Are grocers investing in the right AI?

Back-end AI offers grocers more practical value today than consumer-facing AI

Albertsons shows where AI can create value in grocery retail. At the end of last year, the company launched an agentic shopping assistant. A customer can enter a broad request such as wanting spicy food, and the system can recommend recipes and add the required products to an online basket. Five months later, Albertsons introduced a less visible AI tool. Warehouse employees can use the proprietary system to determine whether pallets of fresh strawberries meet quality standards.

The second use case points to the stronger near-term opportunity. Grocery is a low-margin, operationally complex business. Small errors in fresh-food ordering, inventory, pricing, promotions and labor allocation can quickly consume profit. AI that reduces those errors has a clearer route to financial value than a shopping assistant whose effect on incremental sales or customer retention may be harder to prove.

This distinction should shape investment priorities. Consumer AI can improve discovery and personalization, but operational AI addresses costs and execution problems that grocers already know they have. These include shrink, labor costs, inaccurate inventory, complex promotions and the challenge of setting the right price at the right time. The business case starts with a measurable operating problem.

Curt Prins, Senior Product Manager at Albertsons and a former Kroger employee, takes this position explicitly. He argues that back-end AI can generate a better return on investment than more visible consumer applications. His central point is simple: retailers should focus on impact rather than the latest attention-grabbing technology.

There is an important constraint. Operational AI only creates value when it changes a decision or process in a measurable way. A more accurate forecast has limited value if ordering rules, employee workflows or supplier processes prevent the retailer from acting on it. Executives should therefore evaluate AI against operating metrics such as waste, availability, labor hours, forecast error, gross margin and promotion performance. The relevant question is not whether an AI model works. It is whether the business performs better because of it.

This does not make consumer AI unimportant. Agentic shopping assistants may become valuable as customers become more comfortable delegating product discovery and basket creation to software. But grocers have immediate operational problems with established economic costs. Those are the rational place to concentrate AI investment today.

The fastest route to AI returns is to target specific operational pain points

Grocers do not need to redesign the entire company before putting AI to work. Several retailers are already applying it to narrow, expensive problems. Hy-Vee partnered with Relex to improve forecasting, ordering and replenishment for fresh products. Heritage Grocers Group is using an AI-powered system to scale pricing promotions. Grocery Outlet is incorporating technology from Afresh to improve ordering across departments.

The range of applications is wider than inventory management. Retailers can license AI systems for retail media, circulars, planograms, pricing and stock management. Kroger uses Sage, a virtual assistant that helps employees manage schedules and store tasks. These examples share an important characteristic: the technology is attached to a defined business process.

That is the right deployment model for near-term AI investment. Start with a costly decision that is repeated frequently and supported by sufficient data. Fresh-food ordering is a strong example. Ordering too much creates waste and markdowns. Ordering too little creates out-of-stocks and lost sales. Better forecasts can directly improve the decision. The financial result can then be measured against an existing baseline.

Curt Prins, Senior Product Manager at Albertsons, summarizes the principle as: “AI is a capability, not a strategy.” That distinction matters at the executive level. Buying AI software does not define a business objective. Reducing fresh-food waste by improving order accuracy does. The technology should be selected only after management defines the operating result it expects.

A focused approach also makes ROI easier to test. Executives can establish a baseline, deploy the system in a controlled part of the operation, and measure changes in cost, availability, productivity or margin. A tool that produces no material improvement should not receive broader investment. One that produces repeatable gains can be scaled into additional stores, departments or workflows.

The main bottleneck is therefore not access to AI. Commercial systems already cover many grocery functions. The harder task is identifying where better predictions or faster decisions will materially affect economics, then changing the surrounding process so employees can use those outputs.

There is also a sequencing issue. Retailers should not scatter disconnected AI tools across departments simply because each solves a local problem. Point solutions can produce early gains, but too many isolated systems create integration and data problems later. Near-term pilots should fit a broader plan for data, workflows and system architecture.

For C-suite teams, the investment rule should be clear: fund the problem, not the technology. Define the operating metric first. Establish its current performance and economic value. Apply AI where it can change that metric. Then scale only what produces measurable results.

