AI Fulfillment That Picks
87% More Items Per Hour
Agix built three integrated AI systems for Kroger, a multi-order pick path optimizer, computer vision item verification, and customer-preference substitution AI, transforming grocery e-commerce fulfillment across 2,700+ stores nationwide.
America's largest supermarket chain, where e-commerce growth made fulfillment efficiency critical.
The Kroger Co. is the largest supermarket chain in the United States, operating 2,700+ stores across 35 states under banners including Kroger, Ralphs, Fred Meyer, King Soopers, Smith's, Dillons, and Harris Teeter. Founded in 1883 in Cincinnati, Ohio, Kroger serves tens of millions of households every week with annual revenues of $148 billion.
As grocery e-commerce orders surged from 5% to over 25% of total sales in three years, in-store fulfillment became a critical operational and margin challenge. Picking grocery orders inside a live retail store, while shoppers browse the same aisles, is expensive, error-prone, and difficult to scale with manual processes. At millions of orders per week, inefficiency compounds into a margin problem that threatens the entire e-commerce business model.
Manual grocery fulfillment can't scale with millions of e-commerce orders per week.
When order pickers fill grocery orders manually, they walk whatever path they remember, pick items in whatever order seems natural, and make substitution decisions based on generic rules or personal judgment. At low order volumes this works. At millions of orders per week, it doesn't, inefficient paths, wrong substitutions, and missed items erode both economics and customer trust at scale.
Three AI layers. One fulfillment transformation.
Route intelligence, computer vision verification, and personalized substitution, working together to eliminate picking errors, reduce travel, and delight customers at millions-per-week scale.
Integrated fulfillment AI: route optimization, vision verification, smart substitution.
Agix built three integrated AI systems that work together to transform grocery fulfillment, each solving a distinct failure mode in the manual picking process, each measurable within the first week of deployment.
Multi-Order Pick Path Optimizer
Batches 3–5 simultaneous pickup orders with nearby windows and generates a single optimized pick path that visits each store zone once, covering all items across all batched orders with minimum travel distance. The route accounts for temperature zones, frozen and refrigerated items are collected last to minimize cold-chain exposure before staging.
Real-Time Inventory Integration
Live inventory data prevents pickers from being directed to out-of-stock shelf locations, routing directly to alternate in-store locations or triggering the substitution flow before the picker arrives at an empty shelf. This eliminates wasted walking to shelves that can't be fulfilled and removes a major source of picker frustration and delay.
Computer Vision Item Verification
A vision model running on the picker's handheld scanner verifies that the item scanned matches the order item, catching wrong-flavor, wrong-size, and wrong-brand errors before they reach the customer. In testing, 1-in-30 scans produced a wrong match; at millions of orders per week, this verification layer prevented hundreds of thousands of wrong-item errors per month.
Customer-Preference Substitution AI
When an item is out of stock, the substitution model queries the customer's 18-month purchase history to identify the most likely acceptable alternative, not just the nearest shelf item. A customer who always buys organic receives an organic substitute. A customer with a documented nut allergy never receives a product with nuts. The system generates a ranked list; if the first choice is also unavailable, the picker moves to the second-ranked option automatically.
Order Completion Prediction
Predicts likely completion time for each batch in real time, flagging orders at risk of missing their pickup window so managers can allocate additional picker support proactively. This prevents the cascading delays that occur when a slow pick causes a customer to wait, or worse, arrive to find their order isn't ready, and gives operations managers visibility across all active batches simultaneously.
Picker Performance Analytics
A real-time dashboard shows pick rate per hour, order accuracy, and substitution acceptance rate by picker, enabling targeted coaching and performance management grounded in actual data rather than manager observation. Stores with outlier performance in either direction are automatically surfaced for review, creating a continuous improvement feedback loop across the entire fulfillment operation.
Faster picks. Fewer errors. Customers who actually like their substitutions.
Measured against pre-deployment baselines across Kroger's e-commerce fulfillment network post-launch.
Our pickers used to dread substitutions because customers would reject them and we'd have to issue refunds. Now the AI knows this customer hates Red Delicious apples and always buys organic, it suggests the right substitute and customers appreciate it instead of complaining.
AI that made grocery fulfillment better for everyone, pickers, managers, and customers.
The pick path optimizer succeeded not because it was technically sophisticated, but because it was immediately tangible. Pickers walked 2 fewer miles per shift on day one. Adoption was nearly universal without mandates, the system sold itself through the physical experience of less exhausting work shifts.
The substitution AI succeeded for a less obvious reason: individual purchase history is the only reliable signal for what a specific customer actually wants as a substitute. Category-level rules fail at the individual level. Generic shelf-proximity rules fail completely. Personalization wasn't a nice-to-have, it was the entire difference between a 28% rejection rate and a 7.6% rejection rate.
Batch picking multiplied efficiency
Picking 4–5 orders simultaneously with a single optimized path was 3x more efficient per picker than picking one order at a time. This was a structural efficiency gain that no individual behavior change could achieve, the algorithm unlocked throughput that didn't exist in the manual process.
Personal history beats generic rules
Using 18 months of individual purchase history to generate substitutions, not category rules or shelf proximity, was the single most impactful customer satisfaction improvement. A customer who always buys Honeycrisp apples doesn't want Red Delicious. The data knew this; the old system didn't.
