Because Of Click and Pick, Retailers Now Realize Their On-Shelf Availability Scores Have Been Wrong For Decades

(Note for non-retailers: On-Shelf Availability (OSA) is THE most important metric for a retailer’s supply chain, in-store operations, and customer satisfaction. OSA is calculated as the percent of items that a customer can currently buy in your store without help from an associate.)

Retailers have long searched for a customer satisfaction measure that mattered. Through well documented studies and surveys, many have landed on some pretty basic discoveries, which admittedly vary across retail segments:

• Consumers still go into retailers to buy products and in the vast majority of instances, take the product with them.
• Consumers often tend to look for associates after signage, on-shelf inventory or product packaging fails them.
• Experiential areas tend to rank below 1) having the product on the shelf, with 2) helpful associates there when I choose to engage, and 3) a quick checkout experience.
• Depending on segment, a retailer focused on driving the “experience” would already have fine tuned their store navigation & queue strategy and can set their # 1 metric as On Shelf Availability & #2 metric as Relevant Associate Availability.

Almost all retailers use some form of inventory position in the store to track their service level. We’ll say something like “we are 98% in stock for the customer.” And that’s a great number — most supply chain executives will tell you that’s around the point of diminishing returns. The problem is, when you walk the aisle you intuitively know numbers like 98% can’t possibly be true — even though their corporate data is right there in a black and white report — 98%. The amount of raw data behind this number makes it seem real, allowing us to claim victory and move on to less tangible “experience” topics to improve customer engagement.

However, as click and pick programs get rolled out, and substitution rates hit 10–15%, retailers are realizing that their OSA numbers cannot possibly be 98%. If their own employees can’t pick from their shelves, how can their customers? CPGs, Instacart, click and pick programs, and my own store visits report markedly lower numbers, 90% or sometimes even lower.

Even these inventory numbers are a point in time, in between which out of stock instances occur and are refilled. When you look across time in a category it gets worse. I’ve seen categories where 30% of the SKUs had an shelf level OOS scenario for more than 4 days over a 30 day period. Taking this daily or intra-daily perspective to shelf level out-of-stocks paints a much clearer picture of customer experience. A picture where customer facing out of stock instances happen at an alarming rate, and could cost the typical grab and go retailer 5%+ in lost sales every year. Sales that are rightfully theirs given traffic, sales floor staffing, and system-wide inventory dollars. The report vs reality discrepancies tie back to how, where and when we collect the data.

Twenty-four percent of Amazon’s online revenue comes from customers who experienced an out-of-stock at a local retailer (IHL study). Often, sales floor staffing has been reduced to the point that shoppers in need of assistance will be hard pressed to find an associate. All the while, inventory dollars really haven’t shrunk as much as they’ve become less productive. Luckily, most of the customer facing out of stock issues can be solved without increasing inventory dollars. That requires a new level of data — SKU/store/time level data — that was previously unavailable. It also requires a new level of accuracy, provided by AI sensors. Specifically, computer vision and deep learning, that has only now become affordable.

I have tested this level of highly accurate SKU/store/time data to drive real time shelf restocking productivity, adjust shelf holding positions to localize planograms, and reduce the amount of unproductive inventory in stores — to recapture 5–6% in sales. This same data can change the way operations and merchant executives think about out of stock. This is because analysis at that level creates a truer sense of lost sales that allows for faster, data driven prioritization of solutions.

Manual Scans and Accounting Systems Provide a False Sense of Security

Traditionally, retailers (including myself) have measured OSA by associates walking the store every morning for 3–4 hours, and “shooting the holes”. There are 3 major issues with this habit:

1. The wrong time. Imagine you were trying to estimate the average water level of the ocean against a retaining wall, but you measured the water height at high tide every day. Your average would be very wrong. Manual scans are measured exactly this way. They are done first thing in the morning, not at night after customers shopped the shelves, but right after the store spent 40+ hours of night labor restocking, so this paints the rosiest picture possible. OSA falls by 3–10% in the course of a day (nevermind that most of the time OSA isn’t at 98% to start with). Figure 1 shows OSA data collected in a major retailer on the East Coast for 6 months across 3 stores measured hourly. You can observe average OSA starts at 8am around 97% range and ending at 89% and then the night crew restocks to get them back to 97% again.

Figure 1: The average hourly OSA for three stores of a major retailer shows steep drops throughout the day

2) Inaccurate. Manual scanning is a tedious task that requires long periods of uninterrupted concentration, that associates typically hate. It’s not surprising to hear them describe the process as follows:

“I mean, it’s not very interesting. It’s just something I have to do every day. After an hour, your eyes start to glaze over.”

The store associates are unlikely to put their best foot forward for a task that is perceived as something “I just have to do” and instead will likely end up glossing over the shelves, checking the box, quickly looking for obvious holes and scanning them. This results in very low accuracy. I have measured that associates miss ~15–20% of outs in independent audits. Additionally, during the scan the associate will encounter a number of issues on the shelf that exacerbates the inaccuracy of the scan, such as some outs being faced over, and since the associate is not measuring against a planogram, they miss those outs.

