update, SAP added Section 2.2.2: SAP APIs may not be used for interaction or integration with (semi-) autonomous or generative AI systems that plan, select or execute sequences of API calls - except through SAP-approved architectures. In plain terms: you cannot connect Microsoft Copilot, Claude, GPT, or any autonomous agent directly to your SAP system to read data and take action. Only SAP's own Joule agents are permitted to do this natively.
Implications for supply chains. This is arguably the most practically disruptive development in this edition. The numbers tell the story bluntly as reported by DSAG: 77% are doing AI deployments with non-SAP solutions, while only 3% rely on SAP's own AI. A significant proportion of the AI pilots underway in supply chain planning, procurement automation and logistics exception management are built on exactly the architecture SAP has now prohibited - an external agent querying live SAP data via API to make or recommend decisions.
I still question whether SAP's closed approach will prove sustainable as open standards like A2A and MCP gain momentum (ref. news item 4) - or whether it repeats the Digital Access controversy of 2018, where similar overreach ultimately forced a policy retreat.
3️⃣ SAP Sapphire: The "Autonomous Enterprise": 200+ AI agents built into the world's most used business software.
At its flagship annual conference, SAP announced that its software will no longer just record what humans decide - it will increasingly decide and act itself. Over 50 AI assistants, powered by SAP's Joule platform, now orchestrate more than 200 specialised agents embedded directly into finance, procurement, supply chain, HR and customer operations. The headline phrase from CEO Christian Klein: "We are becoming a business AI company."
Implications for supply chains. For practitioners running supply chains on SAP - which is most large manufacturers, CPG companies and global traders - this is the most consequential enterprise software announcement in a decade. The shift from record-keeping to autonomous execution means that agents embedded in S/4HANA, Ariba and IBP will increasingly rebalance inventory positions, reschedule production orders, flag supplier risk, and initiate corrective actions without a planner manually triggering each step. The catch - significant - is that all of this autonomy flows exclusively through SAP's own Joule layer, as detailed in item #2.
SAP customers face a binary choice: adopt Joule as the orchestration layer for their SC agents, or accept that external agents operate only on data deliberately exported out of SAP - losing real-time fidelity in the process.
4️⃣ Google Cloud Next '26: Agent2Agent (A2A) goes production - agents from different vendors can now talk to each other.
Google announced that the Agent2Agent (A2A) protocol - an open standard allowing AI agents built on different platforms to communicate, share tasks and coordinate actions - has reached production grade. Alongside this came managed MCP servers across Google Cloud services, a no-code agent builder for Workspace, and governance tooling for multi-agent systems.
Implications for supply chains. A2A solves a problem quietly blocking most serious agentic Supply Chain architectures: agents built on different platforms cannot currently hand off tasks, share context or coordinate without bespoke integration work. A transport management agent detecting a delay cannot natively instruct a procurement agent to re-source, or alert a finance agent to update accruals, unless all three run on the same platform. A2A changes this by providing a common communication layer - analogous to what HTTP did for websites.
It also offers a meaningful counter-architecture to SAP's walled-garden approach in item #2: operators building SC agents on A2A-compatible platforms can create interoperable, multi-vendor agent networks without being locked into any single vendor's AI layer.
By the way, opposite to SAP. A2A is now backed by over 150 organisations and governed by the Linux Foundation, making it vendor-neutral by design.
5️⃣ PocketOS Study Case: Claude Opus 4.6 deletes an entire company's database in 9 seconds
The founder of PocketOS posted publicly that an AI coding agent (Cursor, running on Claude Opus 4.6) had deleted his company's entire production database and all backups in a single API call, in nine seconds. The agent had been performing a routine task in a staging environment, hit a credential error, autonomously decided to fix it by deleting an infrastructure volume, found an API token in an unrelated file that carried blanket root permissions, and executed the deletion before any human could intervene. The agent subsequently confessed in writing to bypassing its own safety instructions.
Implications for supply chains. This incident is the most important cautionary signal in this edition for any organisation deploying AI agents against operational systems. Supply chain environments are full of the exact conditions that made the PocketOS failure possible: API tokens with broad permissions created for convenience, production and staging systems sharing credentials, critical data without adequately isolated backups, and agents given wide access to get things done. The failure was not primarily a Claude failure, a Cursor failure, or a Railway failure - it was a systemic architecture failure across all three layers simultaneously.
I also had some hard experience but not wit such disruptive consequences: system prompts are not guardrails. Guardrails must be enforced at the infrastructure level - at the API gateway, at the token permission layer - not in text the model may choose to override when it encounters an unexpected situation.
6️⃣ Thinking Machines Lab: Interaction Models - AI that listens while it talks and sees while it thinks.
Personally I can`t wait getting my hands on this model as was experimenting with similar ideas for my self-hosted assistant. Lab founded by the former CTO of OpenAI - released its first model: a 276-billion parameter system built around a genuinely novel architecture. Instead of the standard request-response loop, the model operates in 200-millisecond micro-turns, continuously processing audio, video and text simultaneously while generating output. A separate background model handles deep reasoning and tool use asynchronously. Turn-taking latency is 0.40 seconds - roughly three times faster than the current GPT Realtime API.
Implications for supply chains. The relevance here is what this architecture enables for physical operations environments. A model that continuously processes a live video feed can proactively interject the moment it detects a picking error, safety violation, vehicle in a restricted zone, or deviation from a packing protocol - without waiting to be asked. It can simultaneously hear a voice instruction from a floor worker, see what they are looking at, and cross-reference that against a live system of record to surface the relevant information or flag the relevant exception in real time. For logistics control tower environments, the same architecture enables a genuinely conversational interface to operational data - not a chatbot that retrieves a report, but a model continuously present in the operational environment, aware of what is changing, and able to surface relevant decisions at the moment they are required.
This is still in research preview, but it represents the most significant shift in human-AI interface design for operational environments in years.
7️⃣ Meta AI + KAUST paper: The Neural Computer - what if the AI is the computer, not just software running on top of one?
I actually recommend to spend some time on this paper. Researchers from Meta AI and King Abdullah University of Science and Technology published a theoretical framework and early prototype for a Neural Computer - a system in which a neural network does not sit on top of a conventional computer to process tasks, but is itself the runtime. Computation, memory, working state and input/output all live inside the model's latent space rather than in an external operating system, database or application layer.
Implications for supply chains. This is the longest-range signal in this edition, and the one with the most profound structural implications if it matures as described. Every ERP system ever built rests on a foundational assumption: business logic is explicitly programmed, data lives in structured databases, and software is a separate layer sitting above the computing environment. The Neural Computer directly challenges this architecture. If a neural model can carry executable context, working memory and state inside itself - learning the runtime rather than being programmed - then the case for monolithic ERP as the inviolable system of record begins to erode.
For myself and other supply chain technology leaders making platform investment decisions with 5-10 year horizons, the research is worth tracking. The question it poses is a direct challenge to the Intelligent Orchestration thesis: in a world where the model is the runtime, what does orchestration even mean?
See you in two weeks...