Home
Welcome
About
Profile & credentials
The Casebook
Real case studies with measured outcomes
Product Playground
Try a concept. Play with a product.
The Radar
Insights on AI in logistics
The Manifest
Daily AI × supply chain signal
Let's connect
Back to The Radar
ai · supply-chain11 min read2026-07-16

Most AI Pilots in Supply Chain Die on the Handoff, Not the Model

Most AI Pilots in Supply Chain Die on the Handoff, Not the Model

Between June and July I kept writing about the same problem from different angles. This piece puts those posts in one place.

AI makes people faster. That does not make the company better.

AI is making people faster.

That does not mean the company is getting better.

Gallup's 2026 workplace research captures the gap clearly: among U.S. workers in organizations that have implemented AI, 65% say AI has had a positive impact on their individual productivity. But only 12% strongly agree that AI has transformed how work gets done.

That gap is where most AI ROI disappears.

In supply chain, I see this often. A planner can summarize demand faster. A buyer can draft supplier emails faster. An analyst can clean a report faster.

But if the forecast handoff is still messy, if purchasing still waits for approval in email, if operations still trusts a spreadsheet more than the system, the business does not move much faster.

My view: AI productivity is a workflow design metric.

The question is not, "Did this tool save someone an hour?"

The better question is:

"Did that hour change a decision, remove a handoff, reduce an exception, or improve service?"

Gallup also found that manager support matters. Employees whose managers actively support AI use are 8.7 times as likely to strongly agree AI has transformed how work gets done.

That tells me the next AI bottleneck is not the model. It is the management system around the model.

The pilot is stuck on the handoff

Most AI pilots in operations do not get stuck on the model.

They get stuck on the handoff.

Freda Duan had a sharp line on X. Everyone is adding AI into the workflow. Almost nobody is asking why the workflow looks that way in the first place.

That matches what I keep seeing in supply chain work.

The hard part is usually not summarizing the exception, drafting the update, or pulling the signal faster.

The hard part is what happens next.

The planner flags the issue. The buyer checks supplier exposure. The ops lead decides whether to expedite.

Same 3 inboxes, same 2 days, same approval path built for a team that doesn't exist anymore.

Nobody has redesigned the transfer point between those decisions.

MIT Sloan summarized this well in April: AI creates the biggest shift when it reshapes how tasks are sequenced, grouped, and handed off between humans and machines. NBER Working Paper 34859 pushes the same idea further: AI creates more value when adjacent tasks can run as a chain instead of breaking across constant review points.

I have sat through enough rollouts to know this is where promising pilots slow down. Teams add AI to one step, but leave the approval path, escalation rule, and exception ownership untouched.

That is not AI transformation.

That is faster work inside the same old coordination loop.

And the reason is uncomfortable: most approval steps don't protect the process. They protect the approver.

Somebody got burned once, so someone added a checkpoint, and now it's how we do things.

AI doesn't dissolve that.

It just gets you to the wait faster.

Adoption is rising faster than redesign

AI use in retail supply chains went from 24% to 40% in two years.

That sounds like momentum. It does not yet sound like transformation.

Inspectorio's State of Supply Chain Report 2026 says 40% of retail supply chain leaders report active AI use in 2026, up from 24% in 2024. The same report says the top barriers are budget conflicts at 39%, integration complexity at 36%, and change management capacity at 36%.

I do not read that as a tooling problem.

I read it as an operating model problem. Adoption can rise long before the workflow gets redesigned.

Teams add AI to quality checks and other bounded tasks. But the harder questions still sit with humans:

  • Who owns the decision when the signal crosses functions?
  • Whose data becomes the source of truth?
  • Who changes the process when the old handoff no longer fits?

The more important signal is this: companies can deploy AI faster than they can unify data, integrate functions, and redesign decision rights around it.

If that gap stays open, adoption will keep rising faster than ROI.

The next wave of winners will connect quality, sourcing, compliance, and operations tightly enough that AI can move from a productivity layer into a decision layer.

Visibility is not execution

Supply chain AI is moving from visibility software to execution software.

That is the product shift I am watching.

For years, the main promise of supply chain tech was visibility:

Where is my shipment? What is the ETA? Which order is delayed?

That visibility layer is useful. But in actual operations, visibility does not close the loop.

If a truck is going to miss its delivery appointment, the dashboard can show the delay.

But someone still has to contact the carrier, reschedule the appointment, update the TMS, notify the customer, and follow up for POD.

That messy middle is where AI products are starting to move.

When I compare project44, FourKites, Vooma, and HappyRobot, I do not see four identical products.

I see different starting points converging toward the same product direction.

project44 and FourKites start closer to the visibility and control tower side. They sit near shipment data, carrier networks, exception context, and enterprise workflows.

Vooma and HappyRobot start closer to the execution layer.

They focus on operational communication: quote requests, carrier follow-ups, appointment coordination, tracking, POD collection, email, phone, and documents.

This is not glamorous work. But it is exactly where many operations teams lose time.

The interesting part is not that all these companies "have AI." It is that their products are moving toward the same workflow pattern:

  1. Pull context from systems, networks, emails, documents, or shipment data
  2. Detect an issue or request
  3. Decide what workflow should happen next
  4. Contact the right party
  5. Update the system of record
  6. Escalate and preserve an audit trail

That is very different from a dashboard. A dashboard asks the user to interpret and act. An execution layer starts taking pieces of the work off the user's plate.

Supply chain teams do not only struggle with lack of information. They struggle with fragmented follow-up.

