Retail AI: From Agent Sprawl to a Coordinated Operating Model

Why Retail Has Too Much AI, Not Too Little

Most large retailers do not have too little AI. They have too much of it, disconnected. A demand-forecasting model, a personalization engine, a fraud-detection system, and an inventory optimizer can each perform well in isolation while producing no combined intelligence about the retail operation as a whole. Gartner projects that the average large enterprise will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025, and warns that without governance this growth produces agent sprawl rather than coordinated capability (Gartner, 2026). Retail, an early and aggressive adopter of point AI tools, is arriving at this problem first, and how it responds will set the pattern other sectors follow.

Executive Summary

Retail entered its first AI wave between 2022 and 2024, deploying point solutions rapidly: demand forecasting, personalization, fraud detection, and inventory optimization each became a separate, often separately owned, AI initiative. By 2025, adoption looked strong on paper. McKinsey found 88% of organizations use AI in at least one business function and 72% report using generative AI, up from 33% in 2024 (McKinsey & Company, 2025). But value has not compounded at the same rate: 71% of merchants surveyed by McKinsey say AI merchandising tools have had limited to no measurable effect on their business so far, and 61% describe their organization as not at all, or only slightly, prepared to scale AI across merchandising (McKinsey & Company, 2026).

The gap is coordination, not capability. Gartner projects enterprise AI agent counts will grow from fewer than 15 per company in 2025 to more than 150,000 by 2028, and expects 40% of enterprises to scale back autonomous agent deployments by 2027 as governance problems surface only after agents reach production (Gartner, 2026). Walmart's response, previewed at its 2025 Converge technology event, is instructive: rather than deploying more agents, it re-architected its Element machine-learning platform around standardized coordination protocols, including Agent-to-Agent and Model Context Protocol, so existing agents can discover, call, and hand off work to one another (Walmart Global Tech, 2025).

Because the disruption reshaping retail AI is a coherence problem rather than a capability gap, this brief argues that retailers must move from independently deployed point tools to a coordinated cognitive operating model, built on shared data contracts and a governed coordination layer, rather than by continuing to add point solutions.

Sector Context

Retail's operating model has historically been organized around discrete functions: merchandising, supply chain and inventory, marketing and personalization, loss prevention, and store operations, each with its own systems, data, and increasingly its own AI tools. This structure made sense when analytics were bespoke projects; each function built or bought what it needed, and coordination between functions happened through people, meetings, and shared spreadsheets rather than shared systems.

That structure is now the primary constraint on AI value. The National Retail Federation's Center for Digital Risk and Innovation surveyed 56 AI leaders at US retailers in summer 2025 and found that most retailers already treat AI governance seriously: 86% have governance policies in place, and 68% report CEO involvement in oversight. Yet retailers still allocate AI conservatively, with 77% spending 5% or less of their technology budget on AI, even as 39% expect AI to exceed 10% of tech spend within three years (National Retail Federation, 2025). Governance maturity is running ahead of coordination maturity: retailers know AI needs oversight, but most have not yet built the technical layer that lets AI tools work together across functional boundaries.

Consumer-facing AI use has, meanwhile, moved faster than the retail operating model has adapted to it. A global study from the IBM Institute for Business Value, conducted with the National Retail Federation, found that 45% of consumers now turn to AI for help during their buying journey, researching products, interpreting reviews, and hunting for deals, often before a retailer's own systems are involved at all (IBM Institute for Business Value & National Retail Federation, 2026). Retailers are being asked to coordinate internally at the same moment external AI agents are starting to intermediate the shopping journey itself.

Four Forces Driving Retail Toward Coordinated AI

Point-tool proliferation has outpaced the ability of retail organizations to connect what those tools learn. Eighty-eight percent of organizations use AI in at least one function, and 72% use generative AI, up from 33% in 2024, yet 71% of merchants report limited to no measurable business effect from AI merchandising tools specifically (McKinsey & Company, 2025, 2026). Adoption breadth is not converting into performance, because each tool optimizes its own function without sharing what it learns. The next investment dollar in retail AI creates more value connecting existing tools than adding new ones.

Agent sprawl is becoming a governance problem before it becomes a productivity one. Gartner projects the average large enterprise will run over 150,000 AI agents by 2028, up from fewer than 15 in 2025, and forecasts that 40% of enterprises will scale back autonomous agent deployments by 2027 once governance issues surface post-production (Gartner, 2026). Retailers that deploy agents function by function, without a shared coordination and governance layer, are building toward the exact failure pattern Gartner describes. Coordination architecture needs to be designed before agent count grows further, not retrofitted after governance problems appear.

