Why MishiPay Will Win - The Computer Vision Piece

Share this with your colleagues

#amazon announced in their last quarterly results that only 3 new Amazon Go stores are expected to go live until the end of the year – this is far from the 2,000 store goal they have put for 2021.

Here is why MishiPay is going to win.

To understand MishiPay, let’s first understand the current landscape. Labor shortages, rising inflation, higher energy costs are all forcing the retail landscape to find new solutions to urgent problems. Stores have tried as much as possible to cut costs, now there is only one solution: more automation.

However, what retailers demand is not what the traditional players can supply.

Traditional kiosks are not satisfying the requirements: bulky and hard to operate, take a lot of physical space and don’t fit in the user journeys for fashion (removing security tags), accessories, shoes, furniture and most other sectors.

On the other hand, computer vision stores (CV stores), read Amazon Just Walk Out, are currently making headlines as the next big thing in retail (link). However, with plans of opening 2000+ locations by 2022, the deployment/expansion has not been achieved (link, link).

In my view, the right way to assess why this is happening is in 3 levels: Scalability, NPS, ROI. Let’s review every aspect:

Scalability

Retail comprises fashion, shoes, furniture, travel, specialty, department, convenience and grocery.

CV stores currently operate small-scale grocery stores.
Only 1 Whole Foods store has been refitted and this took 5 years (
link). No other store has been refitted to work with CV. It only works when you build the store from the ground up, equipping it with the right kind of cameras, sensors and infrastructure. This means, new shelves equipped with specific sensors, cameras covering every sqm, carefully selected merchandise and limited SKUs (link).

Can we find a ZARA CV store? Probably not. The existing stores require a tremendous amount of data to go live, work with limited SKUs, identify items mostly from scales built in the shelves and items are carefully selected. The “sensor fusion” is the biggest limitation today prohibiting the universal application of this technology (link) but also it’s own scalability (link).

NPS

Do people like the experience? CV stores are easy to use and feedback has been great (link) . Small friction on the first time, as customers need to download an App but after that the in-store behavior is natural and seamless. Exit doesn’t have any obstacles. Customers can focus purely on shopping.

MishiPay Scan & Go takes a different twist. Instead of viewing the phone as a point of friction when shopping, we see the phone as the medium to communicate with the customer and enhance the journey. Every scan, is an opportunity to delight the customer.

At a minimum, after every scan, the MishiPay user is provided with a real time tracker of the basket total. Important for the customer who want to stay within budget but also when you carry items that do not have a price tag to remember which items cost what.
Secondly, we also provide information on offers: available and applicable. Did you scan an item that has a BOGOF offer? MishiPay App will notify you and prompt you to scan a second one and take advantage of the offer.
Finally, we also provide product recommendations (blog) allowing you to discover new products which increase the ABV (blog for increase).

Different use cases for different stores. Our NPS is north of 76 averaging 78 over the past 4 months.

MishiPay deployment across US Airports (link), has seen some of the highest NPS. The customer is on his way to the gate/plane, and needs to buy water, sandwich or a book. The store and it’s queue are not the end destination and the customer prefers to abandon baskets compared to missing a flight. Therefore MishiPay has an excellent fit allowing customers to scan, pay and go in less than 22 seconds.

Fashion & accessories are different but equally successful. The store is the destination and customers are ready to spend time. The app understands that customers require more product information (images, reviews, videos, descriptions) and serves them on every scan. This creates an enhanced, more informed user journey which results in better customer experience.

ROI

Saving money:

Traditionally CV stores focus on eliminating labour costs and shrinkage. This can be divided into direct labour savings and indirect (link). The total direct cost saving can be earned from eliminating:

Cashier jobs
Restocking/inventory counting

According to research, for a US store of 10 lanes the savings can be calculated at approx. $400k (link) – although the Amazon Go type of stores are smaller and the actual amount should be (almost) half of that. However, in reality all these stores still employ members of staff for restocking and the alcohol area is always staffed.

MishiPay has validated store cost reductions. Paradies Lagardere has seen a decrease of 9 hours of staff time/store/week per 5% adoption in the store.

At a maximum we can assume:

$12/hour x 9 hours saved x 5% adoption x 20 (to reach maximum of 100% adoption) x 52 weeks = $112,320 in savings per store

Making money:

Hypothesis: Stores without queues will attract more people (link). A study by Focal identified a 15% higher traffic to the store compared to nearby competitors and estimated that for every 1 minute of queue time saved, there is an increase of 1.6% more incoming traffic (link).

MishiPay has proved already that when customers don’t spend time queueing they spend more time shopping, buy more (link) and come back to the store 33% more often. The uplift is on 3 levels:

- Uplift of 25%-35% in value on every basket (link, link)
- Uplift of 65% in value because of recommendations (link)
- And 14% net increase of customer spend due to 33% increase in repeat frequency (link)

Shrinkage (aka theft):

Cost of audits

CV stores need to ensure that when customers walk out, the system has picked up exactly the right products, right quantities and is ready to charge the right amount. If for any reason the system can’t decide, then there is a “low confidence” event (link). For every “low confidence” event there needs to be a manual audit.

(link)

The percentage of the baskets that require an audit can be calculated based on the formula (link):

% requires to be audited = 1 – (1- low_confidence%)^basket_size

For a basket size of 8 items and a low confidence of 1%, this requires 7.73% of baskets to be audited (link). This requires a human to review the video and decide of what was bought – i.e. to label the data.

“Assuming the Amazon Go Platform has a 5% low confidence, this implies that 55,000 shoppers a year would need to be remotely audited per store. The cost of auditing a single 20 minute shopping trip may be ~$2.

This would imply a labeling cost of $110,000 / year / store which is the cost itself of 3–4 full time cashiers.“ (link).

Amazon has tried to resolve this in 2 ways. First, remove the Human In The Loop (HITL) and secondly the reduction of audits needed. The first approach was based on creating synthetic activity data using simulators (link) which augmented the training set. Secondly, by leveraging the low confidence events themselves. After the events were “clarified”, they were labeled accordingly and used as input to train the algorithm again and better. Therefore the system, when encountering the same low confidence event would know how to classify it, therefore reducing the % in the long term.

MishiPay has created it’s own IP to combat the same problem – the Trust Engine.Traditionally stores focus on random audits or sometimes based on heuristics passed during training. MishiPay transforms the “random” to educated audits. While customers enjoy their Scan & Go session, our algorithms analyze every action and determine which customers need to be checked based on their in-app behavior. Similar to the risk analysis performed once you try to use a credit card, MishiPay also calculates a score to determine the likelihood for an audit. Industry papers have also validated that mobile self-checkout brings much lower shrinkage compared to traditional kiosk (fixed) self-checkout (link).