The Promise of AI in Retail - with Eric Bogart
By Jeremy Fajardo
Eric Bogart is the Vice President of Advanced Analytics at Acosta Sales and Marketing.
On the western fringe of Toronto in Liberty Village, Mosaic’s offices are nestled within a larger reclaimed warehouse — the shared home of many other businesses. But small might not be an accurate description of the design and experiential marketing firm, having been acquired by Acosta Sales & Marketing in 2016.
Acosta is no stranger to acquisitions and is an established firm operating between CPG manufacturers and some of the biggest retailers and grocers in North America.
Like many firms who play a consultative role in their industry, Acosta has made recent investments in advanced analytics as a core competency of their enterprise. We sat down with Eric Bogart, VP of Advanced Analytics for an open conversation about his career, the future for AI in retail and the Swiss Army knife nature of his team.
The rise of “Data Science”
Like many analytics leaders today, Bogart can recall what the early days of “data science” felt like, reflecting on the sense of community that exists today contrasted with what it was like in the past.
“[Back then] analytics was a discipline of engineering. So, statisticians were still called statisticians. The idea that we [different industries] are all trying to solve different problems in the same ways with a common toolset and/or philosophy didn’t really exist yet.”
We ask him what’s changed and he lists a few technological leaps. The ease at which he recalls them underscores just how important these two factors have become: the rise of open source and accessible cloud computing.
“Originally, all of the computation power came from your local machine. So, often times your ambitions were greater than your ability to solve them. You ran into a lot of situations where you might have a solution to something but without any way to scale it. ‘As long as the work comes back to me, I can always provide you the output.’ But the idea of shovelling that raw data into a system to do the work for you was sort of unthinkable.”
‘Unthinkable’ would have been a common descriptor for many of the capabilities machine learning powers today. Whether or not the broader promise of AI will live up to the vision it espouses is an ongoing debate.
P&G and the proto-Analytics community
Bogart’s tenure with P&G overlaps with the early days of data science hype. He describes the beginnings of analytics centralization at P&G as a ‘mixing of existing personalities’.
“About a year into my work there, they decided to group analysts into a shared services group. So, this was marketing analysts plus supply chain folks plus retail analytics — all these different personalities now starting to share a single identity. This eventually turned into the embedded analytics professionals P&G has now.”
He suggested that P&G may not have created ‘data scientist’ roles immediately as they had established identities as analysts.
Operations research and the perfect coffee bean
Bogart mentioned an early use case in Operations Research at P&G that reminds us of the optimization problems studied in current analytics hybrid programs like the Queen’s MMA.
“I remember one of the first major success stories back in the 1980s was a blending model for coffee. The idea was fairly simple. Different bean types have different flavour attributes and a brand — like Folgers — has a specific flavour profile. So, the model would look up the daily price of individual beans and optimize the cost to deliver a certain profile.
This ended up becoming one of the first Operations Research (OR) applications at P&G. So, by the time I joined (almost 20 years later) the OR team was already a well established group having evolved from projects like this.”
At the time, this would’ve been a monumental achievement requiring coordination from multiple teams and substantial IT costs. Today, this optimization problem sounds like something that could be achieved in a single python notebook.
Scalable data science
“It feels cliche to say it but investing in data science right now is really about shortening the gap between insight and action. It’s building solutions to embed good decision making into processes.”
Bogart illustrates why that shortening is so crucial in retail.
“You can produce great analysis that says: we have some SKUs with too much inventory. But now you have to go the inventory planner and present that work. Maybe she doesn’t believe you and maybe she’s right, maybe you missed something. The issue isn’t about who’s right or wrong — it’s the delay because you weren’t on the same page somehow.”
Retail has many broken telephone-esque instances like this due to the siloed nature of how retailers operate. To Bogart, the solution is a question of system design and agile execution.
“You need to get everyone to agree to the premises of the analysis and the underlying philosophy behind what we’re doing. Once you have that alignment, then you can build the pipeline to support the MVP of the system that automates the result or insight you were originally digging for, and then — very importantly — iterate.”
This methodology is what Bogart’s team strives to implement in their own work. Considering the scale of what Acosta does with its store reps all over North America, strong consideration for scalability not only makes sense but is necessary for success.
The gamefied store rep
“We have thousands of reps calling on stores all over the country. It’s a workforce with a good mix of part-time and full-time workers. Their expertise is to find problems in the stores and fix them.”
Supporting this work force with traditional analytics outputs like reports and dashboards doesn’t really make sense. So what is the product that empowers an individual rep to address the right problems?
Bogart describes a system that generates a curated list of tasks that are informed by data and can be tackled by individual reps.
“Consider a rep that has a limited amount of time to spend in a given store. They might not be there every day, maybe it’s a weekly visit, maybe there are issues with different item categories. How do you help them optimize their time and address the right issues? How do you empower them with insights from our latest prescriptive analytics that will help them make better decisions quickly?”
Bogart never explicitly mentions incentive systems or details on how a system like this would be productionized (perhaps an app), but the underlying idea of decision automation / support for human labour is a compelling example of scalable analytics and automation with human oversight firmly in the centre.