Platform leaders are converging on standardized agent-to-agent coordination protocols rather than proprietary integrations. Walmart's Element platform was rebuilt around standardized communication protocols, including Agent-to-Agent and Model Context Protocol, plus a super agent, Wibey, that makes domain-owned agents discoverable and interoperable rather than replacing them (Walmart Global Tech, 2025). The technical means to coordinate agents across functional silos now exists as an adoptable pattern, not a custom-build problem each retailer must solve alone. Retailers evaluating their AI architecture should assess whether new tools can plug into a common protocol layer, not only whether each tool performs well individually.

Agentic commerce is starting to route around the retailer's own systems entirely. McKinsey forecasts that AI shopping agents transacting on a consumer's behalf could generate three to five trillion dollars in global retail revenue by 2030, and reports that 43% of consumers already say they trust agents for simple purchases, as of an April 2026 survey (McKinsey & Company, 2025, 2026). Retailers face a second coordination problem beyond their own internal tools: coordinating with external AI agents that now influence or execute purchases. The internal coordination layer retailers build for their own AI tools needs to be designed to also expose clean, governed interfaces to external shopping agents, not treated as a purely internal architecture decision.

Albertsons illustrates a second, complementary path to coordination: rather than building a protocol layer to connect back-end systems, it built a single front-end multi-agent assistant that already spans meal planning, ingredient deduplication, and cart building. Launched in December 2025 on a multi-agent architecture designed for future interoperability with third-party agentic systems, the assistant cut the typical 46-minute grocery trip to about 4 minutes (Albertsons Companies, 2025). Where Walmart coordinates domain-owned agents behind the scenes, Albertsons coordinates them behind one customer-facing interface. Retailers do not need to choose one pattern over the other, but they do need to choose deliberately, rather than let coordination architecture emerge as an accident of whichever vendor sold them their most recent tool.

From Tool-Level Intelligence to System-Level Coherence

The underlying logic of retail competitive advantage is shifting from tool-level intelligence to system-level coherence. In the point-tool era, advantage came from having a better forecasting model, a sharper personalization engine, or a faster fraud-detection system than a competitor. Each tool competed on its own merits, and procurement decisions were made function by function.

That logic is breaking down for two connected reasons. First, individual AI tools are approaching a performance ceiling when they operate in isolation. McKinsey's merchant data shows 71% report limited effect from AI merchandising tools, not because the models are weak, but because a merchandising decision made without visibility into live inventory, fulfillment, and demand signals from other systems is a locally optimal decision inside a system that is not locally optimized (McKinsey & Company, 2026). Second, agent sprawl means the number of AI systems inside a typical retailer is growing far faster than any team's ability to manually coordinate them; Gartner's projection of over 150,000 agents per large enterprise by 2028 makes manual coordination structurally impossible past a certain scale (Gartner, 2026).

The structural shift, then, is from AI as a portfolio of independently procured capabilities to AI as a coordinated operating system for the retail enterprise, where individual tools remain domain-owned but operate through shared data contracts and a common coordination layer, the model Walmart's Element and Wibey architecture makes concrete (Walmart Global Tech, 2025). The unit of competitive advantage moves from which tool is smartest to how well the full system of tools works together.

Retail's Coordination Problem Through the 6xD Framework

Two lenses carry this brief's argument: D2 (Digital Cognitive Organization) and D3 (Digital Business Platform). D6 provides meaningful supporting context; D4 and D5 are treated briefly, as consequences rather than independent drivers.

D2, Digital Cognitive Organization, is the primary lens because retail's problem is defined precisely by D2's core claim: competitive advantage comes not from what any single AI tool knows, but from what the operating model as a whole learns from the interaction of its tools. A demand forecast that never reaches the inventory-optimization system is not organizational cognition; it is a smarter spreadsheet. McKinsey's 71% figure on limited merchandising impact is a direct measurement of this gap: tools that do not feed each other's inputs cannot produce system-level learning, no matter how individually capable they are (McKinsey & Company, 2026).

D3, Digital Business Platform, is the second primary lens because D2's closed loop is not achievable without a shared technical substrate connecting previously independent AI tools. Walmart's rebuild of Element around standardized Agent-to-Agent and Model Context Protocol coordination, plus the Wibey super agent that makes domain-owned agents discoverable rather than centralizing them, is a concrete instance of building that substrate (Walmart Global Tech, 2025). The platform lens clarifies that the fix for agent sprawl is not fewer agents; it is a governed platform layer through which existing, domain-owned agents can discover and coordinate with each other.

