Non classé

The Marginal Cost of Intelligence Is Collapsing. Supply Chains Will Never Be the Same.

Published

2 mois ago

12 août 2026

For most of modern business history, intelligence has been expensive. A transportation planner has only so many hours in a day. A procurement analyst can investigate only so many suppliers. A warehouse manager can examine only so many exceptions. A demand planner can explore only so many scenarios before a decision has to be made. Because human attention is finite and costly, companies have always rationed intelligence. They decide which shipments deserve investigation, which suppliers warrant deeper analysis, which inventory problems require intervention, and which operational exceptions simply are not important enough to justify someone’s time.

Artificial intelligence is beginning to change that economic constraint.

A recent analysis by Christopher Penn illustrates just how far the economics have moved. Penn describes how someone can assemble a surprisingly capable agentic AI environment using roughly $500 of commodity computing hardware, Linux, an open-source AI harness, and inexpensive cloud inference. He estimates that moderate usage could cost only a few dollars per month after the hardware purchase. The exact components will change quickly. Today’s preferred model could be replaced next month. Inference prices will continue to move, and open-source software will evolve.

But the important development is not the $500 computer. It is what the $500 computer represents: the marginal cost of applying intelligence to work is collapsing.

For supply chains, that could prove more important than another incremental increase in model intelligence.

From the Cost of Labor to the Cost of Reasoning

Consider how a supply-chain organization handles a late shipment today. A shipment is projected to arrive two days behind schedule, but that fact alone tells the company very little. Someone still has to determine whether the delay actually matters. They may need to examine the customer order tied to the shipment, available inventory at several locations, production requirements, alternative transportation options, contractual commitments, and perhaps even the importance of the customer involved.

Most companies cannot afford to conduct that level of analysis for every shipment continuously. Instead, they establish thresholds and exception rules. Software identifies something unusual, generates an alert, and places it in front of a planner or manager. At that point, the human becomes the scarce resource.

The same constraint appears throughout the supply chain. Procurement teams cannot continuously reassess every supplier. Inventory planners cannot investigate every potential imbalance. Transportation managers cannot evaluate every possible routing alternative every few minutes. Customer service teams cannot conduct deep root-cause analysis on every order exception. Managers cannot continuously model every possible response to every operational disruption.

Human attention is expensive, so organizations reserve it for the decisions they believe matter most.

AI changes the equation because the cost of another analytical cycle can approach fractions of a cent. An AI agent does not have to decide whether investigating a $500 problem is worth spending 30 minutes of a planner’s time. It can investigate thousands of such problems continuously, which creates a fundamentally different operating model.

The First AI Race Was About Intelligence

Much of the generative AI market has focused on a relatively straightforward question: who has the smartest model? OpenAI, Anthropic, Google, DeepSeek, and others compete across reasoning, coding, mathematics, multimodal capabilities, context windows, and benchmark performance. Those differences matter, but enterprise buyers are likely to become increasingly interested in another question: what is the least expensive intelligence capable of reliably performing this particular task?

That changes how enterprises should think about AI architecture.

A company probably does not need its most capable model to classify a routine transportation exception. It may not need frontier-level reasoning to retrieve a bill of lading, compare an invoice with a purchase order, summarize carrier correspondence, check inventory availability, or identify the appropriate standard operating procedure. A relatively inexpensive model may perform those tasks perfectly well.

More sophisticated models can then be invoked as complexity increases. A higher-end reasoning system might be reserved for ambiguous situations involving multiple objectives, financial tradeoffs, or incomplete information. Humans can remain responsible for decisions where judgment, accountability, or uncertainty warrants intervention.

The objective is no longer simply to deploy the smartest available AI. It is to match the cost of intelligence to the value and complexity of the decision. Once that shift occurs, AI begins to look much less like a software feature and much more like an economic resource.

Supply Chains Contain Millions of Small Decisions

This distinction is particularly important in supply-chain management because supply chains generate enormous numbers of decisions, and most of them are not strategic. They are small, repetitive, operational decisions that occur every minute of every day.