Smaller grocers can use AI without building large in-house technology teams

Walmart and Amazon have major internal AI capabilities. Regional chains and independent grocers do not need to copy that model. Many operational AI capabilities are now available through specialist vendors, cloud platforms and existing enterprise software. Smaller retailers can buy the capability they need instead of building the underlying technology themselves.

Curt Prins, Senior Product Manager at Albertsons and a former Kroger employee, makes the point directly: “[Grocers] don’t need a 2,000-head technology team to do this.” He recommends smart partnerships focused on measurable impact rather than highly visible AI projects. For smaller grocers, this changes the investment decision. The relevant choice is often not build versus do nothing. It is which external capability can solve a specific problem at an acceptable cost and integration burden.

Operational applications are particularly suitable for this approach. A grocer can use external technology to improve ordering, forecasting, pricing, promotions or waste management without developing proprietary AI models. This can narrow some of the operational advantages held by larger competitors. Better fresh-food ordering, for example, can reduce waste while protecting product availability. More accurate pricing and promotion systems can improve execution without adding equivalent manual work.

Tom Furphy, CEO and Managing Director of venture capital firm Consumer Equity Partners and a former Amazon and Wegmans executive, argues that operational AI can make independent grocers and small chains more nimble. This matters because size alone does not determine execution quality. A smaller company with modern tools and simpler processes can improve decisions faster than a larger organization constrained by old systems and workflows.

Bobby Gibbs, Principal at management consulting firm Oliver Wyman, adds an important distinction between technology investment and organizational adoption. He cites a five-store retailer whose CEO requires employees to use Microsoft Copilot in their daily work. As Gibbs explains, “There’s a level of capability that does require major capital investment, but there’s also some cultural change that doesn’t require that kind of investment.”

Executives should still be selective. Buying third-party AI transfers some development work to the vendor, but it does not remove integration, governance or operating risks. Management needs to know who owns the data, how the system integrates with existing applications, how outputs are validated and how pricing changes as usage grows. Dependence on a vendor can also become expensive if a tool becomes embedded in a critical process.

The practical strategy for smaller retailers is therefore targeted adoption. Use partners where the technology is becoming standardized. Reserve internal development for capabilities that provide real competitive differentiation or require proprietary knowledge. Measure results against concrete metrics such as shrink, waste, availability, labor productivity and margin. Limited technology resources make this discipline more important, not less.

Data quality is the main constraint on scaling operational AI

An AI system cannot compensate for unreliable operating data. If inventory records, product attributes, pricing information or transaction histories are incomplete or inconsistent, the system can produce recommendations that look precise but are operationally wrong. Adding more AI on top of poor data can increase the speed and scale of those errors.

The risk is easy to see in personalization. Curt Prins, Senior Product Manager at Albertsons, notes that a retailer might use AI to send a customer a highly relevant personalized promotion. If inventory data is wrong and the advertised product is unavailable, the personalization has failed. Worse, the retailer has created a customer experience problem by promoting something it cannot sell.

The same issue applies to back-end decisions. Forecasting depends on clean sales and inventory histories. Pricing requires reliable information about products, costs, promotions and demand. Automated ordering needs accurate stock positions and lead times. As AI systems begin making decisions across several functions, inconsistencies between those datasets become more consequential.

Prins therefore warns that grocers will struggle to capture the full benefit of AI if their tools depend on fragmented or incomplete company data. His broader statement—“AI is a capability, not a strategy”—is relevant here as well. A retailer cannot solve weak information management simply by buying a more sophisticated model.

Tom Furphy, CEO and Managing Director of Consumer Equity Partners, recommends improving the data foundation while still allowing controlled AI adoption. “You don’t necessarily have to wait completely until your data is fixed or is ready, but don’t go too far with AI on top of that data because it’s going to give you bad signals,” he said.

That creates an important distinction for executives. Perfect data should not become a prerequisite for every pilot. Most large retailers have data defects, duplicated records and legacy systems that will take years to address completely. Waiting for a comprehensive cleanup could delay useful deployments with clear business cases. The better approach is to establish whether the data required for each use case is sufficiently accurate and complete for the decision being automated.

Scaling requires a higher standard. An isolated AI application may depend on a limited dataset and tolerate manual review. Agentic systems can draw information from multiple functions and act with less human intervention. A pricing error, inaccurate inventory record or inconsistent product identifier can therefore affect several connected decisions. Data governance becomes an operating requirement rather than an IT housekeeping exercise.

