Retail is facing a growing AI deployment bottleneck. While organizations are experimenting with artificial intelligence across departments, relatively few have successfully moved beyond isolated pilots and deployed AI consistently across the enterprise.
Carnegie Mellon University’s Software Engineering Institute reports that only 8% of organizations have achieved enterprise-scale AI deployment maturity. A U.S. Census Bureau working paper similarly found that among firms already using AI, 57% deploy it in three or fewer of the 15 business functions tracked by its survey. These findings suggest that AI adoption is often still narrow in scope rather than deeply embedded across enterprise operations.
The challenge is particularly significant in retail, where adoption remains comparatively low and operating environments are highly fragmented. The U.S. Census Bureau’s Business Trends and Outlook Survey found that roughly 14% of retail-trade businesses reported using AI in May 2026, below the 19.8% average across all businesses.
At the same time, leading retailers are moving away from standalone AI experiments and toward integrated systems that can support decisions across inventory management, fulfillment, customer engagement, and operational planning. For organizations operating across hundreds or thousands of locations, multiple business functions, and complex technology environments, the central challenge is increasingly less about building AI models and more about deploying them consistently to deliver measurable business value.
In a series of discussions on the AI in Business Podcast, Emerj spoke with Larissa Schneider, Co-Founder and COO at Unframe, and Chris Slovak, Global Field CTO at Unframe, about the operational barriers preventing enterprise AI from reaching production and the strategies organizations can use to overcome them.
Their insights point to four connected principles for scaling AI in modern retail: establishing a unified AI program, building modular AI components, enabling AI systems to reason across distributed enterprise systems, and creating reusable enterprise context that makes every subsequent deployment faster.
Creating a Unified AI Program for Enterprise-Wide Alignment
One of the biggest challenges facing enterprise AI adoption is that experimentation often happens faster than leadership can track it.
Employees across departments may independently create automations, agents, proofs of concept, and low-code applications to eliminate repetitive work. These initiatives can be valuable individually, but without a central strategy, organizations can end up with disconnected collections of AI projects that are hard to evaluate, govern, integrate, or scale.
Larissa Schneider describes this as a landscape of isolated AI “islands.” Employees are encouraged to innovate, but leadership eventually faces a fundamental question: how do all of these experiments fit together?
She explains:
“You end up with these little islands of projects — trials and POCs running across every part of the organization, which naturally happens. And it’s great; we love people who are eager to try new technology and innovate their work. But how does all of that fit together? That’s the question we’re asking. Because to have reliable, governable, secure AI that truly moves the needle for the business, everything has to work in tandem — everything has to work together.”
The issue is therefore not that experimentation is inherently bad. In fact, experimentation is essential for discovering valuable use cases. The problem occurs when successful experiments have no clear path into a broader enterprise architecture.
A scalable AI strategy requires a central program that provides common standards for architecture, governance, ownership, data access, and performance measurement. Rather than allowing every department to build its own AI environment, organizations can create a shared foundation into which individual use cases can plug.
This gives executives visibility into what AI systems are being built, which enterprise systems they access, what information moves through them, who owns each workflow, and how the quality of their outputs is measured.
That visibility becomes especially important as AI systems become more autonomous. Organizations need to understand not only which models are being used but also what actions those models can take and how their outputs affect business processes.
For Schneider, governance is therefore an operational capability rather than simply a collection of policy documents. Every AI initiative should align with the same enterprise architecture and standards so that successful projects can move from experimentation into production without requiring a complete redesign.
Quality measurement is equally important. Employees are more likely to adopt AI when its outputs are consistent and reliable, while executives need objective benchmarks to determine whether a system is delivering meaningful business value.
A unified AI program creates this common foundation. Instead of scattered pilots competing for resources and operating independently, every initiative becomes part of a larger system that can be evaluated, governed, and expanded.
Building Modular AI Components for Faster Deployment
A second barrier to enterprise AI adoption is the tendency to treat every use case as a custom engineering project.