Should this order be expedited? Does this late shipment actually threaten production? Should inventory be transferred between distribution centers? Does this carrier invoice require review? Is this supplier’s declining performance statistically meaningful? Should this replenishment order be increased? Which customer orders are exposed to this port delay? Does this exception require a planner? Should this load be reassigned? Is there a lower-cost resolution available?

Individually, many of these decisions are not economically significant enough to justify extensive human analysis. Collectively, however, they determine enormous amounts of supply-chain performance. They drive freight expense, inventory levels, working capital, planner workload, expediting costs, service failures, production disruptions, and ultimately customer satisfaction.

This is why falling AI costs matter. AI does not merely make existing analysis cheaper. It potentially makes previously uneconomic analysis economically rational. That is a much larger shift.

Decision-to-Action Latency Becomes the Battleground

This connects directly to another emerging problem in supply-chain management: decision-to-action latency.

Supply chains have spent years improving visibility. Organizations can see shipments moving across networks. They receive estimated arrival times, weather alerts, inventory warnings, supplier notifications, and demand signals. The problem increasingly is not whether the organization can detect something. The problem is what happens after detection.

An event occurs. The organization sees it. Someone evaluates it. Information is gathered. Alternatives are considered. A decision is made. Approval may be requested. Someone else executes the decision. Every handoff introduces latency.

The economic promise of agentic AI is therefore not simply better prediction. It is the compression of the interval between signal and action.

Imagine that a shipment agent identifies a likely two-day delay. Instead of simply generating an alert, it determines which orders are affected. An inventory agent evaluates available stock. A production agent determines whether manufacturing schedules are exposed. A transportation agent examines alternative routes and expedited options. A procurement agent determines whether substitute material is available elsewhere. Another agent calculates the economic impact of the available responses.

Within seconds, the system can potentially transform a simple message — “Shipment delayed” — into something far more useful: “Three customer orders are exposed beginning Thursday. Transferring 240 units from Atlanta prevents two service failures. Expediting the remaining 80 units costs $1,840. No production schedules are affected. Recommended action: initiate the inventory transfer and expedite the remaining quantity.”

That is not simply visibility. It is decision architecture. And when the intelligence required to perform those calculations becomes inexpensive, that architecture becomes viable across far more exceptions.

Intelligence Is Becoming a Stack

Penn’s cheap-AI example also illustrates another structural change: AI is becoming modular.

The model providing reasoning can be separated from the inference provider supplying compute. The orchestration layer can be separated from the hardware. Databases, applications, containers, networking, and security can all exist as distinct components. That matters because enterprise AI does not have to be built around a single vertically integrated vendor.

The model becomes one component of a larger architecture rather than the architecture itself.

A company could use one model for inexpensive classification tasks, another for coding, another for complex reasoning, and perhaps smaller locally hosted models for workloads requiring low latency or greater data control. Models could increasingly be routed dynamically according to task complexity, cost, risk, and performance requirements.

In other words, enterprises may eventually operate portfolios of intelligence. That is a very different architecture from the idea that one flagship model should do everything.

The Model Is Not the Moat

This leads to an uncomfortable possibility for parts of the technology industry. If capable intelligence becomes inexpensive and models become increasingly interchangeable, simply having access to an advanced model becomes less differentiating. Something else becomes more valuable: the surrounding architecture.

A highly intelligent model that knows nothing about your business is interesting. A somewhat less intelligent model that understands your orders, shipments, inventory, suppliers, facilities, operating procedures, and customer commitments — and can act on them — may be far more valuable.

The durable competitive assets therefore become enterprise data, workflow integration, domain knowledge, operational history, knowledge graphs, governance, business rules, and execution capabilities.

The model provides intelligence. The enterprise provides context. And context is what turns generic intelligence into operational capability.

Why MCP, RAG, and Graph RAG Matter

This is also why the surrounding AI infrastructure is becoming so important.

Supply-chain AI cannot operate effectively if every agent exists in isolation. Agents need controlled access to enterprise tools and information. They need mechanisms for interacting with other agents. They need access to current knowledge that was never contained in the model’s original training data. Increasingly, they also need to understand relationships.

