Walmart’s 2025 delivery playbook—plus the AI layer

Walmart reported a 50% increase in U.S. store-fulfilled delivery sales in 2025. That’s not a “nice improvement.” That’s a signal that the winning logistics model for retail is getting clearer: stores, micro-fulfillment, and smarter orchestration are beating the old “ship it from a distant DC and hope for the best” approach.

What I like about Walmart’s 2025 delivery upgrades is that they’re not just flashy pilots. They’re a portfolio: drones for specific lanes, dark stores for dense demand, same-day coverage from stores, and order management integrations that make it easier to scale. If you run transportation, supply chain, or procurement, this matters because it shows where the cost and service curve is heading—and what capabilities you’ll be expected to match.

This post is part of our AI in Transportation & Logistics series, so I’m going to do two things:

1. Break down five moves Walmart made in 2025 to accelerate delivery.
2. Add the missing layer most companies still underinvest in: AI-driven optimization across routing, inventory placement, labor, and supplier alignment.

1) Store-fulfilled delivery scaled because stores are the new nodes

Answer first: Walmart grew fast delivery by treating stores like a distributed fulfillment network, not just a place shoppers walk into.

The headline number—50% growth in U.S. store-fulfilled delivery sales—points to a practical truth: proximity wins. A store within a few miles of the customer can beat a regional DC on speed and often on cost, especially when you can consolidate picking, packing, and dispatch in a repeatable way.

But store fulfillment is messy when it’s managed like an add-on. You’re balancing:

• Store labor vs. picking workload
• Shelf availability vs. promised inventory
• Substitution rules vs. customer satisfaction
• Delivery windows vs. driver supply

Where AI makes store fulfillment actually scalable

This is where AI in logistics stops being theoretical.

• Demand forecasting at hyperlocal level: Store-level forecasts that understand neighborhood seasonality (December peaks, weather swings, local events) reduce substitutions and late orders.
• Real-time inventory accuracy: Computer vision and anomaly detection can flag “phantom inventory” (system says 5, shelf says 0) before it turns into missed SLAs.
• Pick-path optimization: Small minutes add up. AI can optimize pick routes inside the store and batch orders intelligently.

A stance I’ll take: most retailers don’t have a “delivery speed problem.” They have an orchestration problem—and AI is the fastest path to fixing it.

2) Dark stores are a blunt tool—unless you use data to place them

Answer first: Dark stores speed up delivery because they remove shopper interference, but the economics only work if you place them using demand density and SKU strategy.

Walmart’s test of dark stores (fulfillment locations not open to the public) is a classic move when store picking starts to hit limits. Dark stores can improve:

• Pick speed (no navigating customers)
• Inventory reliability (more controlled replenishment)
• Order quality (less substitution)

The risk is also obvious: if you put a dark store in the wrong place or stock it poorly, it becomes an expensive warehouse with short hours.

The AI decision layer: “Where should a dark store exist?”

Dark store placement should be a network design problem, not a real estate experiment.

AI and optimization models can evaluate:

1. Demand heatmaps by time of day and day of week
2. SKU velocity curves (which SKUs drive most orders and which drive most margin)
3. Service-time constraints (same-day vs. 2-hour windows)
4. Last-mile cost curves (driver pay, congestion, failed delivery rates)

A practical rule: if you can’t quantify how a dark store reduces cost per delivered order while protecting fill rate, you’re not running a network strategy—you’re running a pilot.

3) Drone delivery is real now—but it’s a lane strategy, not a blanket strategy

Answer first: Drones help where the constraints are clear (short distance, lightweight items, high urgency), and Walmart’s 2025 expansions show it’s moving from novelty to operations.

Walmart expanded drone delivery and launched service in metro Atlanta, with plans to expand to additional cities. That’s not about replacing vans. It’s about carving out specific deliveries where drones can win:

• Urgent, lightweight items (health and wellness, small household needs)
• Short-distance drops where road travel time is unpredictable
• Scenarios where customers value speed enough to justify cost

The AI layer: make drones part of the same planning brain

The biggest missed opportunity I see in drone programs is running them as separate operations. If drones are a service option, they should be orchestrated by the same intelligence that runs everything else.

AI can:

• Predict drone-eligible orders before checkout (by basket weight, destination, promised window)
• Select mode dynamically (drone vs. driver vs. next-day) based on cost, capacity, and SLA risk
• Manage exception handling (weather, airspace restrictions, battery constraints) and re-route automatically

If you’re thinking about drones (or autonomous delivery of any kind), the procurement angle matters too: you’re buying into a new supplier ecosystem—aircraft, maintenance, software, compliance, insurance—and it needs a governance model.

4) Same-day reach depends on contractors—so treat driver supply like a market

Answer first: Walmart’s scale in last mile relies on independent contractors (like Spark), which means capacity is volatile and must be managed with pricing, incentives, and prediction.

