Why are AI agents stuck at 20% adoption in retail?

The blocker isn't model quality or budget. Only 20% of retailers have agentic AI running because most retail estates can't guarantee the store-level operational data feeding those agents is accurate, current and trustworthy. Disconnected systems, unreliable device data and fragmented visibility mean agents can't reason over data they can't trust.

What is the operational truth layer in retail AI?

The operational truth layer is the real-time, connected operational data that AI agents need to function — stock levels, device health, energy consumption, network performance and payment system status across every site. Without it, AI agents make decisions on incomplete or untrustworthy data.

What is blocking agentic AI in retail?

Agentic AI in retail is blocked by poor operational data quality at store level. Most retail estates run disconnected monitoring systems that don't share data in real time, making it impossible for AI agents to act on accurate, current information across every site.

Why AI Agents Are Stuck at 20% in Retail - And What's Actually Blocking Them

Ross Palmer
Jul 3
6 min read

Nine in ten retailers are raising their AI budgets this year.

Just one in five have agentic AI actually running.

That gap, confirmed in NVIDIA's third annual State of AI in Retail and CPG survey, is the single most important number in retail technology right now. Not because it reveals a failure of ambition.

Every retailer in the survey wants this to work. It reveals something more specific and more fixable: most retail estates cannot yet feed agentic AI retail data it can trust.

This is not a model problem. It is a plumbing problem. And understanding it changes how every retail technology leader should be thinking about AI investment in the second half of 2026.

The platform fight nobody asked for

Two of the biggest retail technology stories this year have been about who controls the agent layer.

Amazon began licensing its AI shopping technology to outside retailers this year, the same engine behind Alexa for Shopping, packaged on AWS, already live with retail partners and expanding. Walmart took the opposite approach, partnering with Google on open commerce-protocol standards designed to let any AI agent transact across any store, rather than building a closed system of its own.

Closed and fast versus open and interoperable. Every other retailer now has to choose a side, or wait to see which model wins.

But underneath that platform fight sits a more important question that almost nobody is asking publicly: who owns the operational data that makes either approach actually work?

An AI agent's intelligence is only as good as the data it's reasoning over. Most retail estates can't yet guarantee that data is accurate, current, or even available.

The real bottleneck: data, not models for agentic AI retail

NVIDIA's survey is the most useful reality check the industry has had this year. The headline numbers tell a story of genuine momentum, 94% of retailers say AI has already cut operating costs, and 87% say it has lifted revenue. Adoption is not the problem.

But dig one layer deeper and the picture changes. Only 47% of retailers are using or actively assessing agentic AI at all. Of those, the leading use cases are internal workflow automation and knowledge retrieval, not the customer-facing AI agents that dominate the headlines. And critically, just 20% have agentic AI actually running in production, with another 21% saying it is coming within the next year.

That leaves more than half of retailers with agentic AI still firmly in the planning or piloting stage - despite budgets rising across the board.

The stated goals retailers want from agentic AI are revealing in themselves:

• Speed and operational efficiency - 57% of respondents

• Better customer experience - 40%

• Faster decisions based on real-time data - 40%

Every one of those goals depends on the same precondition: an AI agent needs current, accurate, connected data about what is actually happening across the estate right now. Stock levels. Device status. Energy consumption. Labour availability. Shelf condition. Payment system health.

Most retail estates cannot guarantee that data exists in a usable form. It sits in disconnected systems, updates on different schedules, and often cannot be trusted to reflect reality at the moment an agent needs to act on it.

An agent cannot reason over data that does not exist, or that it cannot trust. That precondition, not model quality, not budget, not ambition, is what is holding the 20% number down.

The reason isn't model quality. It's that store-level operational data is messy, disconnected and untrustworthy. Agents can't reason over data that doesn't exist or can't be trusted.

Where the early wins are already happening - and why

The areas of retail where agentic AI is gaining real traction this year share one characteristic: the operational data underneath them is unusually clean, structured and real-time already.