That is where technologies such as agent-to-agent communication, the Model Context Protocol, retrieval-augmented generation, and graph-enhanced reasoning become important. The key point is not the acronym. It is what these technologies collectively enable. This reflects the broader architectural direction toward connected intelligence described in ARC’s AI in the Supply Chain, where A2A, MCP, RAG, and Graph RAG are treated as complementary elements of an AI-enabled supply-chain architecture.

An AI system needs to know not merely that Supplier A exists. It needs to understand that Supplier A produces Component B, that Component B goes into Products C and D, that those products are manufactured at Plant E, that Plant E currently has six days of inventory, that replacement Supplier F requires twelve days, and that the material normally enters through Port G, which has just been disrupted.

Supply chains are networks of relationships.

An AI architecture capable of reasoning across those relationships can begin answering much more consequential questions: What happens next? Who is affected? What alternatives exist? What should we do?

That is where Graph RAG and related architectures become especially interesting. The objective is not simply better document search. It is system-level reasoning.

Cheap Intelligence Does Not Mean Cheap Architecture

There is an important caveat.

The falling cost of AI does not mean enterprise AI transformation suddenly becomes trivial. In many cases, intelligence may actually become one of the least difficult pieces.

The hard work remains. Enterprise data is fragmented. Product identifiers do not match. Supplier records contain duplicates. Transportation systems and ERPs disagree. Legacy applications lack modern APIs. Permissions are inconsistent. Operational knowledge lives in spreadsheets, email, documents, and people’s heads. Business processes contain exceptions that were never formally documented. And autonomous actions introduce serious governance questions.

An agent capable of recommending an inventory transfer is one thing. An agent authorized to execute one is another.

As AI becomes less expensive, these architectural and organizational problems become more visible. Companies may discover that they do not have an AI problem. They have a data problem, an integration problem, a process problem, or a governance problem.

The intelligence itself may increasingly be available on demand.

The Economics of Experimentation Change

There is another consequence that should not be overlooked: cheap intelligence dramatically reduces the cost of experimentation.

A supply-chain organization does not necessarily need to begin with a multimillion-dollar autonomous supply-chain transformation. It can start with a bounded workflow. Transportation exception management is an obvious example.

Give an agent controlled access to shipment events, order information, inventory availability, and the appropriate operating procedures. Allow it to retrieve relevant historical exceptions. Have it identify exceptions, determine business impact, diagnose probable causes, and recommend responses. Keep a human responsible for approval.

Then measure everything. How many exceptions did it correctly prioritize? How much planner time did it eliminate? How often were recommendations accepted? How frequently did humans override them? Did resolution time improve? Did transportation costs decline? Did service improve?

If the economics work, increase the scope. Then increase the agent’s authority. Then add another workflow.

The decreasing cost of AI makes this iterative approach increasingly practical.

A New Supply-Chain Metric: Intelligence per Dollar

This ultimately suggests that enterprises may need to think differently about AI economics.

The relevant measure may not be how much they spend on AI. It may be how much useful intelligence they obtain from that expenditure.

Call it intelligence per dollar.

How many decisions can the organization economically analyze? How many exceptions can it investigate? How many scenarios can it evaluate? How many decisions can it improve? How much decision-to-action latency can it remove? And what does each incremental unit of useful reasoning cost?

Once framed this way, the implications become significant.

An organization that can apply reliable intelligence to one million operational decisions for the same cost that a competitor spends analyzing 100,000 has an advantage even if both companies technically “use AI.”

The competitive advantage is not access. It is deployment efficiency.

The Next AI Race Is About Orchestration

For the past several years, the AI industry’s center of gravity has been the model: bigger models, smarter models, longer context windows, higher benchmark scores, more parameters, and better reasoning. Those advances will continue, but enterprise competition is likely to move upward through the stack.

The question becomes less “Who has the smartest AI?” and increasingly “Who can put useful intelligence against the largest number of economically meaningful decisions?”

That requires much more than a model. It requires enterprise context, reliable data, retrieval, agent orchestration, integration, governance, execution, and an architecture capable of determining which intelligence should be applied to which problem at which cost.