C-suite leaders should assign clear accountability for critical datasets. Management needs agreed definitions, ownership, quality thresholds and controls for how information moves between systems. AI performance should also be monitored against actual business outcomes, not only model-level technical measures.

The priority is not to make every dataset perfect. It is to make the data behind high-value decisions reliable enough that management can trust the resulting actions. That provides a practical basis for expanding AI without allowing poor information to undermine the economics of the investment.

Agentic AI could shift grocery retail from task automation to connected decision-making

Most AI deployments in grocery today solve a defined task. They forecast demand, recommend prices, improve orders or help employees manage work. Agentic AI goes further. An AI agent can gather information from multiple systems, evaluate options and take actions within rules set by the retailer. This moves AI from producing recommendations toward participating directly in operating decisions.

Walmart is already applying this approach to inventory. The retailer uses agentic AI to create a unified view of inventory across stores and supply facilities. That matters because inventory decisions rarely exist in isolation. Store demand, warehouse availability, replenishment schedules and supply constraints all affect whether products are available where customers want them.

Category management is another area likely to change. Bobby Gibbs, Principal at management consulting firm Oliver Wyman, says agentic technology could improve vendor negotiations by gathering data from across the enterprise and testing several scenarios. A category manager could assess different combinations of pricing, promotional terms, volumes and other commercial variables without manually assembling each analysis.

Gary Hawkins, Founder and CEO of the Center for Advancing Retail & Technology, expects advanced AI to automate activities such as assortment planning and price optimization within parameters established by category managers. This is an important distinction. Agentic retail does not necessarily remove managers from decisions. It changes where their time and judgment are applied. People can define objectives, constraints and exceptions while AI handles more repetitive analysis and execution.

Hawkins describes the change as retail moving “from human speed to machine speed.” For executives, faster execution should not be treated as the objective by itself. A system that makes poor decisions faster creates more risk. The value comes from combining speed with reliable data, clear constraints and measurable business objectives.

That creates a governance requirement. As AI receives greater authority, retailers need explicit rules covering which decisions an agent can make independently, which require human approval and what conditions trigger escalation. Price changes, purchase commitments and vendor decisions can carry material financial consequences. Audit trails and the ability to identify why an action occurred will become increasingly important.

Agentic AI could also change the economics of managerial work. Employees who spend significant time gathering information, preparing routine analyses and executing recurring decisions could shift toward exception management, commercial judgment and setting operating rules. That creates an opportunity to increase managerial capacity without simply increasing headcount.

The near-term priority is readiness rather than maximum autonomy. Retailers should identify decisions that are frequent, measurable and sufficiently constrained. They can then increase automation as data quality, system reliability and management confidence improve. The companies that build this discipline early will be better positioned as agentic capabilities mature.

Agentic retail requires connected data and a different operating model

Technology alone will not deliver agentic retail. The central constraint is organizational fragmentation. Grocery companies typically distribute responsibility across merchandising, pricing, supply chain, store operations, marketing and other functions. When those groups use separate data, systems and approval processes, an AI agent cannot easily optimize decisions across the whole business.

Gary Hawkins, Founder and CEO of the Center for Advancing Retail & Technology, argues that retailers need a “data fabric” connecting information across the enterprise. In practical terms, this means making relevant data accessible and consistent across business functions. Inventory, pricing, product, customer, promotion and supply information must be usable together when a decision requires it.

The objective is not to place all information into one physical database. Modern enterprises will continue to operate multiple systems. The business requirement is consistent definitions, reliable interfaces, appropriate permissions and clear ownership. An agent calculating a pricing action, for example, needs confidence that product costs, inventory levels and promotion data represent the same products and the correct time periods.

Organizational design must change alongside the data architecture. Capri Brixey, Partner at consultancy The Partnering Group, says traditional top-down structures may become cumbersome as AI accelerates decision-making. She notes that many companies are moving toward matrixed organizations, where authority and expertise are distributed across functions rather than routed exclusively through a single hierarchy.