Retailers may have dozens or hundreds of potential AI applications, from inventory planning and demand forecasting to customer service and supply-chain optimization. If each initiative requires teams to rebuild integrations, interfaces, reasoning systems, and governance mechanisms from the ground up, deployment quickly becomes slow and expensive.
Schneider argues that many apparently different AI applications actually rely on remarkably similar underlying components.
She explains:
“What we noticed very quickly is that AI use cases can come from all different parts of an organization, from all different industries and verticals and teams, but the underlying components that you need to put these AI use cases into practice are actually very similar.”
Those components can include user interfaces, dashboards, enterprise connectors, data-lake integrations, reasoning systems, and AI auditability mechanisms. The business problems may look completely different, but the technical foundations often overlap.
For example, an inventory-planning application in retail and an automated claims-processing system in insurance may have entirely different business objectives. Yet both may require connections to enterprise data, workflow orchestration, reasoning capabilities, user interfaces, and mechanisms for monitoring AI behavior.
This creates an opportunity to approach enterprise AI more like assembling reusable building blocks than constructing every application from scratch.
Schneider compares the process to working with a collection of Lego pieces. Once an organization has developed a large library of reusable components, teams can select the pieces required for a particular workflow and configure them for the business problem at hand.
The result is a fundamental change in how AI projects are developed. Instead of repeatedly building the same infrastructure, teams can focus their engineering effort on the unique aspects of each workflow.
Enterprise connectors are a good example. If an organization has already built a reliable connection to SAP or Salesforce, that integration should not need to be recreated for every department that wants to deploy an AI workflow. It becomes shared infrastructure that new applications can inherit.
The same principle applies to user interfaces, reasoning modules, audit systems, authentication, and other common capabilities.
This modular architecture can significantly shorten deployment timelines because the organization is not starting from zero with every new initiative. Each project benefits from the technical and operational lessons of previous implementations.
The unique business logic of the workflow becomes the primary area requiring customization, while the underlying infrastructure remains reusable.
Schneider’s broader observation is that enterprise AI deployment is often less about writing large amounts of new code and more about clearly defining the workflow, required outputs, user experience, and expected business value.
Once those requirements are established, modular components can make the implementation process much faster.
Enabling Distributed Reasoning Across Retail Systems
Even with a unified AI strategy and reusable technical components, retailers face another major challenge: their operational data is rarely located in one place.
Enterprise systems such as ERP, CRM, point-of-sale platforms, warehouse-management systems, and other operational applications each contain different pieces of the business picture.
An order’s status may exist in one system, inventory information in another, customer context in a CRM, and supply information somewhere else. Humans routinely combine these pieces of information when making operational decisions, but conventional AI architectures often assume that data needs to be centralized before intelligent applications can be built.
Chris Slovak argues that this assumption can become one of the biggest obstacles to AI deployment.
Retailers may spend years consolidating data, cleaning information, developing unified schemas, and migrating systems before allowing AI applications to interact with the resulting environment. By the time the centralized architecture is ready, business requirements and underlying systems may have already changed.
Slovak argues that agentic systems can take a different approach by reasoning across distributed sources directly.
He explains:
“Agents and agentic systems in particular don’t necessarily need one source of aggregated data truth. They can reason like a human can across multiple systems, so long as there’s context and semantic linking. You have ERPs and CRM, and then you have your point of sale systems, and they probably all semi-talk, but the truth is the state of order, supply, inventory probably to some extent live a little bit in each.”
This approach reflects how people already work. A retail employee trying to understand why an order has not been fulfilled may look at the customer’s information in the CRM, examine the order in the ERP, check inventory availability, and review warehouse information.
The employee does not necessarily need all of that information consolidated into a single database. They need the ability to access the right information from each system and understand how the pieces relate to one another.
AI agents can potentially operate similarly.
Instead of requiring organizations to complete massive data-centralization projects before deploying AI, agents can retrieve the specific context they need from existing enterprise systems. This allows retailers to work with their current technology environment while gradually improving their data architecture over time.
The benefit is not merely technical. Avoiding a requirement for complete centralization can reduce disruption to existing workflows and allow organizations to begin generating value much earlier.
