Closing the Supply Chain AI Execution Gap: What the 23% With a Formal Strategy Do Differently

The oddest number in supply chain AI right now is not a forecast for the size of the market. It is 23%. Gartner found that only 23% of supply chain organizations already deploying AI had a formal AI strategy, in a 2025 survey of 120 supply chain leaders.[1] That is not a survey of companies still wondering whether AI matters. These are organizations already in the game.

That is the execution gap in supply chain AI in 2026: ambition is no longer scarce, but operating discipline still is. More teams are funding pilots, adding optimization tools, experimenting with copilots, and asking planners to trust predictions. Yet too many of those efforts still sit outside the real planning, logistics, procurement, and warehouse workflows where margin, working capital, and service cost are actually made or lost.

The problem is not that supply chain AI lacks potential. The problem is that many companies try to scale it before they have standardized the work, embedded the output into decision systems, and defined the financial result they expect the initiative to move. When those three pieces are missing, the model may improve, the demo may impress, and the business may still struggle to prove what changed.

The investment story and the operating story no longer match

PwC’s 2026 Digital Trends in Operations Survey makes the contradiction harder to dismiss. Among 767 US operations leaders, 89% said technology investments had not fully delivered expected results, and 87% said poor data quality had hampered value from digital initiatives. Only 4% said AI was fully embedded with no barriers to scaling.[2]

Those numbers are useful because they come from operations leaders, not just technology sponsors. They describe the layer where AI leaves the slide deck and meets item masters, transport exceptions, supplier lead times, warehouse labor constraints, planning calendars, approval rules, and the everyday habit of exporting a file because the system does not quite support the decision.

BCG’s logistics research, cited in Supply Chain Management Review, points to a similar pattern: roughly 40% of logistics providers and shippers identified unclear ROI as the top barrier to AI adoption, while only 13% of logistics service providers and 7% of shippers could point to measurable value.[3] That is not a capability shortage by itself. It is a traceability problem. If a team cannot connect the AI intervention to a changed decision, and then connect that decision to cost, revenue, inventory, or service, the ROI discussion becomes a debate instead of an operating fact.

There is also a timing issue worth treating honestly. Some AI returns in supply chain take years, not quarters. A network planning model, a replenishment transformation, or an AI-supported procurement workflow may need several planning cycles before the business sees a clean financial pattern. Slow ROI is not automatically failure. But slow ROI with no baseline, no owner, no workflow change, and no P&L bridge is usually not patience. It is drift.

The first difference: disciplined teams standardize the work before they automate judgment

AI exposes process variation faster than most organizations expect. A forecasting model trained across business units may discover that one team defines lost sales differently, another overrides demand at a different level of hierarchy, and a third treats promotional uplift as an offline spreadsheet adjustment. A procurement model may score suppliers using fields that are mandatory in one region and optional in another. A warehouse labor model may be asked to predict productivity while shift coding, task interleaving, and exception handling vary by site.

In those conditions, AI does not create one better operating system. It often becomes another layer of reconciliation. The team spends its time explaining why the recommendation is wrong for this lane, that supplier, this customer class, or that plant calendar. The fix is not to ask data science to smooth over every process difference. The fix is to decide which differences are legitimate operating realities and which are just local workarounds that should never have become permanent.

This is where the 23% strategy gap matters. A formal AI strategy is not valuable because it exists as a document. It is valuable only if it forces uncomfortable choices: which planning level is the system of record, which master data fields are governed, which exceptions planners are expected to handle manually, which decisions can be automated, and which performance metrics determine whether the model survives past the pilot.

The companies that move faster later often move slower here. They do the unglamorous work first: clean up SKU-location logic, align planning calendars, remove duplicate exception codes, define forecast consumption rules, standardize supplier lead-time treatment, and decide how overrides will be captured. None of that sounds like AI. All of it determines whether AI can be trusted inside the business process.

Poor data quality is often discussed as if it were a database problem. In supply chain, it is usually a process ownership problem. PwC’s finding that 87% of surveyed operations leaders said poor data quality hampered value from digital initiatives should not be read as a call for another data lake by default.[2] It is a sign that operational definitions, handoffs, and accountabilities are still too loose for AI to produce repeatable business value.