This does not mean every retailer should reorganize around AI. Structural change should address a specific constraint. If AI generates a useful recommendation in seconds but implementation requires several departments and multiple approval levels, the company has improved analysis without improving execution. Executives need to identify where governance is necessary and where approval processes merely create delay.

Brixey says, “The companies seeing the best outcomes are investing as much in enablement and process redesign as they are in technology.” That is a useful allocation principle. AI budgets should account for workflow redesign, employee training, data integration and performance measurement.

Senior leadership is critical because these changes cross functional boundaries. Hawkins argues that AI transformation “has got to be driven by the CEO. It’s got to be driven by the boards. It’s got to be driven by the owners.” Dollar General and Ahold Delhaize have also recently hired leaders for AI-specific roles, showing that retailers are beginning to formalize responsibility for the technology.

An AI executive alone, however, cannot resolve conflicting incentives between business units. C-suite teams must set shared objectives and decide which outcomes matter at enterprise level. A supply chain system may seek lower inventory while merchandising prioritizes product availability. An AI system operating across both areas needs an explicit business framework for balancing those goals.

Measurement is equally important. Faster decisions do not establish that the operating model is improving. Retailers need to track business outcomes such as availability, waste, gross margin, promotion effectiveness, inventory productivity and labor efficiency. They should also monitor exceptions, overrides and failed automated actions to understand where AI remains unreliable.

The executive priority is therefore broader than selecting an agentic AI platform. Retailers need connected data, clear decision rights, redesigned workflows and leadership alignment. Those capabilities determine whether agentic AI remains a collection of isolated experiments or becomes part of normal retail operations.

Thin margins and uncertain ROI are slowing grocery AI adoption

A 1.7% average profit margin changes how grocers can invest in AI. A technology project does not need to fail completely to destroy value. Cost overruns, weak adoption or small operational gains can be enough to make the investment unattractive. This explains why many food retailers are approaching AI more cautiously than technology vendors or suppliers.

The adoption data shows a clear gap. A 2025 survey from FMI, The Food Industry Association found that 47% of grocers reported using AI in their operations. Among suppliers, the figure was 93%. Scale also matters. Among large grocers, 77% reported using AI, while 74% said they use generative AI.

Steve Markenson, Vice President of Research and Insights at FMI, connects this caution directly to industry economics. “Food retailers operate on a narrow profit margin, currently an average of 1.7%, which means their investments need to be strategic, and they are not often fast followers,” he said.

Bobby Gibbs, Principal at management consulting firm Oliver Wyman, identifies ROI uncertainty as another barrier. He says many grocers are reluctant even to establish partnerships with AI technology companies because management cannot yet determine whether the financial return will justify the investment. This is a rational constraint in a business where capital competes with store upgrades, logistics, labor, e-commerce and other operational priorities.

The core problem is measurement. AI vendors can demonstrate improvements in model accuracy or task completion, but those gains do not automatically become operating profit. A better demand forecast only has financial value if it reduces waste, improves availability, lowers inventory or produces another measurable result. The same rule applies to pricing, promotions and labor applications.

Executives should therefore require business cases built around operating economics rather than AI performance alone. Each deployment needs a baseline, an expected improvement, implementation costs and a defined measurement period. Total cost should include integration, data preparation, employee training, vendor fees, computing usage and ongoing oversight. A low initial software price does not establish a low total cost of ownership.

Small deployments can help resolve uncertainty. A retailer can test an AI system in selected stores, categories or workflows and compare the results with existing processes. This creates evidence before the company commits to a broader rollout. The test should measure financial outcomes and operational performance, not only whether employees use the system.

Executives should not interpret caution as a reason to postpone AI indefinitely. Waiting also has a cost. Competitors that improve ordering, pricing or labor productivity can compound those gains over time. The stronger approach is controlled investment: prioritize use cases with measurable economics, establish clear thresholds for expansion and stop initiatives that fail to produce repeatable value.

Workforce concerns and rising usage costs can limit the scale of AI deployments

AI creates two costs that executives cannot treat as secondary issues: workforce disruption and computing expense. Both become more important as retailers move from small AI pilots to systems used continuously across large employee populations and business processes.

Employee concern is already high. A joint survey by Express Employment Professionals and Harris Poll found that 90% of job seekers had growing concerns about AI in the workplace. For retailers, this matters because operational AI directly touches activities performed by planners, category managers, store employees, warehouse workers and corporate teams.

