AI in the Supply Chain: TraceLink Puts Agent Plans Up for Approval, Zalando's Returns Robots Get a Simulation Training Loop, and Tecsys Forecasts Hospital Drug Purchasing
Title: AI in the Supply Chain: TraceLink Puts Agent Plans Up for Approval, Zalando's Returns Robots Get a Simulation Training Loop, and Tecsys Forecasts Hospital Drug Purchasing
Date: 22 October 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2888 words
Reading time: 11 min
Published: 22-10-2026
TraceLink said on 21 October that its supply chain agents now work from plans that staff can inspect, test and approve before the agents go live. The same day, Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando described a simulation loop that retrains the robots sorting Zalando's returns at two CEVA sites. On 20 October Tecsys released AI tools that forecast hospital pharmacy stock and flag drug purchases that could break US 340B programme rules. This paper sets out how each use works, what is known about cost, and where it breaks.
What happened this week
This is the fourth of SAASiQ's Thursday papers on how AI is used in practice, sector by sector. Finance came first on 1 October, HR on 8 October and the public sector on 15 October. This week is the supply chain: planning, warehousing, transport and the purchasing that feeds them.
Three announcements fell in the week. TraceLink, which runs a network for pharmaceutical and life sciences supply chains, released a change to how its OPUS agents plan their work on 21 October. Sereact, a German robotics company, announced a training pipeline with NVIDIA and the cloud provider Nebius on the same day. Its robots already handle fashion returns for Zalando at sites run by CEVA Logistics. Tecsys, a Canadian supply chain software company, made AI tools for hospital pharmacies generally available on 20 October.
Two earlier items set the context. On 6 October Descartes Systems Group launched an Agent Control Plane at its Innovation Forum in Chicago. The same day the 2027 Third-Party Logistics Study was published at the CSCMP EDGE conference, with a section on AI adoption.
Use case one: agents that run recurring supply chain work
Much supply chain work is the same task done many times. An order changes, a shipment is late, a supplier sends a new date. Someone reads the message, checks the system, updates the record and tells the next party. Descartes put this plainly in its 6 October release: teams spend a lot of time interpreting updates from emails, messages and portals, rekeying information and coordinating next steps.
Agents are being sold to take over that loop. TraceLink's OPUS agents are set up without code. A team gives an agent an objective and the tasks, decisions and rules it may use. Until now, according to TraceLink, an agent of this kind would work out its approach afresh each time a similar event arrived. TraceLink argues that this adds processing time and cost and can give inconsistent results.
The 21 October change adds a reasoning layer TraceLink calls OPUS Brain. It turns the agent's instructions into what the company calls a Semantic Inference Plan. The plan sets out how the agent will reach its objective, using only approved tasks, decisions and rules. OPUS Brain generates the plan, evaluates it and refines it. Staff can then inspect it, test it and approve it before it runs in production. Once approved, the plan guides the agent each time the same kind of work comes up. TraceLink says this lets agents handle recurring work in seconds.
The release also lists the records that go with it. Agent Profiles are versioned, so a change to an agent's instructions can be tracked. Evaluation Logs and Work Logs record what each agent did. Dashboards show performance, completed tasks and the costs attached to them.
Descartes is taking a similar line in logistics. Its Agent Control Plane sits on the Descartes Global Logistics Network and coordinates agents across applications and trading partners. Ken Wood, Descartes' executive vice president of product management, said agents act on events within permissions and approval points that customers define. The product keeps a traceable record of agent actions. It can also make Descartes agents available inside AI platforms that customers already use. Neither company published a price or named a customer running the new features.
Where supply chain agents break
The first weak point is the evidence. Runtime Wire noted on 21 October that TraceLink's release contains no independent benchmarks and no production customer results. The speed and reliability claims are the company's own. A buyer should ask for figures from a live customer, on the buyer's own transaction types, before relying on them.
The second is that an approved plan describes the expected case. Supply chains produce a steady stream of unexpected ones: a partial shipment, a lot number that does not match, a supplier that replies in a free-text email. The question is what the agent does when the event does not fit the plan. It can stop and hand over, or it can improvise. The approval only covers the plan, so improvisation needs its own limits, set in writing.
The third is the trading partner. These agents act across company boundaries. When an agent updates a shared record or sends a message to a supplier, the supplier's systems treat it as an instruction from the buyer. If the agent is wrong, the error moves down the chain. The contract with each partner should say who is liable for an agent's message, and how a partner can tell an agent from a person.
The fourth is drift in the instructions. A versioned profile shows that an agent's rules changed. It does not show whether anyone checked the change. Somebody has to own each agent, review changes before release, and read the logs. That is a staffing cost, and none of the 6 or 21 October releases put a figure on it.
