AI in retail: a blueprint for implementation

Arvid Stenback

CRO & Co-Founder

Thom Iddon-Escalante

Marketing Director

Change is everywhere

COVID has catalysed entirely new retail practices; take in-store social distancing or dedicated shopping hours. But more often it has accelerated change already underway, like online grocery shopping and click and collect. This has confronted Retailers with two often conflicting challenges:

  • Meeting the short-term demands of shifting consumer behaviour and government regulation
  • While grappling with the pandemic’s longer-term implications for retail in general

AI: momentum is gathering

Grappling with both these challenges has accelerated the adoption of AI tools and technology. Back in 2019, Capgemini forecasted that by 2022, in operational improvements alone, AI would have an annual $340bn industry-wide impact. In retrospect, these numbers may well prove to be conservative.

A blueprint for implementation

We’ve previously tackled the issue of whether retailers should build, buy, or subscribe to AI technology, but here we’ll focus on the nitty gritty of implementation.

Recognising the need for AI - even kicking off an AI programme - is not the same as making effective use of it. But for an industry already facing obstacles and upheaval, how can retail executives chart a successful course through these uncertain waters? 

#1 Data first. AI second. 

If you take one thing away from this article it should be this: 

The quality of insight AI serves retail is dependent upon the quality of data retail serves it.

In other words, AI is only as reliable as the data upon which it is trained. Retailers generate vast, vast amounts of data. Building a coherent data strategy and developing complete and correct data models is an investment that will pay off manyfold; not least because retailers then open themselves up to a rapidly growing market of quality third-party AI solutions.

At the same time, the challenge of arriving at this point should not be underestimated. As Arvind Krishna of IBM has stressed, 80% of the work of an AI project is collecting and preparing the data. Some companies are not prepared for the work associated with that. Others need urgent support to be data-ready. 

It’s time to get your (data ware)house in order. 

#2 Prioritise problem areas 

While straightening out the data, it’s time to prioritise areas for improvement. Most retail executives we speak with could generate a list of problems longer than an Xmas shopping receipt. This is a good starting point, but to determine whether a problem is conducive to an AI solution, it should be tested against the following criteria:

5 criteria for determining whether a problem warrants an AI solution: 

  • Is the problem sufficiently valuable to warrant AI investment? Many aren’t and can be better solved by more manual means, as Walmart recently concluded.
  • Is the problem sufficiently complex to warrant AI investment? Occam’s razor: simple problems are usually best served by simple solutions.
  • Is the problem sufficiently narrow in scope? Contrary to what we see on the big screen, there's no such thing as general AI. AI needs to be targeted towards specific problems and deployed within set parameters to be successful.
  • Is there enough quality data on which to train the AI? AI is only as good as the data on which it rests. The higher quality the data, the more likely an AI solution is to meet your objectives.
  • Can an AI solution deliver measurable improvement within 3 months? Retail executives’ most frequently cited obstacle to AI adoption is cost. One way to avoid this is to pursue “quick wins”. Start with one core problem and prove AI effectiveness before moving on to the next.

Any problem that fails to meet these criteria is likely a poor candidate for a retailer’s first foray into AI.

#3 AI is challenging. Help is at hand.

You’ve whittled down your Xmas shopping list to a top-3 wish list. What now?

Retailers sit along a broad spectrum of technology competence. Tech-first giants like Amazon and Alibaba occupy one extreme. With their massive R&D budgets, giants like Walmart and Carrefour aren’t far behind. The majority of smaller, regional retailers cluster towards the other end. This is not to diminish these retailers; far from it. It’s a reflection of retail’s intensifying state of competition and diminishing returns. And it’s precisely these retailers that Formulate and other AI providers seek to help; offering targeted AI solutions that can compete with the best big retail can conjure up.

Why not go at it alone?

It’s a fair question. While it’s relatively easy to create an AI proof of concept within a lab environment, it’s much more complex to develop a solution from scratch, put it into production, and maintain it thereafter, all the while ensuring it is accurate, accessible and accountable. Consider, for example, how much of a curveball COVID-19 has proven for prediction modelling. We’d argue this is not a problem retailers should be preoccupied with. Moreover, the current arms race for retail-specialised data scientists makes finding and retaining talent all the more challenging, especially for smaller retailers.

Case study: solving the promotions challenge with AI

Let’s test Formulate against our 3-step theory.

#1 Data first. AI Second

Since AI is only as good as the data upon which it rests. We always work with retailers to ensure their data is optimised before we get to work. It’s in everyone’s best interests.

#2 Triage the problem areas

Promotions are certainly ripe for AI investment. Judge them by the 5 criteria:

  • Is the problem sufficiently valuable to warrant AI investment? Yes. Retailers typically run thousands of promotions every week across hundreds of stores. They’re enormously valuable, typically accounting for 20-50% of total revenue.
  • Is the problem sufficiently complex to warrant AI investment? Yes. to evaluate the effectiveness of a promotion, it’s necessary to calculate the baseline, and account for switching, halo, fixed and variable supplier funding, weather, calendar, and, as of 2020, COVID-19 effects. Failing to accurately account for any one of these factors could easily skew the analysis.
  • Is the problem sufficiently narrow in scope? Yes. It’s easy to define what a promotion is. They operate under very specific rules and conditions.
  • Is there enough quality data on which to train the AI? Yes. It’s possible to run highly sophisticated promotions analysis using receipt data and structured promotions data alone, both of which retailers possess in abundance.
  • Can an AI solution deliver measurable improvement within 3 months. Yes. With sufficient data it’s possible to glean valuable insights from AI promotions analysis within days and act upon them within weeks.

#3 AI is challenging. Help is at hand

We’ve spent years honing our models to better evaluate historical promotions and forecast future ones. But even we still get surprised sometimes! The truth is AI is hard, even for those who specialise wholly in it. 

Still, our models have been proven to improve the profitability of promotions by more than 15%, saving retailers hundreds of millions Euros every year.

If you want to learn more about how Formulate’s AI could power your promotions, schedule a meeting with an advisor.

Book a demo

Like the article? Spread the word

Sign up to receive stats, data, and insights that will help you keep up with retail trends

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
English