Management needs to be specific about how those jobs will change. Some systems will automate tasks. Others will provide recommendations, accelerate analysis or reduce administrative work. Employees need to understand what decisions remain theirs, which tasks will be automated and which new skills will be required. Without that clarity, adoption can suffer even when the underlying technology performs well.

Training is therefore an operating requirement, not simply an employee-relations measure. A retailer that buys AI but leaves workers uncertain about when to trust, question or override its output introduces new execution risk. Managers also need processes for reporting incorrect recommendations and improving the system over time.

Cost creates a separate scaling problem. Advanced AI services can charge according to consumption, including the amount of computing and model processing used. Agentic systems may generate additional expense because they can perform multiple model calls and actions to complete a task. As usage expands across thousands of employees or frequent operational decisions, costs can grow quickly.

Walmart provides a significant example. Bloomberg reported in June that the retailer capped the amount of agentic AI employees could use after usage costs surged. The Economist separately warned that AI’s substantial processing requirements could cause costs to “spiral out of control.” These reports highlight a problem that can be obscured during limited pilots: unit economics can change materially at enterprise scale.

C-suite teams should therefore evaluate AI on cost per useful business outcome. The relevant question is not how much an individual AI interaction costs. Management needs to know whether the complete workflow saves more money or creates more value than it consumes. An agent that costs more to operate but eliminates substantial manual work can still have strong economics. A frequently used assistant that generates little measurable productivity improvement may not.

Cost controls should be designed before broad deployment. Companies can establish usage limits, assign different models to tasks based on complexity and monitor consumption by function. High-cost models should be used where their additional capability creates measurable value. Routine work may not require the most computationally expensive option available.

The executive issue is ultimately one of sustainable deployment. AI must work economically when thousands of employees and millions of decisions are involved, not just during a controlled demonstration. At the same time, workers need clear roles, training and authority as automation expands. Retailers that manage both constraints early will have more room to scale successful AI systems without creating uncontrolled costs or avoidable organizational resistance.

Leadership readiness and organizational culture will determine whether AI produces business value

AI adoption is becoming easier. Organizational change is not. Retailers can buy AI software, use cloud services and work with specialist vendors without building large internal engineering teams. The harder problem is changing how people make decisions, share data and measure results.

Capri Brixey, Partner at consultancy The Partnering Group, makes this distinction clear. “Finding the resources to be able to implement AI is not difficult. It’s about making your organization ready to take advantage of it, and being able to effectively measure the outcome,” she said. She also notes that companies achieving the strongest results invest as much in enablement and process redesign as they do in the technology itself.

For executives, organizational readiness starts with decision rights. AI can produce an answer quickly, but value is lost if employees do not know whether they can act on it. Management must define which decisions AI can make, which employees can approve its recommendations, when human review is mandatory and how exceptions are handled. Without those rules, faster analysis can simply create another approval burden.

Workflows also need redesign. Adding AI to an inefficient process does not necessarily remove the inefficiency. If a forecasting system improves demand predictions but ordering still depends on slow manual processes, much of the potential benefit remains unrealized. Leaders need to evaluate the full workflow around each AI application and remove steps that no longer serve a clear purpose.

Bobby Gibbs, Principal at management consulting firm Oliver Wyman, highlights the role of culture. He cites a CEO of a five-store retailer who requires employees to use Microsoft Copilot in their daily work. Gibbs distinguishes this type of behavioral change from advanced AI capabilities that demand major capital investment. As he explains, “There’s a level of capability that does require major capital investment, but there’s also some cultural change that doesn’t require that kind of investment.”

That example does not mean mandatory AI use is appropriate for every task. Usage itself is not a meaningful success metric. Leaders should define where AI is expected to improve productivity or decision quality, teach employees how to use it safely and measure whether those gains occur. Otherwise, adoption targets can reward activity without establishing business value.

Executive sponsorship is especially important because operational AI crosses established functions. Gary Hawkins, Founder and CEO of the Center for Advancing Retail & Technology, argues that AI transformation cannot remain an isolated technology initiative. “This has got to be driven by the CEO. It’s got to be driven by the boards. It’s got to be driven by the owners,” he said.

