Bad Decisions Faster: Why Store Inventory Accuracy Decides Whether Retail AI Works

The Fastest Transformation in Thirty-Four Years

Greg has watched every retail technology wave since networked point of sale. This one moves fastest, and it is splitting the industry.

Greg: “There’s such a disparity between where retailers are in readiness for AI that it’s almost scary, in the sense that there’s a lot of retailers being left behind, mainly because they’ve underinvested into things that get to clean and accurate data for their inventory.”

The store sits at the center of that disparity. Walk-in traffic is declining, yet the store keeps its role in almost every transaction as the fulfillment point for local delivery, buy online pickup in store, and click and collect.

Greg: “Still roughly about 89 percent of all transactions have a store component. That store level inventory accuracy and the quality of that data is critical to success in those areas. If you put AI on bad data, all you do is get really bad decisions faster.”

Garbage in, garbage out describes a static failure. Bad decisions faster describes an accelerating one: AI on inaccurate inventory multiplies the damage at machine speed, across every store, before anyone notices.

AI Is a Recipe More Than a Technology

AI is finally connecting the store and e-commerce, and the constraint sits somewhere most executives are reluctant to look.

Greg: “There’s no deficiencies in the AI itself. The deficiencies are in the data. So that’s the first step, clean data. Do you have an accurate view of what you have for inventory in your stores? Are the descriptions accurate?”

Data quality and network quality decide the outcome, which is why AI works as a recipe. And the data now has a second audience beyond human shoppers.

Greg: “More importantly now, are the full descriptions on the website ready for agents to be shopping at, not just a human being looking for solutions there? I found out that just having FAQs and a list of answers doesn’t necessarily get you found. In the age of agentic, you actually have to have it in a JSON format so that you’re seen by the search engines and by the bots.”

Sathish notes that CommerceShop built ConversionBox for exactly this problem. Product data now serves two readers: a person tolerates gaps and infers from photos, while an agent reads structured attributes and drops any product missing a specification.

The Journey Most Retailers Left Unoptimized

The biggest near-term opportunity lies in a journey most retailers have neglected.

Greg: “Most people have optimized that walk-in store journey, but they haven’t optimized the journey where the store is a fulfillment engine for an online store, whether that’s picking the items for buy online pickup in store or click and collect, whether it’s local delivery. Do you have the items in stock? Do you have the layout in the right order?”

Every online order fulfilled from a store adds labor that a walk-in sale avoids entirely. Delivery fees used to offset it, and increasingly those fees get waived to stay competitive, while the added labor remains.

Labor during walk-in peaks is the second lever, and IHL is finishing research on the third: using AI to reduce shrink from theft or honest mistakes, including merchandise locked up so tightly that shoppers choose the exit over waiting for help.

A store running as a fulfillment center needs the discipline of a warehouse, and most stores were built and staffed for browsing.

The $1.7 Trillion Inventory Problem

Demand forecasting predates generative AI, so has the inventory problem improved? IHL’s own figure answers that.

Greg: “We lost 1.7 trillion dollars last year due to not having in stock what the shopper wants to buy, or it’s not in the right place, or we don’t know where it is.”

IHL’s research defines that figure as inventory distortion, the combined cost of out-of-stocks and overstocks, equal to about 6.5 percent of global retail sales, with out-of-stocks accounting for roughly $1.2 trillion. The store makes the problem hard because it keeps changing.

Greg: “It’s a living, breathing organism. You have theft, you have people ripping things off, you have stuff getting broken there, and if you’re not getting to an accurate view of what’s actually there, you can’t sell it.”

Then the number that frames this entire episode.

Greg: “The typical consumer thinks the retailer is in stock about 75 percent of the time, where the retailer themselves thinks they’re in stock 90 percent of the time or more. That delta is the difference in expectation.”

Generative AI makes zero difference here until the data is right. Once it is right, the upside is concrete: promotions timed to what is on hand, prices nudged when a hot day lifts demand for cold items, impulse items for walk-in shoppers, and one more relevant suggestion at online checkout.

Amazon shows the last point with a grocery trick: every bag of frozen or fresh items arrives with a frozen bottle of water, costing around four cents and protecting the strawberries more cheaply than a cold pack. Greg stopped buying bottled water years ago and has cases of it anyway.

A retailer believing it is in stock 90 percent of the time plans promotions, staffing and AI projects on that belief. The shopper’s experience is fifteen points worse, and every system trained on the retailer’s view inherits the error.