A late shipment, missing POD, quote request, or appointment change is rarely one task. It is a chain of small tasks across carriers, facilities, customers, systems, documents, and billing.

This is why "agentic supply chain" should be evaluated workflow by workflow.

The buyer question is not: does this product have AI?

It is: which operational loop does this product actually close?

Can it move a late shipment from alert to resolution? Can it chase a missing POD without creating more work? Can it escalate when the action is outside policy? Can someone audit what the agent did afterward?

The best supply chain AI product is not the one that claims maximum autonomy. It is the one that knows where autonomy should stop.

Where autonomy should stop

I do not think the first useful supply chain agents will replace whole jobs.

They will take over the loops that are narrow enough to trust.

That is why project44's freight procurement agent caught my attention.

Not because it means every logistics role is about to be automated.

Because freight procurement has a shape that agents can work with.

There is a lane. There is available capacity. There is a set of carriers. There is market and contract context. There are guardrails. There is a recommendation or award. There should be an audit trail.

That is still real operational work.

But it is bounded enough to define what "good" looks like, what "unsafe" looks like, and when a human needs to step in.

Demand planning is harder in a different way.

The output may look like a number.

But the real work is often the reasoning around the number.

Why did the forecast move? Which customer are we protecting? Which node now carries the inventory risk? What assumption changed? Which tradeoff needs sales, finance, or operations in the room?

If an AI system jumps straight to the answer, it may remove the exact process artifact the organization needed.

That is where the coding-agent analogy starts to break.

In code, the end state can often be tested.

In planning and operations, the end state is only part of the work. The rationale, exception trail, escalation logic, and human commitments matter too.

So the question I would ask is not:

"Can AI replace this role?"

It is:

"Which loop is ready for autonomy?"

A procurement lane with clear constraints may be closer.

A demand planning decision that reallocates risk across customers, regions, inventory, and margin may not be.

That does not make AI less important in planning.

It changes the job of the AI.

In some workflows, the agent should close the loop.

In others, the agent should preserve the reasoning, surface the tradeoffs, and make the human decision better.

Execution needs an audit loop

Everyone is racing to show that AI can execute.

The more interesting shift is when it starts to audit.

That is why I paid attention to C.H. Robinson's June 3 launch of Lean AI Engineer.

It already had Lean AI Planner as the execution layer in Managed Solutions. The new Engineer layer is positioned to study network performance, identify improvement opportunities, and feed that intelligence back into the planner.

To me, that is the real story.

Most supply chain AI conversations still focus on speed.

Can it quote faster? Can it route faster? Can it tender faster? Can it clear exceptions faster?

Useful questions, but still execution questions.

The harder problem usually shows up one layer above execution.

Where is the recurring drift coming from? Which exceptions keep repeating? What should change before service actually degrades?

That is the beginning of an audit layer.

And anyone in logistics knows why that matters. The real ops tax is rarely one dramatic failure. It is the daily pileup of tender fallout, appointment misses, document rework, POD chase, and exception triage that quietly eats margin and time.

The architecture I keep coming back to is simple:

  1. Planner. Coordinates the workflow in real time.
  2. Executors. Handle the operational actions across shipments, documents, routing, and exceptions.
  3. Auditor. Studies outcomes, spots patterns, surfaces emerging risk or hidden savings, and pushes that learning back into the next cycle.

That third layer is where supply chain AI starts to get more valuable.

Not because it replaces operators.

Because it starts to encode what strong operators actually do: catch drift early, recognize repeat failure modes, and tighten the system before a KPI breaks.

Execution without feedback is just faster activity. Execution with an audit loop is where operational intelligence starts to compound.

The handoff is becoming the asset

The ocean carrier that moves your cargo now wants your warehouse too.

CMA CGM agreed on July 1 to acquire FedEx Supply Chain for $1.4 billion. According to the companies' official announcement, the deal would nearly triple CEVA Logistics' North American contract logistics operations and expand the combined footprint to about 150 warehouses, 20,000 people, and more than 240 locations.

The number is not the most interesting part.

The structure is.

This is what I think the market is really buying now:

  • ocean capacity
  • warehouse execution
  • tighter handoffs between modes
  • more control over where exceptions, delays, and margin leaks actually show up

CMA CGM and FedEx also said they expect multi-year ocean and air commercial agreements after the transaction.

That matters because the next logistics moat may not be a single transport asset.

It may be the handoff layer between ocean, air, warehouse, and final delivery.

Anyone who has worked across carrier management, 3PL operations, and customer service knows the same truth: the pain is rarely inside one node.

It shows up in the seams.

That is where context gets lost, ownership gets blurry, and service failures turn into cost.

If that trend keeps accelerating, companies still buying logistics mode by mode may find that their competitors are buying something more valuable: control over the handoff.

Where I would start

The first major AI impact in supply chain will not be fully autonomous planning. It will be removing thousands of small coordination loops between systems, teams, carriers, facilities, customers, and finance.

That is where the productivity is hiding.

For operators, the move is practical:

Pick one workflow. Map the decision path. Find where AI speed dies. Redesign that handoff.

So the question isn't where you add AI first.

It's which handoff should stop existing.

Name the one on your team that only survives because someone, once, wanted cover.

Sources

The Manifest · Every other week

Liked this? Get the next issue

Every other week I send one practical article on what holds up when AI meets real supply chain operations. Free.

EH
Esther Ho
AI × Supply Chain
All articlesFollow on LinkedIn