This is why the humble $500 AI server is more interesting than it initially appears. Nobody is going to operate a global supply chain from a bargain mini-PC. But the fact that an individual can now assemble a capable agentic computing environment from commodity hardware, open-source software, and a few dollars of monthly inference illustrates how dramatically the economics have shifted.

Intelligence is becoming abundant. The scarce resource is increasingly the ability to connect that intelligence to the business.

Intelligence Is Becoming a Commodity. Orchestration Is Not.

This may ultimately be one of the defining transitions of enterprise AI. The first phase was about gaining access to machine intelligence. The next phase will be about the economics of deploying that intelligence against work.

As inference prices decline and capable models proliferate, companies will increasingly be able to apply reasoning to decisions that previously could not justify human attention. For supply chains, with their enormous volume of small, interconnected, time-sensitive decisions, the implications are particularly significant.

The winners may not be the companies with access to the smartest model. They may be the companies that build the best decision architecture around increasingly abundant intelligence: companies that can determine what happened, understand what it affects, evaluate the alternatives, select the appropriate response, and execute that response with progressively less latency.

That changes the competitive question. It is no longer simply, “Who has the best AI?”

It becomes:

Who can deploy the most useful intelligence, against the most decisions, at the lowest cost, with the shortest path from decision to action?

That is a much more consequential race.

And it has already begun.

The post The Marginal Cost of Intelligence Is Collapsing. Supply Chains Will Never Be the Same. appeared first on Logistics Viewpoints.

Non classé

From Prompting to Persistent Decision Workflows in Supply Chain Operations

Published

9 heures ago

13 octobre 2026

Supply-chain AI is moving from episodic prompting toward persistent decision workflows. Repeated prompting can help an analyst solve a problem, but an operating capability must maintain context, monitor changing state, preserve decision logic, and connect recommendations to the work that follows.

A persistent decision workflow needs memory of the business object, the current plan, previous recommendations, overrides, policies, and outcomes. Without that continuity, users repeatedly reconstruct context and the model cannot reliably distinguish a new exception from a continuation of an old one. This persistence requirement becomes more important as the AI-agent execution model advances, because an agent cannot act safely if each interaction starts without durable operating state or prior decision context.

A mature design connects detection, context assembly, analysis, option generation, recommendation, approval, execution, and feedback. Some stages may use generative AI; others are better handled by optimization, rules, deterministic integrations, or human judgment. Persistence comes from the workflow and state model, not from the conversational interface.

The system should know what it may observe, recommend, prepare, execute, and escalate. Financial thresholds, customer impact, reversibility, confidence, and policy can determine the autonomy level. These controls make repeated decisions governable at scale.

ERP, planning, TMS, WMS, visibility, risk, and network platforms remain systems of record and execution; decision intelligence sits above and across them where context is assembled, tradeoffs are evaluated, actions are prioritized, and responses are coordinated. The persistent workflow therefore needs controlled write-back or action mechanisms, not only read access and summarization.

A decision workflow should record recommendations, human overrides, actions, and results. That creates an evidence base for improving rules, models, prompts, and autonomy thresholds. A chat transcript alone is not an operating learning system.

The architectural shift is from asking AI for help to designing a governed process in which intelligence participates continuously. That is a much stronger foundation for supply-chain decision support than repeated prompting.

Persistent decision workflows need memory and governance

A durable decision workflow has to preserve operating state across time. It should retain the event, relevant context, alternatives, constraints, prior decisions, approvals, overrides, actions, and outcomes so that the next step does not depend on a user reconstructing the problem in a new prompt. That persistence is what turns AI assistance into an operational capability.

Buyers should test continuity across handoffs. Can the system resume an unresolved decision after new data arrive? Can a different user see why an option was rejected? Can an agent act only within encoded authority? Can the resulting transaction be traced back to the evidence and rule that produced it? These are architecture questions, not user-interface questions.

