Period covered 4 – 10 October 2026Written by Tomo Miyata
This week I chose five stories on software that gives instructions to a facility's machines and people, and on how to decide what to leave to AI.
My lead story is a survey of AI use in logistics facilities by Logistics Reply. The company belongs to Italy's Reply Group and makes software that manages inventory and shipping inside facilities. Only 1% of respondents said AI instructs inventory management, core business systems, transport, labour and automated equipment together. Only 2.5% said AI makes operational decisions on its own.
The survey also asked how companies respond to disruptions such as stock shortages, labour shortages and equipment failures. Suppose an inbound delivery arrives half a day late. Do you reorder shipments, reassign staff, or simply wait? The question was how AI and people share that adjustment. 47.5% answered that AI flags the problem and people adjust by hand. Another 47.5% answered that AI proposes the next step and a person makes the final decision. Only 3.8% answered that AI coordinates the response across systems automatically.
Some companies are now setting out a framework for how much authority to give AI. Logistics Reply announced a framework that divides the authority given to AI into five levels (notify, recommend, execute, coordinate and act autonomously under governance) and sets it task by task. What AI can do and what it is allowed to do are separate questions, and authority widens as a track record builds up. Logistics Viewpoints, an analysis site covering logistics, argues that AI models are easy to swap out, and that the real differences lie in authority, approvals, records of what the AI did, and connections to core business systems.
Walmart, the U.S. retailer, shows what happens when a company keeps adding machines. It is automating about 200 distribution facilities in the United States. It took years to find a way to take items out of cardboard cases one at a time, and the wheeled robots that do the job are now being installed at a dozen or so facilities. The small robots that carry merchandise break down often, and the stoppages cascade. At automated grocery facilities, the monthly power bill can top $800,000 at peak times, against about $250,000 under largely manual operations.
Erik Nieves, chief executive of Plus One Robotics, a U.S. maker of parcel-sorting robots, says people should keep handling the exceptions that call for judgment. He expects robots to resolve more exceptions themselves, so that each supervisor oversees more robots.
I believe robots themselves will keep getting easier to adopt, including by renting them as a service. What will then set operators apart is whether a single piece of software can tell machines from different makers, and people, what to do in what order and where to take it. A second test is whether the operator has drawn a clear line between the decisions handed to robots and AI and those left to people.
Add machines without that layer of instruction and freight will wait between machines and equipment will stop more often. At Walmart, stoppages among the small robots cascaded and disrupted operations. GROUND's GWES is software that tells a logistics facility's machines and people what to do. It is designed with that instruction as its core function from the start, rather than as a combination of other vendors' software added later. Japan's logistics facilities face a serious labour shortage and handle a wide mix of products flowing together. I believe that whether a facility has one system that directs machines and people together will determine the gap in productivity.
GROUND Inc. — Founder & CEOTomo Miyata
This Week's News
Five stories
01Lead storyA vehicle moving freight in a logistics facility (illustrative). Photo: Unsplash
AI adoption surveyWork instructionVendor survey
Logistics Reply — Only 1% of respondents have AI instructing everything from inventory management to automated equipment together; 2.5% let it decide on its own
Logistics Reply, part of Italy's Reply Group and a maker of software that manages inventory and shipping inside facilities, published the results of a survey on AI use in logistics facilities on October 9. The respondents were logistics and supply-chain leaders in retail, food and beverage, manufacturing, third-party logistics (3PL) and automotive, among others. The announcement does not give the number of respondents. 27.9% were considering AI and 19.2% were piloting it in one or two processes, so about half are at the consideration or trial stage. 23.1% do not use AI, and 10.6% have built AI into most of their facility operations.
Asked about current uses of AI, 31.3% named producing reports and dashboards. 2.5% said they let AI make operational decisions autonomously. The survey also asked how respondents handle disruptions such as stock shortages, labour shortages and equipment failures. 47.5% said AI flags the problem and people adjust by hand, and another 47.5% said AI proposes the next step and a person makes the final decision. 3.8% said AI coordinates the response across systems automatically. 82.4% review AI's recommendations before acting on them, and 5% trust them enough to act without review.
58% said their AI works with a few systems to coordinate work, and 35% said it stays within a single application. 1% said it instructs the warehouse management system (WMS, the software that manages inventory and shipping inside a facility), the enterprise resource planning (ERP) system for core business operations, the transport management system (TMS), labour management and automated equipment together. The top capabilities respondents want next are recommending the best work to do (23%) and instructing work across multiple systems (22%). The top barriers to adopting AI were budget (25%) and data quality (22.1%). Michelle Jones, the company's director of pre-sales solution consulting, says the next challenge is to extend AI's role from presenting information to recommending, coordinating and instructing work.
