Retail is about to spend a fortune solving the wrong problem.
Walk into any transformation office right now and you will find the same shopping list: license an external corpus, stand up a synthetic data pipeline, buy a vertical model. The logic is sound on paper. The use of synthetic AI training datasets is increasing rapidly to supplement or replace real-world machine learning datasets. And the market reflects it. The global AI training dataset market was valued at 3.2 billion dollars in 2025 and is projected to grow to 16.3 billion by 2033.
Here is the contrarian read. The single highest-signal operational dataset in your entire company is not for sale. It is walking the floor right now, and you are throwing it away.
The floor already knows what your model needs
The associate knows the workaround for the register that freezes on split tender. She knows which planogram never survives contact with a real endcap. She knows the exact three sentences that turn a return into an exchange, and the two that lose the customer for good. That is not anecdote. That is judgment under real constraints, captured at the exact moment of decision, tied to a real outcome.
This is precisely the raw material every serious AI practitioner is now telling you that you cannot skip. In 2026 and beyond, the most capable models will still be anchored in human data, because humans are required to define what good looks like, set objectives, establish red lines, and manage trade-offs. The people writing the playbook on synthetic data are blunt about the failure mode. Synthetic data scales human judgement; it does not replace it.
Skip that anchor and you get model collapse. A model trained repeatedly on its own output can drift from reality, rare cases fade first, then the output narrows toward a bland average, and researchers call this model collapse or AI inbreeding. A generic corpus plus a synthetic pipeline gives you a retail agent that sounds fluent and gets the store wrong. Your associates are the correction layer that keeps it honest.
And the associate layer is not niche. Walmart alone employs approximately 2.1 million associates worldwide. Instrumented correctly, that is the largest proprietary reasoning dataset in retail, and no competitor can license it.
The CODN math nobody is running
Cost of Doing Nothing is usually framed as a missed feature or a slow quarter. On training data it is worse, because the asset is perishable and it is actively rotting.
Two forces compound. First, the associate role itself is being redefined in real time. When a Foot Locker associate packs a DoorDash order between helping a customer try on sneakers, something fundamental has shifted. Second, the door never stops spinning. Staff turnover averages 60 percent annually. Every departure that leaves undocumented takes a slice of your best training signal with it.
Meanwhile the upside of getting the frontline right is measurable. McKinsey research shows that top-performing retailers with AI-enabled frontline operations achieve three percentage points higher same-store sales compared with peers. So the CODN is not abstract. It is three points of comp you are leaving on the table while paying vendors for a thinner, generic version of data you already generate for free and then discard.
Three ways to instrument the floor without building surveillance
The reason nobody captures this is fear, and the fear is legitimate. Without intentional design, AI systems risk amplifying bias, misusing sensitive data, or shifting risk onto workers. Instrument the associate the wrong way and you build a productivity panopticon that your best people quit over. So build it the right way.
One. Capture the decision, not the person. Log the escalation, the choice, and the outcome. Do not log keystrokes, idle time, or location. The unit of value is the judgment call, not the human making it. For frontline AI to deliver on its promise, safety, transparency and human oversight must be built in from the start.
Two. Make the loop pay them. When an associate’s correction demonstrably improves the model, that is a contribution, and it should be compensated, credited, and visible. This flips the frame from extraction to authorship. Trust is the currency. It makes more sense to bring people on board who can learn and adapt quickly and arm them with trust than to manage the tech and force employees to conform.
Three. Ship value back to the floor same-week. The fastest way to earn the next correction is to give the associate something the model learned last week that makes their shift easier. The tools already exist to close this loop. Walmart is equipping store associates with a suite of AI tools available through the associate app, designed to eliminate friction, simplify actions, and make work more efficient. Point that same pipe backward and the app becomes both the capture layer and the payoff.
Stop treating the floor as a cost center that consumes your models. It is the mine that produces them. The retailers who win the agentic era will not be the ones who bought the biggest corpus. They will be the ones who figured out that their best training data clocks in every morning, and built the respect to keep it coming.