Shelf-monitoring computer vision is a clear example. The technology itself is not new, what changed in 2026 is that edge processing made it economically viable at scale. Cameras now process images locally, flagging out-of-stocks, planogram mismatches and misplaced products without streaming video to the cloud.

The data produced is structured, immediate, and directly actionable. 97% of retailers plan to hold or increase AI investment in vision technology this year, because the operational data layer underneath it already works.

Self-checkout is following the same pattern. Modern systems use AI vision to catch mis-scans and flag shrink in real time, turning every transaction into a structured data point about stock, pricing accuracy and loss, rather than a standalone payment event.

The checkout has effectively become a sensor, generating clean operational signal as a byproduct of something retailers were already doing.

AI-driven energy management tells the same story from a different angle. Vendors are now claiming 20 to 40 percent reductions in retail energy costs by dynamically tuning HVAC, lighting and refrigeration systems to real-time conditions.

That only works because HVAC and refrigeration systems generate continuous, structured sensor data, compressor load, temperature variance, fan performance, that is far easier to trust and act on than, say, footfall predictions stitched together from three disconnected systems.

Supply chain is where agentic AI's impact is expected to be most disruptive of all, analysts forecast AI-driven forecasting can cut inventory costs by 20 to 30 percent while improving availability, and reduce perishable waste by up to 30 percent.

But that scale of autonomous rebalancing, dynamic pricing and even vendor negotiation only works if the store-level data feeding those decisions is accurate. Inventory sitting on a shelf only becomes a balance-sheet improvement if the agent making decisions about it can trust what it's being told about that shelf.

Every early agentic AI success this year shares the same precondition: clean, structured, real-time operational data. That's not a coincidence. It's the rule.

The operational truth layer agents run on

As the agent layer itself becomes increasingly commoditised, Amazon licensing its technology, Walmart pursuing open standards, dozens of vendors building agentic tools on top of both, the durable competitive advantage moves one layer down.

It moves to whoever owns clean, real-time signals about what is actually happening across a retail estate. Stock. Energy. Devices. Labour. Shrink. Network health. Payment system performance.

This is the layer that determines whether any agent - whichever ecosystem a retailer eventually chooses - can actually deliver on its promise. Not the smartest model in the world. The most current, trustworthy, connected operational data underneath it.

That is precisely the layer OpSiteAI was built to provide.

OpSiteAI continuously monitors operational signals across every site in a retail or forecourt estate: POS systems, network infrastructure, IoT devices, energy consumption - correlating them into one connected, real-time operational picture. Not a dashboard that shows what happened yesterday. A live truth layer that any system, human or AI, can act on with confidence right now.

Whether a retailer eventually builds on Amazon's agent infrastructure, Walmart's open standards, or develops something proprietary, the question that determines success is the same: can the data feeding that agent be trusted?

The agent war will be won by whoever the retailer picks. The data war underneath it is won by whoever has the operational truth layer ready first.

What this means for retail technology leaders right now

For retail IT and operations leaders evaluating where to invest in the second half of 2026, the NVIDIA data points to a clear sequencing problem. Many organisations are attempting to deploy agentic AI before they have solved the operational visibility problem underneath it, effectively trying to build the second floor before the foundation is poured.

The retailers seeing the fastest agentic AI returns this year, in shelf vision, self-checkout intelligence, energy management, are succeeding specifically in the areas where operational data was already clean and connected. That is not a coincidence. It is the precondition working as expected.

For estates where that operational visibility does not yet exist, where stock data, device health, energy consumption and labour availability sit in disconnected systems updating on different schedules, building the operational truth layer first is not a delay to the AI agenda. It is the AI agenda.

The retailers who solve this now will be ready the moment they choose an agent ecosystem. The retailers who don't will spend 2027 discovering that their AI investment underperformed for reasons that had nothing to do with the AI itself.

If you want to understand what an operational truth layer looks like across your estate, we are ready to show you.

Get in touch at opsiteai.com/contact or call +353(0)818 900 000