D6, Digital Accelerators, is a meaningful supporting lens: standardized protocols like Model Context Protocol and Agent-to-Agent function as accelerators precisely because they let retailers adopt a coordination pattern already proven at scale, rather than building bespoke integration for every pair of AI tools. This directly lowers the cost of the D3 platform investment retailers need to make, though it is not the central argument.

D4, Digital Transformation 2.0, is supporting: once a coordination layer exists, further AI investment can be evaluated by whether it plugs into that layer rather than whether it is impressive in isolation, turning transformation from a sequence of point deployments into a cumulative, architecture-governed program.

D5, Digital Worker and Workspace, is also supporting: as agents coordinate more of the routine decision flow between forecasting, inventory, and fulfillment, retail associates' and merchandisers' roles shift toward exception management, agent oversight, and judgment calls that cross the domains individual agents still cannot bridge, a redesign question retailers need to address deliberately as coordination increases.

The Tradeoffs of Coordinating Retail's AI Agents

The opportunity is significant: retailers with a working coordination layer can convert the 88% adoption and 72% generative-AI-usage baseline that already exists into measurable performance, closing the gap behind McKinsey's 71% limited-impact figure, and can additionally position themselves to serve, rather than be bypassed by, the external shopping agents McKinsey expects to route three to five trillion dollars in transactions by 2030 (McKinsey & Company, 2025, 2026). Each opportunity carries a matched risk.

Building a coordination layer that connects previously siloed AI tools also expands the blast radius of a single point of failure: an error propagating through a shared data contract can now affect forecasting, inventory, and fulfillment simultaneously, rather than being contained to one tool. Coordination architecture needs matched incident-response and rollback design, not only integration engineering.

Exposing internal coordination interfaces to external shopping agents creates new commercial opportunity but also cedes some control over the customer relationship and pricing dynamics to third-party agent platforms the retailer does not control, a tradeoff that needs an explicit governance decision rather than a default acceptance driven by competitive pressure.

Standardized protocols like Model Context Protocol and Agent-to-Agent accelerate coordination-layer build-out, but adopting external protocols also means inheriting their security and versioning dependencies. Retailers should treat protocol governance as a supply-chain risk category, similar to how they already treat vendor software.

Gartner's own forecast that 40% of enterprises will scale back autonomous agents by 2027 due to governance problems is itself a warning that the sector may be over-rotating toward agent deployment before coordination governance catches up (Gartner, 2026). Retailers should read that projection as a reason to build governance now, not as evidence to slow adoption altogether.

Six Moves to Build a Coordinated AI Operating Model

Inventory existing AI tools and their data dependencies before commissioning new ones. Most retailers cannot yet answer, tool by tool, which systems each AI deployment reads from and writes to; that inventory is the prerequisite for any coordination-layer design.

Establish a shared data-contract standard across forecasting, inventory, personalization, and fraud systems. Define common entity models for customer, SKU, and location that every AI tool must read and write against, rather than allowing each tool to maintain its own.

Adopt or pilot a standardized agent-coordination protocol, such as Model Context Protocol or Agent-to-Agent, rather than building bespoke point-to-point integrations. Treat this as an architecture decision owned by IT leadership, not a byproduct of any single AI vendor selection.

Name a single accountable owner for the coordination layer, distinct from the owners of individual AI tools, with the authority to require new AI deployments to integrate with shared data contracts before going into production.

Set a governance gate on new agent deployment tied to Gartner's sprawl warning: no new autonomous agent moves to production without a defined data contract, a named exception-handling owner, and a rollback plan, closing the exact gap Gartner attributes to the coming 2027 scale-back.

Decide deliberately how much of the coordination layer to expose to external shopping agents, rather than defaulting to either full exposure or full lockout, given the scale of agentic commerce McKinsey projects by 2030.

Closing Perspective

Retail does not need a new generation of smarter individual AI tools. It has enough of them already, and Gartner's trajectory toward more than 150,000 agents per large enterprise by 2028 suggests the sector will soon have far more than enough (Gartner, 2026). What it needs is the operating architecture that lets the tools it already has learn from each other.

The leadership divide in retail AI is forming between organizations that treat each new AI capability as a standalone procurement decision and those that treat every new capability as a candidate for a shared, governed coordination layer. The first group will keep adding tools and keep reporting limited impact. The second group is building the system-level intelligence that actually compounds.