AI agents might smooth some of retail’s worst data problems

Every vertical has its own unique technical challenges. In retail, for instance, incorrect or incomplete product data has for decades plagued the world’s largest retailers.

Here are some of the retail nightmares faulty product data has caused:

Retailers have tried and failed to alert shoppers to recalls of products they recently purchased. 

Retail sites and mobile apps constantly tell shoppers that an item is available for purchase when it isn’t. 

Retailers can’t alert customers when a product they searched for two months ago becomes available.

Apps tell customers the location of a product in a brick-and-mortar store, but the item is nowhere near that location.

Much of this stems from unreliable data from retail suppliers, such as Procter & Gamble, Unilever, Colgate-Palmolive, Johnson & Johnson, Kraft, Mars, General Mills, and thousands of others across the planet.

Unreliable data is a broad category that includes incorrect information, but also data placed in the wrong field and data that is materially incomplete or not included at all. Have you ever visited a grocery site and clicked on the “ingredients” button and were only shown a nutrition chart with no ingredients listed at all? That is an example of data placed in the wrong field or cell.

Various genAI vendors — with Google arguably taking the lead — say that genAI and agentic AI may be able to improve the product data situation, in much the same way that agentic may dramatically change ecommerce interactions. Could an army of software agents that tirelessly scout out and correct faulty data be the solution?

The implications stretch far beyond retail. Observing how agentic AI can address specific technical challenges in retail may show IT leaders in a variety of industries how AI might help them fix their own tech headaches.

Product data (un)reliability

How bad is the quality of data being shared by retail suppliers today? That question is almost impossible to meaningfully answer because of the vast difference in the kinds of products being sold, as well as the sub-verticals (pharmacies versus groceries versus apparel chains), geographies, and types of information sought by those different retailers. 

Given those variables, retail experts understandably offer very different estimates. 

Melody Brue, a VP/principal analyst for Moor Insights & Strategy, puts the supplier data reliability percentage at between 65% and 80%. 

Abhilasha Sharma, practice director at Everest Group,offers a “supplier data reliability range of 60% to 65%, with higher reliability levels [of] more than 85% in areas such as identifiers (GTIN/UPC, brand) and pack/size/dimensions/weight” and less than 50% for suppliers that need to deliver informational images.

“Accuracy on these parameters is increasing with gen AI/AI adding to the equation to flag inaccuracies,” Sharma noted, “but additional parameters like videos are adding complexities” and therefore reducing typical accuracy. 

Sam Vise, CEO of Optimum Retailing, was the most pessimistic, saying the accuracy of supplier data is “at best 50% or slightly under that.”

Paul Tepfenhart, global director of retail at Google, was the most optimistic, scoffing at the 50% accuracy figure. “I would say 80% to 90% is typical where I have been. Granted, these are top-shelf retailers, but please: 50%?” Tepfenhart said. “I know of one [supplier] that is about there, but [that is] certainly not the norm.”

Even 10% to 20% unreliability in product data can cause headaches. Tepfenhart tried to put 80% reliability into context.

Roughly, “80% percent means that every online basket has a couple of wrong or OOS [out of stock] items, which is a constant pain to shoppers,” he said. “Large global retailers are going to look at every item with 120 to 250 attributes. Multiply that matrix by two million items. Walmart, for example, has about a half-billion items. That is a staggering math problem.”

AI to the rescue?

Even if AI can improve retail data reliability, can it improve it enough that retailers are comfortable changing aspects of their operations? For example, retailers briefly tried — and stopped — alerting customers to product recalls because the data was so flawed. They found themselves alerting lots of customers who never purchased the item and not reaching out to many who had. It was a lawsuit waiting to happen.

Another example: Most retailers won’t stress how many items are still in stock in a store because that data is also of low reliability. Once retailers are confident in the data reliability, they will revisit these situations.

Getting retail IT executives to that level of confidence in supplier data means that the reliability has to hit — depending on the retailer — anywhere from 90% to 99%, which is higher than where almost any supplier is today.

Some retail experts question whether suppliers will invest sufficiently to get their accuracy that high, but Tepfenhart suggests that even small increases are worth it.

“I think ‘great’ is the enemy of ‘good.’ There is much to be gained by a step change in accuracy. By setting a lofty and unachievable initial goal, it paralyzes the organization to bother making any progress,” Tepfenhart said.

Google’s approach to fixing supplier data reliability issues is to use agentic systems to evaluate supplier data, flag any issues, and then directly share those concerns with the suppliers.

The initial phase, Tepfenhart said, is to have the AI “evaluate what’s missing or doesn’t make sense and to catch errors, flag missing or incorrect information, and to then trigger it back to the supply base: ‘These five items have to be revisited. Here is our suggestion.’”

That is phase one. Phase two is where the agent doesn’t even need to contact the supplier. It makes its best guesses and tries to fix the data on its own.

“After a while, the AI learns,” Tepfenhart said, “and we are then going to just start fixing it ourselves.” Retailers are currently spending “a ton of money” manually fixing errors, he noted. What’s more, he added, there are critical health considerations at play. “Some of these [ingredients] are deadly allergens.”