Walmart has thousands of independent contractors on its Spark Driver platform. This is the reality of modern last-mile delivery: capacity comes from flexible labor markets, and those markets move.

Peak season (hello, mid-December) makes this painfully visible:

• Order volumes spike
• On-time expectations tighten
• Driver availability competes with other gig platforms

AI for last-mile capacity: reduce cancellations and late deliveries

AI helps when you stop treating dispatch as a static routing problem.

• Predict no-show/decline risk by zone, time slot, incentive level, and order complexity
• Dynamic incentive pricing to stabilize acceptance rates during peaks
• Route construction that considers human behavior (drivers prefer certain store locations, drop types, apartment deliveries vs. houses)

This is also where a lot of companies mis-measure performance. They focus on average delivery time and miss the metric that customers actually feel: promise accuracy (did you deliver when you said you would?). AI improves promise accuracy by quantifying risk at the order level.

5) Order management integration is the quiet move that makes everything else work

Answer first: Integrating delivery services into order management systems turns logistics speed into a repeatable capability instead of a patchwork.

Walmart GoLocal integrated with IBM’s order management system, aiming to make it easier for retailers to implement white-label delivery. On the surface, that’s a tech partnership story. Operationally, it’s the “glue” move.

Because faster delivery isn’t one initiative. It’s dozens of decisions across:

• Inventory sourcing (which node fulfills?)
• Promising (what window can we commit to?)
• Carrier selection (who delivers?)
• Exceptions (what happens when something breaks?)

The AI layer: smarter orchestration in the OMS/TMS stack

When AI sits on top of order management and transportation management, you can automate decisions humans can’t make fast enough:

• Multi-objective fulfillment optimization: minimize cost and maximize on-time probability and protect margin
• Inventory rebalancing triggers: identify when a zone is trending toward stockouts and move product earlier
• Supplier-aware ETA predictions: incorporate inbound variability so your customer promises don’t ignore upstream reality

If your OMS can’t support real-time decisioning and feedback loops, your delivery strategy will plateau—no matter how many pilots you launch.

What Walmart’s 2025 lessons mean for procurement leaders

Answer first: The next delivery advantage comes from procurement enabling data, flexibility, and shared metrics—not just negotiating rates.

A lot of delivery “transformation” gets stuck because procurement is asked late—after operations has chosen platforms, signed pilots, and created dependencies.

Here’s what I’ve found works better: procurement sets the conditions for scale.

Procurement checklist for AI-enabled delivery performance

• Data rights and interoperability: Can you access event-level data (pick time, handoff time, delivery proof, exception codes)? If not, you can’t improve with AI.
• Outcome-based SLAs: Pay for on-time performance, promise accuracy, and exception resolution—not vanity metrics.
• Model risk controls: If you’re using AI for dispatch or promise times, define auditability, bias checks (e.g., incentive fairness by zone), and human override rules.
• Supplier ecosystem strategy: Drones, gig platforms, and OMS integrations create lock-in. Manage it intentionally.

The punchline: if procurement only optimizes unit cost, it will accidentally increase total cost by driving late deliveries, substitutions, and churn.

A practical 90-day plan to apply the “Walmart playbook + AI”

Answer first: You don’t need Walmart’s budget; you need Walmart’s sequencing—instrument first, optimize second, automate third.

If you’re building an AI-optimized supply chain for last mile, here’s a realistic 90-day path:

1. Weeks 1–3: Instrument the workflow

   • Standardize event timestamps across pick/pack/handoff/delivery
   • Create a single exception taxonomy (late pick, driver decline, inventory short, address issue)

2. Weeks 4–7: Build “promise accuracy” as a core metric

   • Track: promised window vs. actual arrival
   • Segment by zone, store, carrier/driver type, and SKU category

3. Weeks 8–10: Deploy AI where it’s easiest to win

   • Predict late orders and re-route early
   • Improve inventory accuracy for top 200 velocity SKUs
   • Optimize batching and pick paths for peak hours

4. Weeks 11–13: Lock in procurement and governance

   • Update SLAs to reflect promise accuracy and exception handling
   • Add data access clauses and model transparency requirements

This isn’t flashy. It is effective.

Where delivery is heading in 2026: speed is table stakes, orchestration is the moat

Walmart’s 2025 delivery upgrades show the direction of travel: distributed fulfillment + multiple delivery modes + strong orchestration tech. Competitors can copy pieces of that. The harder part to copy is the decision engine that ties it together.

For teams working in transportation, logistics, and procurement, the opportunity is straightforward: build an AI capability that improves promise accuracy, lowers cost per delivered order, and reduces exceptions. That’s what customers feel, and that’s what boards fund.

If you’re planning your 2026 roadmap now, here’s the question I’d put on the first slide: Which decisions in our last-mile operation are still being made by habit—when they should be made by data?