For retailers, this can make it possible to build AI applications around immediate operational needs rather than waiting for a multi-year transformation program to reach completion.
The architecture therefore becomes more flexible. AI systems can draw information from the systems where operational truth already exists, provided they have appropriate access, context, and semantic relationships between those sources.
Creating Reusable Enterprise Context
The final piece of the strategy is the creation of reusable enterprise context.
Slovak describes a recurring problem in enterprise AI: organizations spend enormous amounts of time building foundational infrastructure before solving a single business problem. They may focus on data migrations, centralized warehouses, new models, or enterprise-wide transformation programs while delaying actual deployment.
A more effective approach is to begin with a specific business problem and build outward.
When an organization successfully deploys its first AI use case, it does more than solve one problem. The project also creates valuable knowledge about business entities, system relationships, data sources, workflows, permissions, and operational rules.
That knowledge can then become part of the foundation for the next AI application.
Slovak describes this compounding effect:
“The concept that the core context is going to evolve, that has to be core to your design decisions, if my first use case gives me exposure to 60% of the major business entities and tools that I use today, use case number two is already 60% of the way there.”
This means the value of enterprise AI infrastructure can increase with every successful deployment.
The first use case may require significant work to understand a company’s systems and business processes. The second use case can reuse much of that knowledge. The third can build on both. Over time, implementation becomes faster because the organization has accumulated a growing layer of reusable enterprise context.
This creates a compounding advantage.
Instead of treating every AI deployment as an independent project, organizations can design each implementation to contribute something to the next one. New system integrations, business rules, data relationships, workflow definitions, and governance mechanisms become reusable assets.
The result is a deployment model in which AI adoption can accelerate over time rather than repeatedly returning to the starting line.
This also changes how organizations should think about transformation. The objective is not necessarily to design the perfect enterprise AI architecture before deploying anything. It is to create an architecture capable of learning and evolving as real use cases are implemented.
From AI Pilots to an Enterprise Deployment Engine
The insights from Schneider and Slovak point toward a common conclusion: enterprise AI deployment is fundamentally an operating-model challenge.
Retailers do not necessarily lack AI models or potential use cases. They often lack the architecture, governance, reusable infrastructure, and enterprise context required to turn individual experiments into scalable production systems.
A unified AI program gives organizations a common structure for managing their initiatives. Modular components allow teams to assemble new applications without rebuilding foundational infrastructure. Distributed reasoning allows agents to work with information across existing enterprise systems instead of waiting for every piece of data to be centralized. Reusable context ensures that every successful deployment makes the next one faster.
These principles reinforce one another.
A centralized AI program creates the standards that allow components to be reused. Modular components make it easier to connect agents to distributed systems. Those deployments create additional enterprise context. That context then reduces the effort required to build future applications.
The result is a potential shift from isolated AI experimentation to a repeatable enterprise deployment engine.
For retailers, this distinction could become increasingly important. Competitive advantage will not necessarily come from having the largest number of AI pilots. It will come from the ability to identify valuable use cases, deploy them quickly, measure their results, and use what is learned to accelerate the next wave of applications.
Conclusion
AI adoption in retail is entering a new phase. The question is no longer simply whether organizations can experiment with artificial intelligence. The more difficult question is whether they can turn those experiments into reliable, governed, production-grade systems that operate across the enterprise.
The experience shared by Larissa Schneider and Chris Slovak suggests that scaling requires a different mindset.
Organizations need one coherent AI program rather than disconnected projects. They need reusable components rather than bespoke implementations. They need AI systems capable of reasoning across the distributed technology environments that already run their businesses rather than waiting for perfect data centralization. Most importantly, they need to treat enterprise context as an asset that compounds with every deployment.
The companies that solve these challenges can move beyond the pilot stage and build an increasingly efficient system for deploying AI at scale.
For modern retailers, that ability may become one of the most important competitive advantages of the AI era: not simply having artificial intelligence, but having the organizational and technical infrastructure to put it into production repeatedly, reliably, and at enterprise speed.