Persistence changes the economics of decision support

When a workflow is persistent, the organization stops paying the same context-assembly cost every time a decision changes hands. New events can be evaluated against the existing state, prior alternatives remain visible, and approvals or overrides do not have to be reconstructed from email or chat history. That can reduce decision latency while improving traceability. It also creates a better foundation for automation because the system can distinguish a new decision from a continuation of an existing one. Buyers should therefore evaluate memory, state management, versioning, and auditability as operating capabilities rather than treating them as secondary AI features.

The 2026 Market Map is designed to help organizations understand the structure of the Decision Intelligence market, evaluate provider differences, and identify the capabilities most relevant to their operating environment. If your organization is evaluating Decision Intelligence platforms or clarifying where decision intelligence fits within the broader technology architecture, I would be glad to provide the Market Map brochure and discuss the evaluation questions and provider differences most relevant to your requirements.

Request the Decision Intelligence Market Map Brochure

For technology providers

Providers may request the brochure, discuss the research framework, or contact me to confirm how their capabilities are represented in the market assessment.

Discuss the research or confirm your profile

Non classé

Why SKU Selection Can Make or Break an AS/RS Business Case

Published

11 heures ago

13 octobre 2026

Why Sku Selection Can Make Or Break An As/rs Business Case

Automated Storage and Retrieval Systems (AS/RS) have become a cornerstone of modern warehouse automation strategies. Whether the technology involves shuttles, cube-based systems, cranes, or goods-to-person solutions, the business case is typically built around a common set of objectives: increasing throughput, reducing labor requirements, improving accuracy, and maximizing storage utilization. Yet one of the most important decisions in any AS/RS implementation often receives far less attention than the technology itself: SKU selection.

Companies frequently spend months evaluating vendors, software platforms, and automation architectures. However, even the most advanced AS/RS can fail to deliver expected returns if the wrong products are placed into the system. In many cases, the difference between a highly successful deployment and a disappointing investment comes down to a simple question: Which SKUs should be automated?

The Purpose of AS/RS Is Productivity, Not Storage

One of the most common misconceptions about AS/RS is that its primary purpose is to store inventory. While increased storage density is certainly a benefit, the real value of automation lies in improving warehouse productivity.

AS/RS creates value by reducing manual labor associated with inventory handling. Every automated retrieval eliminates walking, searching, traveling, and material movement. The more frequently a product is handled, the greater the opportunity for automation to generate measurable savings.

This is where SKU velocity becomes critical.

Not all inventory contributes equally to warehouse activity. In most distribution centers, a relatively small percentage of SKUs generates a disproportionately large percentage of order lines, picks, and replenishment activity. These high-velocity items create the majority of warehouse work and therefore present the greatest opportunity for automation.

Why High-Velocity SKUs Belong in AS/RS

The strongest AS/RS business cases are generally built around automating the products that move most frequently.

Consider two products:

SKU A is picked 2,000 times per week.

SKU B is picked 10 times per month.

Both consume storage space within the AS/RS, but only one generates meaningful operational value.

Every time SKU A is automatically retrieved, labor is eliminated and productivity improves. Over the course of a year, those savings accumulate rapidly. SKU B, on the other hand, occupies valuable automated storage capacity while contributing very little to throughput improvement or labor reduction.

Simply put, high-velocity SKUs produce the transactions that justify the investment.

When organizations prioritize their fastest-moving inventory for AS/RS storage, they maximize:

Labor savings

Pick productivity

Throughput performance

Equipment utilization

Return on investment

This is why leading warehouse operators often use ABC inventory analysis to determine AS/RS eligibility, with A-items receiving priority placement within the automated system.

The Hidden Cost of Selecting the Wrong SKUs

Many automation projects encounter challenges not because the technology underperforms, but because the SKU strategy was flawed from the start.

A common mistake is attempting to place too many slow-moving products into the AS/RS simply because capacity is available. While this may increase storage utilization, it often reduces economic utilization.

An AS/RS filled with low-velocity inventory can create several problems:

Lower Labor Savings

Products that rarely move generate few picking events. As a result, automation replaces little labor activity, weakening the original business case.

Reduced Throughput Benefits

System resources become occupied storing inventory that contributes minimally to order fulfillment activity, limiting the throughput impact of the investment.