Use case two: robots that sort returns
Fashion returns are hard to automate. Each returned item is different. Traditional warehouse robots are trained on fixed products with known shapes.
Sereact's dual-arm robots went into live operation at two CEVA Logistics sites handling returns for Zalando, as reported on 7 October. The sites are Greven in Germany and Świebodzin in Poland. The robots run on Sereact's Cortex software, which is designed to identify and grasp items it has not seen before. It does not need separate training for each product or a fixed profile for each stock-keeping unit. The robots are supplied as a service, so CEVA pays for robotic capacity rather than buying the machines. Zalando invested in Sereact's Series B round in July as a strategic investor.
The partners did not publish the number of stations at either site or a figure for items handled per hour. The announcement describes the deployment as the first milestone in a plan to extend robotic returns handling across Europe.
The 21 October announcement covers how the robots get better. Each time a robot works at a returns station, the video and the arm's movement data are saved to Nebius object storage. NVIDIA's Isaac Lab software turns those recordings into thousands of simulated variations. NVIDIA's Cosmos models then re-render the variations with different garment states, packaging, materials and lighting. The resulting dataset is used to train a new version of Cortex on NVIDIA Blackwell GPUs.
The new version is then tested against held-out simulated cases. It has to pass a threshold that Sereact sets. If it passes, it goes to the real stations for a live test. If it fails, the system generates more variations of the cases it got wrong and trains again. An AI agent runs parts of this loop. It previews the GPU cost of each training run and decides which failures need more simulated data. Engineers set the goals and must approve any run above a set GPU spending limit.
The companies say a fix for an error spotted on a Monday can now come back as a tested update within days, where the old process took weeks. Nebius is offering the pipeline to other robotics teams as a reference design on its AI cloud.
Where returns robots break
A pass in simulation earns a live test, as one outlet covering the announcement pointed out. The test that matters is the rate of errors at the station after each update, measured against the version before it.
Updates also carry risk. A training loop that ships a new robot policy every few days changes the behaviour of equipment working next to people. Warehouse operators already manage software releases for their management systems. Robot policy updates need the same discipline: a release note, a rollback route, and a named person who signs off the change at each site.
Under robotics as a service, the operator avoids buying the machines and pays a fee instead. The partners did not publish the terms. An operator should know whether the fee is per robot, per hour or per item. Each one rewards different behaviour. A per-item fee, for example, pays the supplier more when volumes rise, whether or not the robot improves.
The last weak point is the data itself. Every returns episode is video of a customer's returned goods, recorded in a third-party cloud. The operator, the retailer and the robot supplier need to agree who owns those recordings, how long they are kept, and whether they can be used to train robots for other customers.
Use case three: forecasting hospital drug purchases
Hospital pharmacy supply is a supply chain with compliance rules on top. In the United States, the 340B Drug Pricing Program, run by the Health Resources and Services Administration, lets eligible hospitals buy outpatient drugs at reduced prices. Some types of hospital in the programme are also barred from buying covered outpatient drugs through a group purchasing organisation. Tecsys' new tools are aimed at the point where compliant stock runs short and the next order could break that rule.
Tecsys' tools, released on 20 October, run on TecsysIQ, the AI engine of its Elite platform. The 340B Integrity Agent forecasts when a hospital's compliant purchases of a drug are likely to run out. It helps pharmacy teams decide how much to buy now. It flags hospitals that are subject to the purchasing ban and are nearing a compliance risk, so buyers can raise a corrective purchase order before a breach. Tecsys described this in June as a capability not currently available in the market. That is the company's own claim.
The second tool, TecsysIQ Inventory Visibility, brings inventory data into one view. It draws on automated dispensing cabinets, pharmacy software and ERP systems. The first version covers pharmacy environments. A third tool, TecsysIQ Image Capture, has been available since June.
Where pharmacy forecasting breaks
The forecast depends on the feeds. A single view of stock across dispensing cabinets, pharmacy systems and the ERP depends on every one of those systems reporting on time and using the same drug identifiers. Where a cabinet is offline or a product code differs between systems, the view will be wrong without saying so. A hospital should test the tool against a physical count before trusting the totals.
The legal duty stays with the hospital. If a hospital breaches the purchasing ban, the hospital answers for it, not the software supplier. So the agent's alerts need an owner in the pharmacy team, and the hospital needs a record of each alert and what was done about it.
Tecsys published no accuracy figures for the forecasts and no price. A hospital buying the tools should ask how often the forecast was wrong in testing, and in which direction. A forecast that runs early wastes compliant stock. One that runs late produces the breach it was bought to prevent.