Walmart — Automation of about 200 U.S. facilities runs into trouble, with power bills topping $800,000 a month at peak
The Wall Street Journal reported in October that Walmart's effort to automate about 200 distribution facilities in the United States is running into technical and financial obstacles. More than 140,000 people work in these facilities. Rob Montgomery, who oversees supply-chain operations for Walmart U.S., described the company as being at "peak complexity" and pointed to the difficulty of re-engineering facilities without taking them offline.
One difficult task is taking items out of cardboard cases and shipping them individually. Walmart spent years looking for a solution and is now installing small wheeled robots with motion and obstacle sensors at a dozen or so facilities. The cases it had used were too large for the automated systems, so it moved to shorter ones that must be taped shut. Separately, the small robots that carry merchandise fail often enough to disrupt operations, and breakdowns from causes such as dust build-up have led to cascading stoppages. Bulky products such as large bags of dog food are still handled by hand.
At automated grocery facilities, the monthly power bill can exceed $800,000 at peak times, compared with about $250,000 under largely manual operations. Walmart's main partner for non-perishable goods is Symbotic of the U.S. (which makes automation systems for logistics facilities), in which Walmart holds a 12.6% stake, but Symbotic has struggled to meet throughput targets. Walmart has invested nearly $1 billion since 2016 in automating the picking of online orders inside its stores. It bought the U.S. robotics company Alert Innovation for $400 million in 2023 and sold it to Symbotic for $200 million. Symbotic is expected to be able to offer a workable in-store system in 2028. Walmart expects its automation spending to peak in 2026 and 2027, and Montgomery says: "If we wait till perfect, it'll take too long."
Logistics Reply — Announces a framework that divides the authority given to AI into five levels and sets it task by task
On October 5, Logistics Reply announced the "LEA AI Agent Authority Model," which sets out how much authority to give AI agents (AI that carries out work on its own) in logistics facility operations. LEA is the company's execution platform for logistics facilities. The model decides how much to leave to AI using two measures: the organization's AI maturity (four stages) and the authority given for each task (five levels). The five levels are notify, recommend, execute, coordinate and act autonomously under governance. The company says that what AI can do and what it is allowed to do are separate questions, and that authority widens as a track record and trust build up.
The same day, the company began offering five ready-to-use AI agents in LEA's AI feature, Dynamic Intelligence, with no additional development needed. Three of them work as follows. One investigates the causes of stock-outs and separates real shortages from temporary problems. One spots uneven workloads and proposes reassigning staff so deadlines are met. One recalculates the ABC classification, which ranks products by how often they ship, from shipping data. Once a person approves the recalculated result, it is written back to the product master in the WMS.
Of the remaining two, one takes dock appointment bookings for trucks through a conversational interface. The other investigates the cause of delays from camera footage, including finding items that block the path of transport robots. Each agent operates under different authority and human oversight for each task.
Logistics Viewpoints — Commentary argues that AI results depend more on the operating framework around it than on the model
On October 5, Logistics Viewpoints, an analysis site covering logistics, published a commentary arguing that what surrounds an AI model matters more than the model itself in turning AI into results in real operations. It says the model has become the most easily swapped component of an AI system. Lasting advantage, it argues, comes from trustworthy operational data, system actions under controlled authority, approval mechanisms, records of what the AI did, and connections to core business systems.
Once AI can create shipments, change orders, contact carriers, allocate inventory and approve exceptions, the commentary says, governance becomes an operational necessity. It argues that limits on authority, spending caps, approvals, retries and the ability to reverse actions separate a demo from real operations. It also warns that AI can return plausible answers based on stale or incomplete information.
On measurement, the commentary says to measure operational performance rather than the fluency of AI's answers. Specifically, it lists time to decision, the amount of manual effort spent understanding the situation, how often people override the AI, time to recover from errors, the failure rate of system actions, and the business metric being improved. It says these measures give the basis for widening autonomy as results warrant. The commentary presents no new survey data.
Plus One Robotics — Chief executive says human oversight remains necessary even as humanoid robots spread
Erik Nieves, chief executive of Plus One Robotics, a U.S. company that automates parcel sorting and similar work, said in an interview with Robotics & Automation News that robots should keep handling routine tasks while people handle the exceptions that require judgment. In a survey by the company, 26% of Americans named a lack of human judgment as their biggest concern about humanoid robots. About 58% of respondents who were uneasy about certain robot tasks said they would feel better if a trained person could monitor the robot and step in.
Nieves says that foundation models, general-purpose AI trained on large volumes of data, shorten the time it takes to deploy or redeploy a robot. They do not make the robot move faster, he says. For high-volume work such as parcel induction and depalletizing, he expects purpose-built equipment to stay faster, cheaper and easier to maintain. Humanoids suit work where tasks change often, volumes are low and the facility cannot be rebuilt around a machine.
Five years from now, he expects purpose-built automation to still do most of the work. The bigger change, he says, is that robots will resolve more exceptions themselves, so that each supervisor oversees more robots.