Longer Payback Periods

If labor savings and productivity gains fall below projections, the expected payback period can extend significantly beyond the original ROI model.

Capacity Constraints

As demand grows, operators may discover that high-velocity items no longer have sufficient space within the AS/RS because slower-moving products are consuming capacity.

The technology itself may be performing exactly as designed, but the financial return can still disappoint.

Focus on Activity, Not SKU Count

One of the most important lessons for warehouse leaders is that AS/RS success should not be measured by the number of SKUs stored within the system.

Instead, it should be measured by the percentage of warehouse activity flowing through the system.

For example, placing 1,000 high-velocity SKUs into an AS/RS that captures 80% of total picking activity will typically generate greater value than storing 5,000 SKUs that account for only 40% of picks.

The goal is not to automate the greatest amount of inventory.

The goal is to automate the greatest amount of work.

That distinction has significant implications for system utilization and overall project economics.

The Importance of Continuous SKU Evaluation

SKU selection should not be treated as a one-time exercise conducted during system commissioning.

Demand patterns change. Product popularity shifts. New items are introduced while others decline.

As a result, warehouse operators should periodically re-evaluate which products belong inside the AS/RS. High-velocity inventory today may not be high-velocity inventory next year.

Organizations that continuously align automated storage capacity with current demand patterns are often able to sustain higher utilization rates and stronger ROI throughout the life of the system.

ROI strategy

Consider a simple ROI example. Assume a company invests $5 million in an AS/RS and expects to save $1 million annually through labor reduction and productivity improvements. The original business case assumes the system will handle 75 percent of warehouse picking activity.

If the company correctly slots high-velocity products into the system, the projected labor savings can be achieved, resulting in a payback period of approximately five years. However, if management fills a large portion of the AS/RS with slow-moving inventory, the system may only process 40 percent of warehouse picking activity. Annual savings could fall to $600,000, extending the payback period from 5 years to more than 8 years. The technology has not failed. The SKU strategy has.

Final Thoughts

AS/RS technology continues to play an increasingly important role in warehouse automation strategies. Yet successful implementations depend on much more than selecting the right hardware or software platform.

The most effective AS/RS deployments are built around careful SKU selection.

High-velocity products generate the majority of warehouse activity and therefore create the greatest opportunity for labor reduction, throughput improvement, and productivity gains. Conversely, filling automated systems with low-velocity inventory can significantly dilute returns and extend payback periods.

For warehouse and supply chain executives, the takeaway is straightforward: an AS/RS should be designed around activity, not occupancy. The objective is not simply to store products in an automated system. The objective is to automate the work that drives operational performance. In most facilities, that means ensuring the highest-velocity SKUs receive priority access to valuable AS/RS capacity.

The difference between a successful automation investment and an underperforming one may ultimately come down to a decision that appears deceptively simple: choosing the right SKUs.

The post Why SKU Selection Can Make or Break an AS/RS Business Case appeared first on Logistics Viewpoints.

Non classé

What the Descartes Innovation Forum Said About the Next Phase of Logistics Execution

Published

11 heures ago

13 octobre 2026

After several days at the 2026 Descartes Innovation Forum, the clearest takeaway was not simply that artificial intelligence is coming to logistics. That part is already obvious. What stood out more was where the technology is moving and why it matters now.

Across the keynote, a transportation leadership panel, a Walmart Marketplace session, and a detailed discussion of carrier fraud, the same operating problem kept surfacing in different forms. Logistics organizations have more data, more applications, more connectivity, and increasingly more AI, but the real value comes when those capabilities help someone make a better decision or execute the next step faster, more reliably, and with the right controls in place.

That matters because the operating environment is not getting easier. Transportation capacity is tightening, rates and fuel costs remain under pressure, fraud and cargo theft are serious concerns, customer expectations continue to rise, and companies are being asked to improve productivity without adding equivalent labor. Against that backdrop, the most interesting technology at the Forum was not technology that simply explained what was happening. It was technology aimed directly at the work itself.