Why Smaller Retailers Hold an Advantage

Can a small retailer afford any of this? The budget differs, and the fundamentals stay identical.

Greg: “The key blocking and tackling here for the small retailer is exactly the same for the large retailer. And actually the large retailer has several disadvantages that the smaller retailer doesn’t have.”

Organized retail crime targets large chains because identical store formats cluster regionally, letting crews hit several quickly. The smaller retailer usually knows its inventory far better, and once that inventory is electronically categorized and accurate, the same AI tools apply: forecasting, pricing, markdowns and labor allocation.

Theft follows traffic. The 4 to 7 p.m. rush at a grocery or drug store lifts both employee and consumer theft, and hot promotional items get pilfered more. As retailers watch the front of the store, more loss moves out the back door, where rushed staff sign off on deliveries before counting them.

Large retailers carry one more loss unrelated to theft: vendor-managed inventory that fills a shelf to look full while ignoring the space plan.

Greg: “Anybody that drinks Diet Coke knows that Coke Zero is not the same as Diet Coke. If you fill up Diet Coke with Coke Zero, your sales of Diet Coke are going to go down and you’re not going to see the relative increase in Coke Zero as a result.”

A full shelf with the wrong product records as in stock and sells like an empty one, and only a check of what actually sits on the shelf reveals it.

Personalization and AI on the Store Floor

Online personalization draws on clicks and browsing history. In the store it relies mostly on coupons and offers, and real-time personalization in the aisle faces a legal headwind more than a technical one.

Greg: “It’s being perceived and being promoted in the media as if your price is going to be higher because of the color of your skin or your income level or how you’re dressed, rather than your actual sales level within the organization.”

So personalization has moved earlier in the journey. The app tells a frequent buyer about extra coupons or a two-for-one before the visit, avoiding the scrutiny an offer triggered in the aisle attracts.

On the associate side the case is straightforward. AI customer support brings new staff to expert level quickly by drawing on the history of past calls, even if it adds little for the best veterans. The same logic applies to a store associate carrying a handheld with product, configuration or health information.

Zebra points the way with slim models running on the device, fine-tuned to a specific retailer and associate, with no general chatbot licenses. A small model trained on one retailer’s products answers that retailer’s questions better than a general model that knows a little about everything.

Shopping Agents Are Already Here

How fast is agentic shopping moving? A domain name bought as a joke answers that.

Greg: “I bought the domain Retail Agent Con in March as a joke, to say for April Fools we’re going to put out that there’s a new conference that’s agents only, no people allowed. Well, the reality is the world has changed so much that we’re actually putting on a real conference using that domain in two months.”

Agents now handle shopping, pricing, promotion and marketing inside retail systems, and consumers deploy their own agents to shop. Few retailers are ready for either, and the stakes follow from where shopping begins.

Greg: “If 80 percent of all shopping starts with an online visit of some sort for investigation, if you don’t do this, you are effectively invisible.”

At minimum, retailers need agent-optimized sites and product data. The harder question is who owns the customer, and delivery platforms show how that plays out. When an order goes through DoorDash or Instacart, the platform holds the relationship and the data. When it passes through ChatGPT, Anthropic, Google or Walmart, a mid-size retailer carries far less leverage than the giants.

Greg: “We used to think it was SEO. It is nothing to do with SEO anymore. It’s all about AEO and GEO.”

Sathish raises Perplexity’s stated goal: tell an agent to order a pizza, and it finds the nearest location, orders and arranges delivery, with the shopper indifferent to the source. Does that make retailers interchangeable? Expect a fight, decided by how much value each party brings to the model, and there is a cautionary precedent.

Catalina had the technology and the manufacturer and retailer relationships to run retail media networks fifteen years ago, through the coupons its printers produced. CPG companies and retailers each wanted to own it, and the tug of war delayed retail media by about fifteen years. Agentic commerce will allow far less time.

Every party insisted on ownership, and the value sat unclaimed. Retailers treating agentic commerce as a negotiation to win before participating risk repeating that while the agent platforms build the relationship themselves.

The Store Becomes a Fulfillment Center

Does agentic commerce finally replace the store? The agent places the order, and fulfillment still happens somewhere physical.

Greg: “For the large format retailers, their stores are their advantage for distribution. Grocery, health and beauty, those mass merchants, it’s more related to that local piece of it. Proximity is the key.”

