An AI tool helped people shopping for groceries on behalf of customers work faster, but the biggest gains did not come from replacing their experience. In a large US field experiment on the Instacart platform, real-time guidance helped workers locate products and plan their routes around stores. The peer-reviewed study in Information Systems Research found that experienced workers using AI for grocery workers were often better placed to take advantage of the technology. The result challenges a simple assumption: if a computer knows the route, a seasoned human's judgement must become less valuable.
The investigation was based on a randomised experiment involving nearly 6,000 shoppers across roughly 160 grocery stores in the United States, as discussed in an explanation by the study authors. The tool offered product-location advice and route guidance rather than carrying out the shopping independently. On average, the researchers reported better picking speed and fewer refunds for customers. But the detail that makes the story worth understanding is how differently workers used an apparently identical aid.
Why AI for Grocery Workers Needs Experience
Finding items efficiently may sound like a problem a computer should solve easily. A supermarket has aisles, shelves and a stock list, so why not calculate the shortest route and tell everyone where to go? Real stores are less tidy than that description. Goods move, displays change, popular items run out and the most direct path may not be the most practical one at a busy time.
An experienced shopper can recognise when an item is likely to have been moved or when the recommended aisle is blocked. They may know that an awkward product is usually found in an unexpected area, or that searching for a substitute requires checking its size and suitability rather than grabbing the nearest package. Those judgements come from repeated contact with real conditions. A route can be mathematically attractive and still be the wrong plan for the trip in front of the worker.
The Instacart tool supplied a useful layer of information about locations and picking order. The experiment indicates that it made work more efficient overall. According to the paper, average efficiency improved by 3.29% and service quality by 3.83% on its reported measures. Those findings describe improvements in picking speed and quality, respectively. The precise percentages are tied to the design and measurement of this experiment, not a promise that every shop or worker would receive the same improvement.
Experience Helped People Know When to Ignore the Screen
The researchers found that more experienced workers often combined the guidance with their own understanding of store conditions. Less experienced workers had more reason to rely on the suggested route, but might follow it when practical judgement called for a change. That matters because the ability to use information well is different from simply receiving information. An instruction can be correct in general and still need adjusting for the situation.
This suggests a model of human and machine working together that is less dramatic than full automation, but possibly more realistic. The computer can narrow down likely locations and reduce repetitive searching. The person can see the shelf, recognise the customer's order and deal with exceptions that the system may not have captured. Each contributes something different. The most useful result may come from recognising that difference instead of trying to remove it.
The study also found that algorithmic support could help workers with lower skill levels narrow some performance gaps. That is not the same as saying beginners suddenly became experts. The researchers specifically noted the importance of experience in gaining full benefit from the tool. Skill, experience and the difficulty of the particular task interact. There is little reason to expect one simple rule to describe every shop or every kind of work.
The Numbers Describe a Real but Bounded Experiment
Random assignment gives this investigation an important advantage over a survey asking workers whether they felt more productive. The researchers could compare groups given different access to guidance while examining measured behaviour on the platform. That offers stronger evidence of a causal effect in the studied setting than a collection of anecdotes about workers enjoying a new app.
The results nevertheless have boundaries. The experiment concerns an Instacart picking and navigation tool used by gig workers in US grocery stores. It is not an evaluation of general-purpose chatbots, nor does it prove that a fully autonomous robot could complete the same job. The platform, store information and type of order all shape what the tool can usefully recommend. A different system in a different workplace could perform differently.
The authors also reported an increase in work volume on the platform and greater flexibility to shop across stores. The original paper cited a 3.16% productivity-related volume increase and a 32.5% rise in cross-store shopping under its measures. Those figures do not tell us an individual worker's wages, expenses or well-being. Working across more locations might create opportunities, but its financial value to a person would require evidence about pay, travel and the demands of the assignments.
Research interpretation also benefits from acknowledging commercial relationships. One of the authors was associated with Maplebear, the parent company behind Instacart, and the experiment relied on access to the platform. That does not invalidate a peer-reviewed study, but it is a reason to examine the methods and outcomes rather than accept a favourable headline at face value. The interesting evidence is what the experiment measured, not a blanket endorsement of a company's use of AI.
Why Better Navigation Does Not Replace Common Sense
Consider an ordinary customer order. The list may request a specific type of fruit, a particular pack size and an ingredient suitable for an allergy. A route planner can identify a likely location, but it does not by itself decide whether a substitute is appropriate. A worker must understand the order's requirements and the customer's preferences. Efficiency without accurate fulfilment would not be much of an improvement.
The same principle applies to unexpected conditions. A display may be empty, a product misplaced or a queue blocking the best route. A tool that depends on stored information can become less useful as reality drifts away from its data. Workers with experience may notice those changes quickly, which helps explain why guidance and human knowledge can complement rather than cancel one another.
LiveAIWire has previously reported on AI changing the way people search for and buy products. That article concerns recommendations to consumers, a different problem from helping someone pick an existing order accurately. The distinction is useful: deciding what should go into a basket and finding the requested items inside a physical shop require different information and different kinds of judgement.
A further LiveAIWire report examined AI shopping agents and customer trust. For the person paying for groceries, confidence depends not just on a clever recommendation but on receiving the right things. A worker-support tool could improve that experience even if the customer never encounters its interface. The AI is assisting the human doing the shopping rather than asking the buyer to hand control of a purchase to a digital agent.
The Lesson Applies Beyond Supermarkets
Many workplaces combine repeatable routes or procedures with unpredictable exceptions. Technicians follow checklists but discover unusual faults. Warehouse staff work from locations while responding to changing stock. Office workers follow workflows while interpreting requests that do not fit the template. Guidance can remove routine searching in all these settings, but it may also create trouble if treated as an instruction that should never be questioned.
The useful question for employers is how to train people to interpret and challenge automated suggestions. A novice may need both route assistance and explicit guidance about exceptions. An experienced colleague may benefit from a quick overview without wanting the system to dictate every movement. Good design should make it easy to understand the recommendation and recognise when the user needs to override it.
There is a workforce dimension too. If software makes experienced workers more effective, it can reward knowledge rather than erase it. But if a platform measures only speed, it may overlook the judgement that prevented an incorrect substitution or a dissatisfied customer. Performance measures need to reflect quality, not just the number of items collected. The study's combination of efficiency and refund outcomes helps show why more than one measure is relevant.
LiveAIWire's coverage of AI and the changing value of human work raises a similar question. Does new technology make a job more manageable, or does it merely raise the pace expected of the worker? The Instacart study offers evidence of operational changes in one real setting, but it does not settle that wider question for every occupation. Worker experience, control and incentives will all shape the answer.
A Useful Kind of AI Does Not Need to Replace Anybody
The result is less theatrical than a fully automated supermarket, and perhaps more informative. The researchers tested a practical decision-support tool on actual shopping work and found measurable gains, particularly when the software's suggestions met the judgement of an experienced person. They also identified limits to how easily less experienced workers could benefit. That is a more nuanced picture than the idea that a clever algorithm instantly makes everyone equally capable.
For shoppers employed on platforms, the important issue is whether a tool helps them complete accurate orders without creating new frustration. For customers, it is whether the right groceries arrive. And for businesses, it is whether design and training allow workers to use guidance intelligently. The experiment suggests that human experience remains valuable not despite AI being present, but partly because someone needs to know how to use it well.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.