Moor’s Brue said that Google’s approach might work, assuming that both retailers and suppliers cooperate with the software. 

“That kind of data feedback could drive percentages of improvement immediately,” Brue said, adding that suppliers might have to assign personnel to deal with the AI suggestions. “Right now when you are using AI in these scenarios, you may have to have a person who responds to that inquiry.”

Better data leads to QR codes leads to better data (in theory)

One of the key retail changes that could be fueled by a sharp improvement in supplier data reliability is the much-delayed global retail shift away from barcodes to QR codes. QR codes can house far more information than the current ubiquitous barcodes, but it only makes sense to consider the shift if retailers are confident that suppliers are providing reliable data.

“Barcodes are limited to 25 characters, while QR codes can hold thousands of characters,” said Optimum’s Vise. 

And given that QR codes can include lengthy URLs — with the data housed on a web server somewhere — the amount of data that they can present is almost unlimited. 

With a QR code, Vise said, retailers can access everything from shipping method to date of manufacture and how long it can sit on the shelf. It could track the product’s complete history of temperatures, which is critical when deciding whether to accept pallets of perishable items.

“There’s just not that much you can do with 25 characters,” Vise said, referring to barcodes. 

If AI and automation improve the product data coming from suppliers, retailers may at long last be willing to invest in the equipment, software, and infrastructure changes needed to move from barcodes to QR codes — and that in turn would allow for richer, more useful data associated with each product, to the benefit of both retailers and their customers.

That said, the move to QR codes does not rest with retailers alone. It’s a massive global change that requires agreement and coordination among all suppliers, supply chain players, and retailers worldwide, along with a standards group that will dictate the exact specs of the QR code — no mean feat given that QR codes are orders of magnitude more complex than barcodes.

But until retailers are confident in data reliability, why bother?

The planogram problem

Another retail area that Google wants to try and improve is planograms, which are the store maps indicating which aisle and shelf products should be on. In theory, a proper and accurate planogram would indicate precisely where a product is, which would allow a mobile app to point customers to the exact location of, say, a specific cereal. 

The problem is that planograms are very rarely accurate. One key factor is that many stores are given discretion to change product placement to better cater to their customers. But those local store changes are rarely fed back into the planogram system.

“Planogram software doesn’t talk with inventory systems. Most retailers’ planogram tools and inventory systems are siloed,” Vise said. “There’s no live sync between what’s planned and what’s available. Even if the initial planogram is accurate, it’s irrelevant within days because of fluctuating inventory, supply chain delays, or sell-through. Without automation or AI, there’s no way to keep planograms relevant without a huge amount of manual work.”

The retail gold standard for tracking products and mastering the planogram may be Amazon Go, the company’s set of “grab and go” retail stores. Shoppers check into the store with an app, take the items they want, and leave, with the amount they owe for the goods automatically charged to their Amazon accounts. The system uses a sophisticated array of digital cameras, sensors, and analytics to track every product, every shopping cart, and every customer at all times. 

But Amazon Go stores are seen as prototypes of what could be, in that they are reportedly far from profitable and are not expected to see profitability. 

That hasn’t stopped startups such as AiFi from launching similar efforts. The vendor sells its “camera-first spatial intelligence platform” to retail stores to enable checkout-free shopping, with automated inventory management and planograms. The company’s “computer-vision AI is growing quite quickly, and it can handle seeing the entire store,” Brue said.

Google’s approach to in-store product tracking is more scaled-down. Referring to large retail chains implementing an Amazon Go level of real-time data gathering and analysis, Tepfenhart said, “there is no way that is ever going to pay out.”

Instead, Google is trying to deploy “an affordable augmentation” that relies on having scans performed on each aisle at some set interval — say, once a day — rather than constantly. This could be done by an array of fixed cameras or a drone that flies around the store. The software would then analyze the images or video and automatically update the system/site data and the text reflecting that data.

“You don’t really need continuous monitoring,” Tepfenhart said.

Vise sees some budgetary wisdom in Google’s in-between approach. “Computer vision is extremely expensive,” he said.

A system that leverages once-a-day scans would mean a more modest outlay for stores that aren’t ready to go all in on computer-vision AI.

Applications beyond retail

It seems likely that a combination of these approaches, all based on different deployments of AI, will improve product data reliability and comprehensiveness — and ultimately the customer experience, both online and in physical stores.

From a data analysis perspective, what fixes retail’s problems could theoretically work for any other vertical. It’s a matter of IT leaders looking at their vertical’s data problems, then reexamining the root cause in an AI context to see if video analysis, a set of agents, or other AI tools could cost-effectively fix the headache.

In healthcare, for instance, genAI tools might help overworked ER residents pull key data from patients’ massive medical records. In finance, an agentic AI system might verify transactions without introducing a delay, thus improving fraud prevention. The list goes on.

Just about every vertical can learn from retail. What problems might automation and AI agents fix in yours?