More of the actual sales will shift toward e-commerce journeys, which forces a rethink of the store itself: less space for walk-in shopping, more for staged pickup and delivery orders in the back.

Greg: “One of the data points that just slapped me in the face was finding out that 80 percent of all items in a grocery store sell less than one a week. All those spices and all that stuff doesn’t need to be taking up shelf space.”

Reclaim that floor space for optimized picking, possibly with robotics, and costs fall sharply while a smaller footprint shrinks the theft problem.

On robotics, the payoff today lies in inventory accuracy. Electronic shelf labels, RFID and computer vision free up labor to stock shelves, where people still beat robots across the variety of merchandise in a store. Staffing is the pressure pushing automation forward.

Greg: “The average sit-down restaurant is missing about four people that they had in 2020 prior to COVID. That’s why your service is worse. That’s why you see everything moving to kiosks for the front end.”

If stores are the fulfillment footprint and too few people exist to staff them, robotic picking becomes a requirement. The 80 percent figure turns a vague ambition about the store of the future into a specific floor-plan decision: which slow-moving items leave the shelf for a picking zone.

Where to Point AI First

Enterprises are deploying agents across merchandising, supply chain and customer service. The priority is bang for the buck, starting where data is cleanest and opportunity is largest.

Greg: “If I can increase sales, that’s a top line thing. If I’m just reducing cost for the average retailer, that’s only 28 percent of the P&L. So I can go after the hundred percent and increase that, or I can go after reducing the loss at the 28 percent.”

That usually points to sales-side work first: personalization, pricing, delivery options and promotions. Document drafting and invoice processing get attention because ChatGPT does them quickly, yet they move far less of the income statement.

On autonomy, the deciding question is the cost of an error. IHL, a research house, has run about 3,600 AI use cases and averages roughly 15 times faster than a human across them. Some functions that began near 40 percent accurate now run at 99 percent after repeated correction. Others still produce recommendations the team sets aside.

Customer-facing agents already work well for problem solving. When Greg misplaced his Sam’s Club credit card, the issuer’s agent validated his transactions and ordered a replacement in about 45 seconds, work that normally means 15 or 20 minutes on hold.

The 40 to 99 percent progression shows how autonomy gets earned. Accuracy improved because a person corrected the agent repeatedly on one narrow function. Retailers expecting that reliability on day one skip the part that produced it.

Three Mistakes Retailers Keep Making

1) Skipping the data work.

Greg: “Poor performing retailers who have not done the data work are trying to use generative AI as a way of catching up. And what they’re doing is basically running themselves out of business by doing that. You cannot skip the data part.”

2) Going for the home run first. Adoption momentum comes from smaller wins, and one ambitious project that stalls kills it.

3) Ignoring the culture change. Some employees fear AI will take their jobs, others see a tool that makes them more effective. Leaders have to read the room and overcommunicate that the goal is doing more with the same people, with AI taking the work everyone wishes they had time for.

The three mistakes compound. A retailer that skips data work, bets on one big project and neglects its people ends up with a failed project, a skeptical workforce and the same bad data.

The Road to LaGuardia

Some retailers got this right: Tapestry, Walmart, Home Depot, Kroger, Target, Lowe’s, Tractor Supply and, clearly, Amazon. They share a starting date.

Greg: “In about 2018, when Amazon went profitable on their e-commerce business, they saw that and realized that they were going to be in an IT race. And they started decoupling what they spend on IT from just the revenue line.”

The old rule tied IT budgets to revenue at about 2 percent: revenue up 4 percent, IT up 4 percent. The leaders broke that link, spent more and fixed their data. The metaphor places everyone on the road from Times Square during NRF to LaGuardia Airport.

Greg: “These guys are already through town, through the tunnel, on the Long Island Expressway, going full speed towards the airport because they did this work before. Most retailers are stuck between Eighth and Ninth on 42nd Street in bumper to bumper traffic, because they get started, bad data, I gotta fix this. Get started, bad data, I gotta fix this.”

That explains why the gap keeps widening. Every project the leaders start builds on clean data, while laggards restart each one by rediscovering the same problems.

Clean data still leaves the people question, and the answer is to overcommunicate, because reaction depends heavily on business conditions. When business is good, AI reads as help. When it is flat or struggling, it reads as a threat to jobs. A third group holds back because of liability, and one executive at an appliance retailer chose to be late to AI deliberately, with a reason that is hard to argue with.