3) Expensive. Manual scanning is an expensive process to determine your daily on shelf positions. The monthly cost of just one scanner for 4 hours a day @ $15 an hour across 30.5 days is $1,830. To solve for this, I worked with a number of AI companies that attempted to provide the ability to scan our shelves, process the scans to identify problems, triage the scenario and provide real time in-store action. The scan should be the easy part — and we tried robots scanning the aisles, drones hovering around, and static cameras and sensors on the shelf. As can often be the case in mass retail, the simplest solution worked the best to capture the scans. Implementing the shelf cameras and feeding their images into a deep learning data set resulted in huge new insights & actions that blew up old philosophies and allowed us to measure productivity, increase OSA, and grow sales. This technology I believe to be a significant step forward for retailers who are serious about digitizing their stores and improving their operational efficiencies.

My own past projects have met with easily understandable objections:

If people just followed the process, then IMS (Inventory Management System) would be accurate, and then Perpetual Inventory (PI) would solve the problem right?
If retailers could have achieved this with people alone without new hardware sensors, then they would have solved it by now. It is just not practical. This is because inventory touches so many different people and there are so many opportunities for things to go wrong. Inventory systems make a lot of assumptions and each department is guilty of causing a small amount of inaccuracy that in summation results in huge amount of wrong inventory. Many stores perform full store manual inventory counts maybe once a quarter that allow them to see “how wrong we were”. Figure 2 below shows example error rates that colleagues in the big box grocery space have experienced. As much as 66% of inventory is at least 1 unit off, 28% is more than 10 units off, 38% is more than 50 units off and 19% is more than 100 units off.

Figure 2: After taking physical inventory quarterly, colleagues in big box grocery reported how off their inventory counts actually were, as measured by number of units. Only 34% of inventory was perfect. 38% of the inventory was off by 50 units or more.

Here are the typical causes of these inaccuracies.

• Receiving/Warehouse/DSD causes 3–4% inaccuracy. Inbound receipts require either scanning as cartons come off the truck or we can depend on warehouse’s / vendor’s count. Both methodologies result in error to “credit inventory”. Mis-selects (as they are called) average 3–4% of an order. Cartons are incorrectly scanned in as part of the inbound scan or missing as part of an assumed receipt or worse yet the product is completely wrong (i.e. the carton’s contents are not matched to the label on the box.)
• Shrink causes another 3% of inaccuracy. The use of point of sale “scan data” to debit from inventory is also error prone. It assumes a correct starting point and zero breakage, theft, and mis-scans. Depending on segment, shrink can average 3%.
• Click and Pick timing. Some retailers do not debit Click and Pick orders until after they are picked, so the item was already sold, but it still shows up in inventory.
• Cycle Counts cause another 3%. Store managers and stockers may have a habit of “zero-ing out” or “cycle counting” inventory that they can not find in the store. This causes the system to order more, even though it may be within in the four walls. The opposite side of this can be worse — thinking product is somewhere within the store when its not.
• Misplaced products. Customers tend to move products all the time. So you may think you do not have the inventory in the store, but you do, it is just in the wrong place.

Can’t a person do the same work with a scan gun?
Not as efficiently, effectively, or often enough to make a difference. Many have tried entire teams deployed to fight a losing battle better served by simple technology. Additionally, as outlined earlier even with much effort at the end of the day this is a mundane task that associate’s do not want to do and are currently doing with subpar accuracy.

Robots & drones sound amazing — why not let them do this?
It is all about the accuracy and cost of the data. They probably can. But not often enough, fast enough, accurate enough, and/or cheap enough. How much do they cost per scan?

1. Our baseline is humans, which take 3–4 hours to scan once a day and they are at 80% accuracy. Today the average minimum wage in the United States is trending towards $12/hour, so total effective wages and benefits is around $15/hour. That is $45–60 per store scan.
2. A robot for example costs ~$10k a month and can only scan twice a day making the plot in Figure 1 impossible to produce. Assuming the max coverage would be 2 scans per day, 30 days of scanning is about $166.66 per store scan. Also, in bake-off’s I have seen, robots have reported 70–80% accuracy.
3. A drone can stay in the air for limited time, which created questions of how many aisles it could scan per charge and frequency of daily scans. From my discussions with companies the accuracy is worse, and the number of scans is less than a robot. So this is more than $166.66 per store scan.
4. Shelf camera companies charge ~$6k a month, and they can scan 17 times a day to give you the most complete picture of what is happening in your stores, which is $11.76 per store scan.

While robots and drones sound cool, as the Washington Post reports they can scare customers and employees alike. Check out this chart for a comparison between currently available out of stock solutions.

In conclusion, retailers are realizing that their inventory systems & processes are flawed in a number of ways and that the current 95-98% service levels they may report is often a feel good metric that does not represent the reality for the customer in the store. The current methods of scanning in the out-of-stocks manually or relying on perpetual inventory have severe downsides that lead to inaccurate input data.

In order to make On Shelf Availability a useful metric & improve customer experience, better input data is a prerequisite. Advances in technology have brought a number of new options for this purpose to the market in recent years. My first hand experience with shelf cameras, using very inexpensive cameras powered by computer vision & deep learning to monitor a store’s shelves throughout the day, has shown the most benefit & fastest ROI. Accurate & frequent SKU level availability metrics allows efficient, automated micro- and macro-level inventory decisions that improve a store’s inventory position in ways simply not possible with prior approaches.

A new breed of real-time, action-oriented AI tools will empower stores to improve satisfaction where & when it matters. It’s time to start collecting accurate data where customer action happens — at the shelf.

About the Author
Jonathan Luster has decades of experience in the retail industry in a variety of leadership positions at Lowe’s and Staples. Jonathan has also been an invited speaker at numerous retail industry events including the National Retail Federation Big Show, Knowledge@Wharton’s Retail Summit, Feicon Batimat and Retail Spaces.