How far adoption has got
Two recent studies give a sense of scale. The 2027 Third-Party Logistics Study, published on 6 October with NTT DATA, Penske Logistics, the University of Tennessee and CSCMP, found that 93 per cent of shippers rate AI capabilities as important. Only 12 per cent of logistics providers called themselves highly prepared, and 82 per cent said they were moderately prepared. Shippers use AI for predictive analytics and risk planning more often than their providers, at 66 per cent against 43 per cent. The barriers cited were data quality and availability (61 per cent), legacy system integration (54 per cent) and budget (46 per cent).
Gartner published a forecast on 24 September from a survey of 243 senior leaders at organisations with annual revenue of at least $500 million. It found that 83 per cent had spent at least $3 million on automating supply chain planning, AI included. Even so, Gartner predicts that by 2030 only 5 per cent of organisations automating planning will make at least 10 per cent of their planning decisions autonomously. Jan Snoeckx, a Gartner analyst, said full autonomy will stay concentrated in routine, high-frequency and lower-risk decisions.
Manufacturers are moving at their own pace. Unilever said on 16 June that it will work with Accenture to build AI-enabled digital twins at about 40 sites over 18 months. The work builds on five sites already running: one in the United States, one in Poland, one in Vietnam and two in India. Unilever reports a 20 per cent cut in waste and a 10 per cent gain in capacity at its Raeford plant in North Carolina. Its Gandhidham site in India reports a 30 per cent cut in Dove soap quality defects over four years. These are company figures. Unilever did not publish the cost of the programme or the list of new sites.
Sandy Douglas, chief executive of the food distributor UNFI, said at a Goldman Sachs conference on 15 September that UNFI uses Samsara's AI-based platform for visibility across close to 50 distribution centres and 2,000 delivery vehicles. He said it helps cut the miles driven on customer deliveries. UNFI also hedges its fuel and passes surcharges on to customers under contract. US on-highway diesel was $6.18 a gallon in Energy Information Administration data for 12 October, up $2.51 on a year earlier.
What it costs and who pays
None of the three companies published a price. The releases do show where the costs fall.
For agents, the running cost is the model's work on each event. TraceLink's case for approved plans is partly a cost case: an agent that follows a stored plan does less fresh reasoning per event than one that starts from scratch. Its dashboards track the expense attached to each agent's tasks. A buyer should get that cost per transaction in writing, with what happens when volumes double at peak.
For robots, the training loop has its own bill. Every retraining run uses GPU time on a third-party cloud. Sereact's agent previews that cost before each run, and engineers approve anything above a set limit. Under a service contract that cost may sit with the supplier, or it may be passed through. The operator should find out which.
Then there are the costs outside the AI line, and they are mostly staff. Someone owns each agent and reviews its changes. Someone at each site signs off robot updates. Someone in the pharmacy acts on each alert. In SAASiQ's view, a business case that counts licence fees and compute but leaves out those roles will understate the cost.
A framework for supply chain teams
The steps below apply whichever suppliers a team uses. They follow the three use cases above.
Start with an inventory. List every AI feature in use across planning, purchasing, warehousing and transport. Include features that arrived in a supplier update and agents that trading partners run against your systems. For each, record what it reads, what it changes, and who acts on its output.
Decide what each agent may do alone. For each type of event, write down whether the agent drafts, recommends or acts. Where it acts, set the limits: order value, quantity, which partners it may message. Approval of a plan should cover the expected case. Anything outside the plan should stop and go to a person.
Tell your partners. Agree with each trading partner how agent messages are marked, who is liable for an agent's error, and how either side can switch an agent off. Put it in the contract or the service terms.
Treat robot and model updates as releases. Every update to a robot policy or an agent's instructions needs a version number, a test result, a rollback route and a named approver at each site. Measure live error rates before and after each update.
Test the data before the forecast. Before relying on a single view of stock, reconcile it against a physical count. Check that product identifiers match across every system that feeds it. Repeat the check after any system change.
Get the cost model in writing. For each AI service, find out whether you pay per seat, per transaction, per robot, per item or per unit of compute. Ask what happens to the bill at peak volume. Ask who pays for retraining.
Keep the record. For each AI action that changes an order, a stock level or a partner's record, keep what the system was given, what it did and who approved it. Check that the record can be produced for an audit or a dispute with a partner.
What happens next
Tecsys' Inventory Visibility and 340B Integrity Agent are available now. Nebius is offering the Sereact training pipeline as a reference design to other robotics teams on its cloud. The CEVA and Zalando deployment is the first step in a plan to extend robotic returns handling across Europe. Unilever's digital twin rollout runs for 18 months from June 2026.
SAASiQ - Intelligent Solutions for SaaS ©