Greg: “We don’t want somebody typing into a chat bot, your toaster burned down my house, and the chat bot writing back, yeah, sometimes it does that.”

Where a Retailer Should Start Tomorrow

The transition is running faster than anyone planned. One person Greg describes spun up a Grok bot, loaded it with friends’ and family’s birthdays, and set it to research gift ideas two weeks ahead from each person’s social profiles, suggesting where to buy. The recipient is entirely out of the loop, and that capability exists today.

The first advice for retailers comes in three parts.

1) Inventory accuracy and clean data. Go through the income statement and ask where the data is cleanest and where the opportunity is largest.

2) An agent-optimized online presence.

Greg: “We just went past 57 percent of all traffic now on the internet to e-commerce sites being agents. And we just went past that this month. If you don’t have it optimized for that, you’re going to be left out.”

3) Culture, and a new definition of every job.

Greg: “Teach people that they are going to be managers, regardless of what their role is. For them to be most effective in your organization, they are going to be managers of agents.”

Outline your job function by function, turn each into an agent, keep a human in the loop, then chain the agents together. Work that took ten steps might take two, and time drops by as much as 90 percent.

On synthetic data, Greg sees little reason for retailers to worry.

Greg: “Retailers have tons of data. They’re data rich, wisdom poor. It’s more which of this data is relevant.”

An A&W Root Beer executive once described running 20 different combinations of point-of-sale systems across his stores, so every data stream arrived in a different shape. The answer is a single version of the truth: one repository, with tools such as Databricks or Snowflake extracting what each use case needs.

Most retailers already hold plenty of data, and the job is consolidating it into one trusted source.

How Retailers Buy Technology Now

Procurement has changed as much as the technology. Retailers buy more general-purpose tools because AI moves so quickly, and the largest build their own and sell it: Lowe’s and Walmart both operate technology businesses that sell to other companies.

The old buying cycle took 18 to 20 months, which in AI terms reaches back before GPT-4. Retailers that delayed upgrades such as a Windows refresh now face a memory crunch with prices up 30 to 40 percent, turning procurement into a race to lock in prices.

Retailers are deploying more technology than ever across point of sale, payments, self-service kiosks, computer vision, RFID and in-store networks, and now spend an estimated 60 to 70 percent more of their revenue on IT than five years ago.

Laggards now spend heavily too, under price pressure and a compressed timeline, trying to buy their way through traffic the leaders cleared years ago.

The Next Three to Five Years

Calls to slow AI down read to Greg as mostly marketing and coordinated pushback, with some legitimate concerns mixed in. The near-term cost is stated plainly.

Greg: “I believe the job disruption’s going to happen faster, which is going to have a negative impact on the economy before the positives get fully baked in when it comes to AI. And I don’t like that.”

For retail, two things happen at once: poorly run retailers failing at a rate unseen since the start of COVID, and the biggest players growing bigger. The leaders are past the toll booth, opening stores and fulfillment centers more cheaply and taking share from weaker rivals.

Retail stays retail, and some shopping will always involve touching the product. More keeps moving online, driven by time pressure and growing trust in grocery delivery, leaving the store for things needed right now. The gap between technology haves and have-nots widens.

Amazon Go-style stores have a place in hospitals and on college and corporate campuses, with limited reach into standard convenience retail while organized retail crime costs hundreds of billions a year.

Retailers who fixed their store inventory data early are pulling away because every capability the AI era offers depends on that foundation. Everyone else is learning, at speed and at cost, that bad data only produces bad decisions faster.

Your Retail AI Readiness Checklist

Start with the first three. They take a month and show how far you have to travel.

• Measure your in-stock rate from the shopper’s side with store walks, and compare it to what your system reports
• Map which categories leave the store through pickup and delivery, and time each pick
• Check whether your product attributes exist in structured formats an agent can read, such as JSON
• Audit vendor-managed shelves against the planogram for substitutions like the Diet Coke problem
• List your slowest-moving items and model what reclaiming that shelf space for a picking zone would save
• Rank AI projects by how much of the income statement each touches, with sales-side work first
• Pick one narrow function, run an agent on it with a human correcting it, and track accuracy weekly
• Consolidate point-of-sale and inventory feeds into a single repository before adding new AI tools
• Ask every team member to outline their job function by function and flag which parts an agent could take