71% of retail merchants say AI tools have had limited or no effect so far

Seventy-one percent. That is how many merchants in McKinsey's 2026 retail merchandising research reported that, so far, AI merchandising tools had delivered only a limited effect or none at all. The same research found that 61% said their organizations were only slightly prepared or unprepared to scale AI across merchandising.

Pigment's Q3 CFO Index surveyed 2,000 CFOs and other finance executives in the U.S., UK, France and Germany, and 83% reported that consumption-based AI costs had run higher than they expected. Their confidence in AI's business value still averaged 8.2 out of 10. Finance leaders still believe in the technology. What has surprised them is the bill.

Whether the intelligence turns out to be useful can depend heavily on how well commercial data and workflows are connected. For operators, that moves the main question away from which model to buy and toward something less glamorous: whether the data underneath is joined up well enough for any tool to pay off.

Where the merchant's week goes

When merchants spend about 40% of their time stitching data together and 71% report limited or no effect from AI merchandising tools, the cheapest AI gain may sit in the plumbing between systems they already pay for. McKinsey counts consolidating data, preparing reports and reconciling information across systems among those lower-value activities. That is the work an AI tool would need clean inputs from, and it is still being done by hand.

Part of the difficulty comes from how AI has spread within companies. According to The Hacker News, LayerX Security's State of AI Usage Report 2026 found AI activity is increasingly scattered: personal accounts, browser extensions, embedded copilots, connectors and other tools. The Next Web traces the pattern to pressure for quick results. Departments adopt copilots, chatbots and task-specific automations to meet immediate needs, and the technology environment gets more varied each time.

Each of those tools can solve a local problem and still add one more place where data has to be reconciled. Spending is set to rise as well. The National Retail Federation reports that 39% of surveyed retailers expect AI to represent more than 10% of technology spending within three years, so the cost of getting the architecture wrong would grow along with the budget line.

FastQuery's case for one layer, many roles

FastQuery AI is an AI-native operating platform built around retail operations and insights, led by co-founder and CEO Joseph Bullock and founder and CTO Anton Kulikalov. It begins from a simple observation: retailers already have plenty of technology, including inventory and point-of-sale systems, ERP platforms, supplier networks and analytics.

Kulikalov puts the problem as interpretation, not collection. "The data to make great decisions has always been there," he said, describing the challenge as making sense of it, separating signal from noise and answering one question: "what requires my attention today?"

Joseph Bullock frames the shift as decades spent building systems of record, with the next job being systems of intelligence that put that information where the work happens. Kulikalov describes it as a progression toward an enterprise where employees increasingly work through an intelligent layer. The Next Web, which profiled the company, offers a comparison: search engines reduced the need to visit individual websites for every question, and enterprise AI may similarly cut the number of source applications employees open for routine information and decisions.

The role examples make the idea concrete. A category manager needs assortment, pricing, vendor performance and demand at once. A store associate standing with a customer needs product specs, availability and how-to information. An operations leader needs conditions across locations before setting priorities. Same underlying data, three different questions, and in FastQuery's model the systems of record keep running in the background while employees work through one layer.

The argument isn't unique to store operations. Doba CEO Mandy Ji makes the same case for ecommerce in a sponsored Retail TouchPoints piece: the shift is connecting product research, sourcing, inventory, orders and fulfillment so people spend less time on manual coordination. Doba, a U.S.-focused dropshipping operations platform with access to more than 1 million products, synchronizes inventory and orders across connected stores. Two vendors from different corners of retail landing on one pitch suggests where buyers will hear it next. Both are vendor claims, though, and neither reports measured results.

Appinventiv describes the failure this layer is meant to prevent. A retailer can predict a stockout correctly and still lose the sale if the forecast arrives after purchase orders close, sits outside the replenishment system, or reaches a team without authority to act. Accuracy is the easy part to buy. Getting the answer into the right system before the deadline is the part that pays.

Its pricing example shows the same thing from the merchandising side. A model that recommends smaller discounts on products likely to sell at full price is only worth something if merchandising teams can apply those recommendations within promotion schedules and supplier agreements. For a merchant, that means one question before any model: which decision changes, and who controls it?

What connecting systems actually costs

Speaking at the NACS Show session "Shopping for AI Without Breaking the Bank," as NACS Magazine's Noelle Riddle reported, Rick Rigby, CTO at Atlantis Management, named hidden fees, usage-based pricing, subscription sprawl and integration problems as the usual drains on ROI.

His advice comes in three parts, which he calls the three P's: get data integrity right first, know exactly how you're billed and watch it, and pilot in a real store before trusting the demo. On the first, he ties data integrity to people as much as systems: "Data democratization and building the right teams is the first P." He also sorts AI spending into seven cost buckets:

• Acquire: the subscription or license, plus any modules you need
• Implement: discovery work, configuring the system, integrating it
• Prepare Data: cleanse the data, then map and label it
• Operate: pay for compute and storage, and keep up monitoring
• Adopt: train staff and communicate the change
• Govern: set access controls and run a privacy review
• Exit: export your data and get support through the transition

For operators weighing a connected-systems project, those cost buckets suggest where the money could go. Integration and data preparation could make up much of the work, and ongoing operating costs may be where consumption-based charges land. Exit deserves more attention than it usually gets. A layer that becomes the place employees do their daily work is a layer that's expensive to leave.

Appinventiv puts machine learning implementation budgets in retail at $40,000 to $400,000 or more, depending on scope, data readiness and integrations. Because that range depends on data readiness and integrations, a retailer with clean, mapped inventory and supplier data could land nearer the bottom of that spread, and one still reconciling spreadsheets nearer the top.

Appinventiv's sequencing also matches Rigby's. It advises scaling only after a controlled pilot shows business gains alongside manageable operating costs. That second condition is the one Pigment's respondents appear to have missed, and an in-store pilot is the cheapest place to find out what the Operate bucket really costs.

When the fix becomes another tool

An intelligence layer is still software, and if one department buys it to answer one department's questions, it joins the copilots and extensions LayerX described. The thesis holds only where the layer replaces some of the direct trips into source applications, the way Kulikalov describes it, and doesn't just sit on top of them.

The other boundary comes from governance. The NRF reports that 86% of surveyed retailers have AI governance policies, and Rigby keeps his working rules blunt. If you wouldn't email something outside the company, don't paste it into AI. And no AI output should decide a customer, employee, safety or compliance matter until a human has reviewed it.

Those rules carry more weight once a single layer pulls data from every system into one view, because that view now contains everything. Customers are already wary. SAP's Industry Market Report for retail found 73% of consumers are concerned about how AI uses their personal data and only 24% trust AI-led retail services, so a consolidated layer that touches customer records puts the Govern bucket at the center of the project, not the margins.

Rigby also puts the payoff in human terms. In his view, AI can take mundane tasks off employees' plates and train them for higher-level work, whether that means running a food program or a loyalty program, or engaging with a customer.

The number that would confirm or break this read is McKinsey's 71%. If retailers spending more on connection work see that figure fall while their consumption bills hold steady, the plumbing was the problem. If it doesn't move, the gap sits somewhere other than the data.