Agent scale: General Mills’
18h to 30min

18 hours to 30 minutes — General Mills compressed order-to-load optimization into half an hour, not with a stronger model but with three more mundane things: about 96-97% data accuracy, system-to-system order integration, and a bounded action space. Supply chain is currently one of the best-evidenced scale-up scenes for agents, but what buyers should take away is the readiness list — not the phrase “agentic AI.”

Evidence: Project ELF and the $3B cost plan

Paul Gallagher, General Mills’ chief supply chain officer, presented Project ELF (end-to-end logistics flow) at the October 2025 investor day: system-to-system (S2S) order intake with one of the company’s largest retailers plus truck-load optimization; agentic AI layered on top of machine learning, AI and GenAI; order intake and load optimization compressed from 18 hours to under 30 minutes; 15,000 tons of carbon removed so far because fewer trucks are on the road (Constellation Research, 2026-07-12; sources at end).

Two numbers in the same disclosure matter: supply chain master and transactional data accuracy stands at about 96-97% (“one version of the truth”); and proprietary manufacturing algorithms running on that data have already generated more than $40 million across multiple plants — and that is described as early days. The company-level pressure explains why: FY2026 revenue was $18.42B, down 5% year over year, with a net loss of $88M. Management announced a four-year $3B cost program (FY2027 target $750M): $2B from Holistic Margin Management and $1B from global transformation — supply chain redesign and AI are the transformation’s core levers. Earlier background (CIO Dive, 2025-02): AI models assess 5,000+ daily shipments, delivering more than $20M in transportation savings since FY2024.

Our judgment (1): why supply chain runs agents first — three structural conditions

Supply chain is one of the best-evidenced scenes for agent scale-up, and not by accident. It has three structural conditions that knowledge-work agents typically lack: ① quantifiable data accuracy — about 96-97%, the precondition for “letting the data decide,” not “the data is probably fine”; ② system-to-system integration — orders are structured messages, not PDFs or emails someone has to read; ③ a bounded action space — order-to-load-to-shipment is an enumerable state machine with clear decision boundaries. With all three in place, agents have something to run on, errors can be detected, and results can be measured.

Our judgment: knowledge-work agents are not blocked by models — they are blocked by these three conditions: long-tail tasks, unstructured input, and ambiguous decision rights. If you want to replicate supply chain success in knowledge work, replicate the conditions first, not a better model.

Our judgment (2): decomposing the 18h-to-30min lever

Is 18 hours to 30 minutes the work of “agentic AI”? Decomposed, the leverage comes from three layers: deterministic algorithm optimization (load and routing is an operations-research problem), process redesign (18 hours of manual reconciliation becomes system actions), and agentic execution (automating exceptions and coordination). The LLM is the last piece of the puzzle, not the whole puzzle.

Our judgment: do not be led by the “agentic AI go-live” narrative — without about 96-97% data accuracy and S2S integration, there is nothing for the agent to run on. Get process and data right first; agentification then accelerates a flow that already works. The reverse — agentifying messy data — simply automates the mess.

Buyer checklist: the readiness five for supply-chain agents

Action and the decision question for buyers

Thirty-day move: do not rush to agents. Run the readiness five — quantify data accuracy, inventory S2S, enumerate the state space, record baseline cycle times, define escalation paths. Only when all five pass, start a pilot.

The decision question: what is your supply chain data accuracy? If you cannot answer, agentification is premature — this is not a model problem, it is a foundation problem.

References: Constellation Research, “General Mills bets on AI, supply chain redesign to drive $3 billion in savings” (2026-07-12; Project ELF details from the October 2025 investor day, as reported by Constellation) https://www.constellationr.com/insights/news/general-mills-bets-ai-supply-chain-redesign-drive-3-billion-savings · CIO Dive, “General Mills attributes millions in cost savings to AI” (2025-02-19; 5,000+ daily shipments and $20M+ transportation savings background) https://www.ciodive.com/news/General-Mills-AI-cost-saving-strategy/740416/ · Google Cloud customer story (cloud and data migration background) https://cloud.google.com/customers/generalmills