The second difference: AI sits inside the workflow, not beside it

A supply chain AI pilot can look successful while still being operationally irrelevant. The model predicts late orders. The dashboard ranks inventory risk. The copilot summarizes supplier exposure. Then the planner opens the same ERP screen, the transportation analyst works the same tender queue, the buyer follows the same approval chain, and the warehouse supervisor runs the same labor meeting. If the AI output does not change the sequence of work, it is decoration with a subscription fee.

Embedding AI into workflow means the recommendation appears where the decision is already made, at the moment the decision is made, with enough context for the user to act. It also means the system captures what happened next. Did the planner accept the replenishment recommendation? Did they override it? Why? Did the shipment move to another carrier? Was the accessorial avoided or merely shifted? Did the supplier risk alert trigger a dual-source action, an expedite, a commercial conversation, or nothing?

That feedback loop is not a user-experience nicety. It is how the organization learns whether the model is changing behavior. Without it, AI teams measure model performance while operators measure service failures, expediting cost, schedule adherence, and inventory. Two scoreboards emerge, and the business eventually trusts the one tied to the morning meeting.

The workflow test is simple: if the AI system disappeared tomorrow, which operating step would stop, slow down, or become visibly worse? If the honest answer is “nothing, but we would lose a useful dashboard,” the tool has not yet entered the operating system of the business.

This is also where formal change management earns its keep. Training people on a model is not the same as redesigning the work around a new decision path. The practical questions are more specific: who receives the recommendation, what authority do they have, what threshold triggers escalation, what evidence is required to override, who reviews repeated overrides, and which KPI changes if adoption improves?

A planner should not have to become a part-time model auditor just to use AI. A buyer should not have to copy a recommendation from one platform into another and then defend it in a meeting where the old spreadsheet still carries more authority. When AI runs beside the workflow, the burden of integration falls on the user. When it runs inside the workflow, the process itself changes.

The third difference: P&L metrics are defined before the pilot starts

Supply chain teams do not need every AI initiative to produce instant savings. They do need to know which financial mechanism is supposed to move. “Better visibility” is not a P&L metric. “Improved decisions” is not a P&L metric. “Fewer manual touches” may matter, but only if the labor, service, or cycle-time consequence is defined.

A credible AI use case should be tied to one or more hard outcomes before model selection begins:

• Margin: fewer expedites, better price realization, lower premium freight, improved procurement savings capture, reduced waste or obsolescence.
• Working capital: lower inventory without unacceptable service degradation, better allocation of constrained stock, improved safety stock placement, faster release of excess.
• Service cost: fewer stockouts, better on-time performance, lower detention and accessorial exposure, reduced manual rework, fewer avoidable exception touches.

The baseline has to be just as explicit. If the initiative is meant to reduce expedites, define which expedite categories count, which business units are in scope, and what comparison period is valid. If the goal is inventory reduction, define the service constraint and the inventory bucket. If the model is intended to improve procurement spend, separate negotiated savings from realized savings. Otherwise the pilot can always claim directional success while finance sees no clean movement.

This is the reason unclear ROI keeps showing up as a barrier in logistics AI adoption.[3] Many pilots are designed around technical feasibility first and financial attribution later. By the time the team asks whether the tool paid back, the baseline is muddy, the process has changed in several other ways, and the business sponsor has moved on to the next initiative.

For a deeper treatment of the measurement problem, see What the Numbers Actually Say About AI ROI in Supply Chain. The short version is that the financial case must be designed into the operating change, not reconstructed afterward.

Where pilots usually break

Most stalled supply chain AI efforts do not fail in a dramatic way. They fade. A promising forecast model remains in one region because other regions use different planning assumptions. A logistics exception tool is used by analysts who like it, ignored by those who can still clear their queue faster in the old system, and never reconciled with carrier management. A procurement risk model produces useful alerts, but nobody changes sourcing policy, contract review cadence, or supplier development priorities.

The pattern is familiar:

1. The pilot is scoped around a visible pain point but not a governed process.
2. The data team builds a model using what is available rather than what the decision truly requires.
3. The tool is demonstrated outside the system where planners, buyers, dispatchers, or supervisors work.
4. Adoption is measured as usage, not as changed decisions.
5. ROI is discussed after the fact, with finance asked to validate a benefit the process was never instrumented to capture.