Vision verification scales inspection
1-in-30 scans produced a wrong-item match in testing, individually insignificant, but catastrophic at millions of orders per week. Vision verification turned an invisible quality problem into a caught-and-corrected problem before it reached the customer. Human inspection at this scale is impossible; automated verification isn't.
Immediate benefit drove adoption
Pickers walked 2 miles less per shift on day one. The efficiency improvement was so immediate and tangible, less physical exhaustion, faster order completion, that adoption was nearly universal without mandates. When AI makes the job easier in an obvious way, adoption is not a change management problem.
What this system doesn't solve.
E-commerce fulfillment AI is powerful for planned, repeatable operations. It has real limits at the edges of that structure.
Only Works for Planned E-Commerce Orders
The route optimizer requires advance order data to generate batch picks. Walk-in customers and last-minute orders can't be batched effectively, these are handled through conventional single-order picking. The system is designed for the pre-ordered pickup and delivery channel, not the in-store shopping experience.
New Customers Have No Substitution History
First-time and infrequent customers don't have enough purchase history for the substitution model to generate accurate personalized alternatives. These orders fall back to category-level substitution rules, better than shelf-proximity guessing, but not as accurate as the personalized model. The AI becomes progressively more useful for each customer as purchase history accumulates.
Store Layout Changes Require Manual Updates
Planogram resets and store layout changes require manual updates to the store zone graph before the route optimizer reflects new product locations accurately. During the window between a layout change and the zone graph update, pick paths may route to old locations. Store operations teams need a clear process for triggering zone graph updates when resets occur.
High-Judgment Items Still Need Human Discretion
Fresh prepared foods, deli counter orders, and items requiring quality assessment, produce ripeness, seafood freshness, butcher cuts, still require picker discretion that the system supports but cannot fully automate. The AI routes and verifies; the picker still needs to judge whether the avocado is actually ready to eat or whether the fish looks fresh enough to fulfill the order.
Is this the right build for your fulfillment operation?
What powers this system.
AI Predictive Analytics
Purchase history modelling, substitution preference prediction, and customer-level personalization, the data layer that powers smart substitution and customer satisfaction improvements.
AI Process Automation
Pick path optimization, multi-order batch routing, and fulfillment workflow automation, eliminating the manual coordination overhead that limits picker throughput at scale.
AI Computer Vision
Item verification scanning during pick operations, catching wrong-flavor, wrong-size, and wrong-brand errors at the point of scan before they reach the customer staging area.
Retail AI Systems
AI purpose-built for grocery and general merchandise retail, from e-commerce fulfillment and demand forecasting to shrink reduction and assortment optimization across large store networks.
Logistics AI Solutions
Last-mile delivery and fulfillment center optimization, the same route intelligence and completion prediction logic applied to dedicated fulfillment environments beyond in-store picking.
Operational Intelligence
Picker performance dashboards, order accuracy tracking, and substitution acceptance analytics, giving operations teams real-time visibility and targeted coaching capability across the fulfillment network.
Common questions about AI e-commerce fulfillment for grocery retail.
The route optimizer is temperature-zone aware. It sequences picks so that frozen and refrigerated items are collected last before staging to minimize temperature exposure time. The zone graph includes temperature classifications for each store section, and staging assignment routes cold items to refrigerated staging zones. This eliminates the common problem of ice cream and frozen goods sitting in a shopping cart for 45 minutes while a picker works through ambient aisles first.
The substitution model generates a ranked list of alternatives, not just a single recommendation. When the first choice is unavailable, the picker moves to the second-ranked option automatically, no additional decision required. If all suggested substitutes are unavailable, the item is flagged for removal from the order and the customer is notified via app with a refund option. The customer gets immediate transparency about what happened rather than discovering the issue at pickup.
Age-restricted items are flagged in the order and require ID verification at pickup. The system surfaces this requirement to the customer proactively in their pickup confirmation so they arrive prepared with valid ID. Pickers are prompted to verify that the ID check step will be required at handoff, preventing the awkward situation where a customer arrives for pickup and finds they can't receive part of their order without producing ID they didn't bring.
Yes, the same route optimization, vision verification, and substitution intelligence layers can be deployed in micro-fulfillment center environments. The zone graph structure and batch picking logic are applicable to any organized product storage environment, with configuration adjustments for the physical layout differences versus in-store settings. MFC deployments typically achieve even higher pick rates than in-store deployments because the environment is purpose-built for picking rather than shared with live shoppers.
The substitution model improves continuously as it collects more individual customer acceptance and rejection signals, every accepted or declined substitution tightens the model's understanding of that customer's preferences. The route optimizer improves as zone graph data is refined over time and as the system learns typical picker pace patterns by store zone. Day one performance is strong; day 90 performance is meaningfully better, and the gap between new-customer substitution quality and returning-customer substitution quality narrows as purchase histories grow.
Ready to transform your e-commerce fulfillment operations?
Most projects go from kickoff to deployed AI system in 8–16 weeks. Let's talk about what pick path optimization, vision verification, and personalized substitution could do for your fulfillment margins.