The Swivel-Chair Problem Is Still Real

Descartes opened the Forum with a problem that will sound familiar to anyone who has spent time inside transportation operations. Companies may have capable systems and extensive networks, yet much of the actual work still happens in the gaps between those systems.

Someone receives an email, opens a portal, checks a transportation system, moves to another application, sends a message, waits for a response, updates a record, and then starts the process again somewhere else. Descartes referred to this as the “swivel chair tax,” and the phrase worked because it described a problem technology has never fully eliminated.

The company’s answer is its Agent Control Plane, which is designed to expose capabilities from Descartes applications as skills that AI agents can use. As I covered earlier in Descartes Takes Agentic AI Into the Logistics Workflow at the Descartes Innovation Forum, Descartes was careful to distinguish that approach from a basic chat interface or a traditional workflow engine. The intent is to allow agents to work across applications and adapt as conditions change rather than simply follow a rigid sequence of predefined screens.

The live demonstration made the concept more tangible. Descartes followed an order across four representative companies—a retailer, freight forwarder, customs broker, and final-mile provider—with 15 Descartes products supporting 28 process steps behind the Agent Control Plane. Many of those steps happened without requiring a person to manually move information from one application to another.

What stood out to me was not the amount of automation. It was where Descartes chose not to automate.

At one point, an agent identified a carrier option that could save roughly 8 percent. The carrier, however, had a poor trust score, thin operating history, and other warning signs. Rather than make the lower-cost choice automatically, the system surfaced the decision for human review. The human declined the cheaper carrier.

That is a much more practical model for agentic AI in logistics than the idea of simply handing an operating process over to a machine. Descartes repeatedly described autonomy as a dial, not a switch. Agents have identities, their actions are logged, activity can be replayed, and organizations can determine which decisions require human approval.

The significance is straightforward. Once AI begins taking actions inside real operating processes, governance stops being an abstract IT issue and becomes part of the operating model.

The Market Is Making Better Execution More Valuable

The transportation leadership panel I moderated later in the Forum put that technology discussion into a much tougher operating context.

The freight market described by the panelists was not one of strong demand and easy growth. It was one in which relatively weak demand is colliding with meaningful reductions in capacity. One participant cited a decline in the number of trucking companies from roughly 340,000 to 282,000 and dry-van truckload rates up approximately 33 percent year over year. The larger point was that the conversation has shifted. Two years ago, shippers were largely focused on rates. Today, capacity availability and reliability matter much more.

That tighter environment also increases risk. When capacity is harder to find, there is greater temptation to use carriers that may not have passed the same diligence in a looser market. Panelists tied that pressure to growing concerns around freight fraud, cargo theft, safety, insurance, litigation, and carrier qualification.

As discussed in Transportation Leaders See AI Moving From Experiment to Operating Model at the Descartes Innovation Forum, the response is not simply to cut spending. Technology did not sound discretionary in this discussion. One participant described a single repetitive business activity being performed roughly 750,000 times per month, representing the equivalent of about 1,500 employee-hours per day. That is exactly the kind of workload where automation can move from being interesting to economically significant.

The panel also showed how quickly expectations around technology development are changing. Voice agents, email automation, operational workflows, data science, and agentic AI were all discussed as practical tools rather than distant possibilities. One participant described a proprietary TMS containing more than 30 million lines of code and roughly 2,000 database tables. A 12-person team working with AWS and Anthropic was then tasked with recreating the system using AI-assisted development and reportedly produced a replication in approximately six weeks.

Replicating a system is not the same as proving it in full production, so that example deserves some caution. Still, it shows how quickly the economics of software development may be changing.

Transportation technology has traditionally changed slowly because systems were complex, deeply integrated, and expensive to modify. AI may compress that cycle dramatically. For transportation leaders, that creates an interesting tension: the operating environment is getting harder at the same time that the technology available to respond to it is improving faster.

Walmart Brought the Conversation Back to the Customer

A Walmart Marketplace session provided a useful counterweight to the AI-heavy discussions elsewhere at the Forum. Bryce Caswell, Director of Ship with Walmart for Walmart Marketplace, focused on something much more basic: the customer promise.