“If I give a list of 100 things for a person to do everyday, that’s probably not efficient. But if we look at the average time to completion for every individual task, crossed with each rep’s own historical productivity at different times in different places, can we curate a list that’s more optimized but manageable? More importantly, can we drive a sense of satisfaction for own employees?”
The key idea:
Where AI will be most effective in retail
We shared an article with Eric that’s made the rounds in the analytics community to get his take on it. He laughs, noting that he’d seen it earlier today — his boss had forwarded it to him.
“There’s a few things that drive those recommendations [from the article]. The reality in retail is that there’s a lot of effort that goes into repeated processes with a lot of rules-based decisions being made along the way.”
Bogart suggests that AI’s most valuable trait for retail is the flexibility it can create within existing systems and the potential to replace the current rigid, rules-based systems with AI-driven ones. A less sales-pitchy way of saying it is: analytically automating the simpler decision points to the best of your abilities (or within your acceptable error tolerance) can help to expose the parts of the process where human intervention is most valuable.
“All these things that are AI-friendly but are currently being done by humans is a big part of why the valuation of AI in retail is so high.”
The second key trait of AI that Bogart believes can be better exploited are feedback cycles.
“Ultimately, this is the higher promise of AI right now, right? The idea that any sort of model(s) that is put into production somewhere can always be tweaked, revisited or retrained provides some amount of scalability into the future.”
These are the bigger picture ideas. There is, of course, mention of the potential of robotics and vision systems that will certainly be disruptive in the future. Bogart acknowledges this as well.
“ […] these are the things outside of the more nuanced side of retail; things as simple as counting out-of-stocks, for example.”
The more nuanced side of retail that Bogart alludes to is the world of marketing, forecasting and branding. Broadly speaking, the consumer facing side of the industry.
“This side of the business deals with things like: what is the effect of adding this new item to a category given its own set of characteristics? Yes, there are models for this already that could be augmented by AI. But ultimately, this is the side of the business that is core to brand and identity. ‘Who are we and how do we attract and retain shoppers?’ This is not a question that can be directly answered by throwing AI at it.”
The search for in-store data
We then chatted at length about the challenge in capturing in-store data, but the sense we got after circling around cool use cases and potential solutions is that, there is really no clear and obvious winning approach yet. A shared sentiment that was echoed in another interview we had with Peter Cuthbert (Director of Customer Loyalty @ Loblaw — article coming soon).
“Lots of solutions right now rely on bluetooth or WiFi. There’s also video capture and proximity sensing, but these are things that don’t really address the who. It’s more about tracking the ‘blob’. How long do people stay in front of a display? What parts of the store get the most traffic?”
Bogart concedes that there is a lot of promise in solutions like these but the right convergence of methods hasn’t been found yet. He identifies one reason why finding concrete value is difficult.
“Scaling anything is hard. Anything that goes from a pilot to multi-store wide execution becomes magnitudes more difficult.”
If scaling is out of the question, proving value needs to be done with an MVP. But to Bogart, the tradeoff needs to come at the cost of trying to answer a harder question.
“Driving convergence between behavioural and sales analytics is one of the biggest frontiers. Mapping an individual shoppers journey through a store with some granular level of detail and linking it to the their final basket is huge. Being able to classify a shopper’s mission when they walk into a store, these are the harder, much more valuable problems to solve.”
If we consider a point-of-sale dataset, a small basket could be associated with two different consumers. One could be a simple stock up trip and the other could be a dissatisfied shopper who left without buying everything they wanted. The difference between these two outcomes is that one is a valid mission and the other isn’t. In-store data is the key to separating the threads here.
So when it comes to in-store data, it seems the endgame outcome is clear. The difficulty comes from not having a roadmap of the best intermediate solutions along the way. Slowly but surely, it seems like retailers are willing to try things out.
What retailers need to do to win online — frictionless
“Wal-Mart.com might not be the best online shopping experience but it’s not too dissimilar from what shopping at an actual Wal-Mart is like. It’s not overly premium online and sometimes the aisles can get a bit confusing, but online and in-person don’t feel like two different stores.”
Bogart is quick to point out that building a carbon copy of your in-person experience online is not the point.There are retailers with vastly different online and in-store experiences and are majorly successful.
Sephora, for example, adopted a showroom model wherein the retail experience serves a fundamentally different mission — the experience of testing and trying products. It continues to succeed because, while the retail experience isn’t primarily built to facilitate the final purchase, the overall experience of Sephora never gets in the way of it. In fact, reps in Sephora stores can easily facilitate online purchases and are encouraged to do so.
When assessing how to execute eCommerce, Bogart offers the following, “Retailers need to understand why that shopper is in their store and make sure to always support that mission. Then they have to figure out ‘What is the consumer’s appetite to shop online and whether competitors are offering similar experiences that are threatening?’
If the answer is ‘yes’, they need to ask themselves, ‘Does my brand extend itself in the same way?”
Lastly, Bogart makes a distinction about where AI is most effective in this domain.