This does not require every CEO to become an AI specialist. It requires senior leadership to set priorities, resolve conflicts between functions and make managers accountable for outcomes. Dollar General and Ahold Delhaize have recently appointed executives to AI-specific leadership positions. Such roles can coordinate execution, but responsibility for transformation still belongs with the broader leadership team.

Measurement is the final requirement. AI programs need operating metrics tied to the original business problem: waste, product availability, gross margin, forecast accuracy, labor productivity, inventory performance or decision cycle time. Executives should distinguish adoption metrics from outcome metrics. More employees using AI does not prove that the company has become more productive.

The management priority is clear. Technology access is no longer the main constraint for many retailers. The harder work is redesigning processes, assigning decision rights, developing employee capability and measuring economic results. Companies that solve those problems will extract more value from the same AI technology available to their competitors.

The gap between AI adopters and laggards could become a material competitive threat

The competitive impact of AI will depend on accumulated operating improvements, not the number of AI products a retailer deploys. Better forecasting can reduce waste. Better ordering can improve availability. More effective pricing can protect margin. Faster analysis can increase managerial capacity. When these improvements operate across multiple functions, the difference in performance can become significant.

Gary Hawkins, Founder and CEO of the Center for Advancing Retail & Technology, expects the pace of change to accelerate sharply. “I think we’re going to see more change in the next 24 to 36 months than we’ve seen in the last three or four decades,” he said. Hawkins believes the performance gap between companies that adopt advanced AI and those that do not could become existential.

Executives should treat the specific 24-to-36-month timeframe accordingly. The strategic argument behind it is more important: AI can increase the speed and scale of operational decisions, so differences in organizational capability may become more consequential as the technology improves.

The effect will not necessarily divide the market between large and small retailers. Tom Furphy, CEO and Managing Director of venture capital firm Consumer Equity Partners and a former Amazon and Wegmans executive, argues that operational AI could make independent grocers and small chains more nimble. Smaller companies can access commercial AI systems without recreating the technology infrastructure of Walmart or Amazon.

Capri Brixey, Partner at The Partnering Group, also frames organizational readiness as more important than simply having investment capital. “Finding the resources to be able to implement AI is not difficult. It’s about making your organization ready to take advantage of it, and being able to effectively measure the outcome,” she said.

This creates a different competitive test for C-suite teams. The question is not whether the company has the largest AI budget. It is whether it can identify valuable use cases, provide reliable data, change workflows and scale applications that produce measurable returns. A smaller retailer that executes those steps well can benefit from AI. A larger company can still waste capital by deploying disconnected tools without clear objectives.

Timing matters, but indiscriminate investment is not the answer. Grocery retailers operate on narrow margins. Executives therefore need to balance the cost of moving too slowly against the cost of scaling unproven technology. Both can damage competitiveness.

For executives, the practical response is neither passive observation nor an uncontrolled AI spending program. Build the capabilities that allow successful applications to scale: reliable data, measurable objectives, clear decision rights and workflows designed for AI-assisted execution. The advantage will come from repeatedly converting AI capability into better operating results.

The bottom line

The strongest AI opportunity in grocery is operational. The economics make that clear. In a business with an average profit margin of 1.7%, reducing waste, improving inventory accuracy and making better pricing or ordering decisions can matter more than adding another customer-facing AI feature.

The harder problem is not buying AI. It is preparing the business to use it. Fragmented data, slow decision processes and unclear ownership will limit even capable systems. Agentic AI raises the stakes because software will increasingly move from recommending actions to executing them within defined limits.

Executives should therefore fund outcomes rather than AI projects. Start with a costly, measurable problem. Establish the baseline. Fix the data required to make the decision. Define human oversight and decision rights. Then scale only when the economics hold up under real operating conditions.

This also means AI cannot remain an IT initiative. CEOs and boards need to decide where automation creates strategic value, how much autonomy systems should receive and what organizational changes are required. Technology teams can implement the tools. Leadership must redesign the operating model around them.

Grocers do not need the technology budgets of Walmart or Amazon to compete. They do need discipline. The advantage will go to retailers that can repeatedly turn available AI capabilities into lower costs, faster decisions and better execution.