None of those failures mean the algorithm was useless. They mean the deployment was incomplete. The model was asked to create value in an operating environment that had not been prepared to absorb it.

This is why broad AI adoption intent should be treated carefully. Intent signals that leaders see the opportunity. It does not prove readiness, effectiveness, or financial impact. A company can have a dozen pilots and still lack one production workflow where AI reliably changes a decision that finance can trace.

What the minority are doing differently

The minority of companies getting measurable value are not simply buying better tools. They are making different management choices around the tools. They narrow the scope, define the financial mechanism, clean up the decision process, and put AI where work happens.

In practice, that looks less like a grand transformation announcement and more like a controlled operating redesign:

• They choose a decision with economic weight, such as replenishment quantities, carrier tendering, production sequencing, supplier risk response, or inventory positioning.
• They standardize the inputs and rules around that decision before asking AI to optimize it.
• They embed recommendations into the existing planning, execution, procurement, or warehouse system rather than adding a separate analytical destination.
• They assign decision rights: who can accept, reject, override, escalate, or automate.
• They capture the reason for overrides and review repeated exceptions as process signals, not user resistance by default.
• They track financial outcomes against a baseline agreed with finance before launch.

That last point is the one that keeps pilots honest. A demand-sensing tool should not be judged only on forecast accuracy if the business case was lower inventory and better service. A logistics AI tool should not be judged only on predicted delay accuracy if the promised outcome was lower premium freight or fewer missed customer appointments. A procurement model should not be judged only on supplier risk classification if the business case depends on avoided disruption, lower spend, or better compliance.

For teams stuck between promising pilots and production scale, Why Most AI Supply Chain Planning Pilots Stall — and the Methodology That Scales Them is the useful companion question: not whether the model works in isolation, but whether the organization has designed the path from recommendation to adopted decision.

The value is real, but it is not automatic

There are enough credible examples and analyses to accept that AI can improve supply chain economics when applied well. McKinsey has estimated that AI-enabled distribution can reduce logistics costs by 5–20%, inventory by 20–30%, and procurement spend by 5–15%.[4] Those ranges are useful as a reminder of the prize, not as a promise that every pilot deserves funding until it finds its way.

Vendor case studies can also be useful, especially when they show the process changes behind the result. But the headline metric is rarely the whole story. A large reduction in forecast error, for example, may depend on data maturity, product mix, demand volatility, planning discipline, and the willingness of the business to change how it uses the forecast. Treat those cases as evidence that value is possible, not as a median outcome.

Market-size optimism deserves the same treatment. It may explain why boards, investors, and executives are paying attention. It does not tell an operations leader whether a planner will adopt a recommendation, whether a warehouse supervisor will change labor deployment, or whether finance will see the savings hit the right line.

A practical test before the next AI rollout

Before approving another supply chain AI initiative, leadership can ask a few questions that cut through most of the theater:

• Which exact decision will change?
• Where is that decision made today, and will the AI recommendation appear inside that workflow?
• Which process inputs must be standardized before the recommendation can be trusted?
• Who owns the decision rights, override rules, and exception review?
• Which P&L or working-capital metric should move, and what is the baseline?
• How will the organization know whether the model changed behavior, not just whether users opened the tool?

If those questions cannot be answered, the initiative is not ready to scale. It may still be worth exploring, but it should be funded and governed as exploration, not presented as transformation.

The execution gap in supply chain AI is not mysterious. Gartner’s 23% strategy figure shows that many organizations deploying AI have not yet built the management system around it.[1] PwC’s data shows that technology investment, data quality, and embedded adoption are still blocking value for a large share of operations leaders.[2] BCG’s logistics findings show how quickly the conversation turns to unclear ROI when measurable value is not built into deployment.[3]

Slow ROI is tolerable when the operating path is clear. No ROI is what happens when AI is scaled on top of unstable processes, parked outside the workflow, and justified with benefits nobody can trace. The minority succeeding in 2026 are not luckier. They are more disciplined.

References

1. Gartner Survey Shows Just 23% of Supply Chain Organizations Have a Formal AI Strategy — Gartner, June 11, 2025.
2. 2026 Digital Trends in Operations Survey — PwC.
3. AI in the supply chain: From pilot programs to P&L impact — Supply Chain Management Review.
4. 2026: The age of the AI supply chain — Supply Chain Management Review.