Walmart described four foundations for marketplace success—assortment, listing quality, pricing, and shipping—and emphasized that the four are interdependent. Good assortment will not solve poor content. Good content will not offset uncompetitive pricing. And none of it matters if the seller cannot deliver the product when promised.

That may sound obvious, but it was one of the more useful reminders of the Forum. Logistics technology discussions can quickly become dominated by architecture, optimization, visibility, and AI. The customer evaluates something much simpler: did the order arrive when you said it would?

As I wrote in Walmart Is Turning Marketplace Logistics Into a Platform, Walmart’s emphasis was not simply on speed. Reliability and consistency mattered just as much. During holiday peaks, sellers were encouraged to protect inventory availability, leave enough time to move product into the fulfillment network, understand what they are promising, and then meet that promise.

The company is also trying to simplify the mechanics underneath that promise. Walmart Fulfillment Services provides pick, pack, and ship capabilities through Walmart’s fulfillment network, while its multi-channel model allows sellers to support multiple sales channels from a single inventory pool.

Ship with Walmart extends that logic into transportation by offering access to negotiated carrier rates, including national, regional, and LTL options, while adding certain protections around on-time delivery and lost shipments.

The larger point is that simplification has value only if it improves execution. Whether the tool is an AI agent or a fulfillment network, the objective is the same: remove unnecessary friction so people can focus on the decisions and exceptions that actually matter.

Fraud Changes the Meaning of Visibility

The Forum’s carrier fraud session added another layer to the discussion because it showed how operating data is being repurposed for trust.

Descartes demonstrated how MyCarrierPortal and MacroPoint can be used together to strengthen carrier vetting before a load is tendered. The process extends beyond checking insurance or basic carrier status. It can incorporate configurable risk assessments, tracking history, lane history, inspection information, validated users, equipment data, and driver-level tracking statistics.

That is a subtle but important shift in how transportation visibility data is used. Historically, tracking information primarily answered a familiar question: where is my load? Increasingly, that same information can help answer a different question: should I trust the party moving it?

The same theme runs through Descartes Expands Freight Fraud Prevention Across Carrier Vetting and Shipment Execution. Descartes showed how historical lane activity and tracking behavior can be used as additional carrier-vetting signals. Once a shipment is in transit, the Notification Center can surface route deviations and other exceptions that warrant attention without requiring someone to monitor every load manually.

ELD Plus adds another layer by allowing two tracking sources to be used within the same tracking session. If mobile-device data and telematics data begin to diverge, that mismatch becomes a signal that someone may need to investigate.

Carrier lookup goes further by combining information such as power units, active loads, tracking history, spoofing flags, and recent lane activity. The example used during the session was simple: if a carrier has 100 power units but appears to be active on 600 loads, that discrepancy may deserve a closer look.

More automation does not reduce the need for trust. It raises the stakes. If systems are going to make recommendations or take actions faster, the quality of the signals behind those decisions becomes even more important.

Execution Is Becoming the Real Battleground

Looking back across the Forum, I do not think the central story was simply AI. It was execution.

Descartes is trying to reduce the swivel-chair work that still exists between logistics applications and companies. Transportation leaders are looking for technology that can improve productivity and help them operate through a tighter, riskier freight market. Walmart is focused on simplifying fulfillment so sellers can keep the customer promise. Carrier-verification tools are turning tracking data into a source of trust as well as visibility.

Put together, that points to a broader shift in logistics technology. Software is moving closer to the actual transaction, the actual decision, and the actual exception. The value of that technology will increasingly be judged by whether it helps organizations act better, not simply see more.

The companies that benefit most will not necessarily be the ones with the most AI, the most dashboards, or the most data. They will be the ones that can connect intelligence to execution without losing control of the process.

For me, that was the real takeaway from the 2026 Descartes Innovation Forum. The next phase of logistics technology is not just about making systems smarter. It is about knowing when the system should act, when a person should decide, and how to make both work together fast enough to keep freight moving and the customer promise intact.

The post What the Descartes Innovation Forum Said About the Next Phase of Logistics Execution appeared first on Logistics Viewpoints.