Four Considerations For Implementing AI In Ecommerce & Retail

Matt Payne
November 10, 2021

The landscape for using AI in ecommerce and the retail industry has changed a lot recently. Some of the most popular products and approaches have been compromised or undermined in a very short time by a new global impetus for privacy reform, and by the way that the COVID-19 pandemic has transformed the nature of retail.

These events have thrown a huge spoke in the long-term plans of both the AI in ecommerce industry and the forecasting sector alike. The previous surge of interest in facial recognition systems for retail has given way to a focus on improving the online shopping experience, leveraging AI-driven predictive analytics, and improving the management and configuration of bricks-and-mortar stores.

In the meantime, the continuity of market forecasting models has been catastrophically disrupted by a transformed business environment. With a massive increase in ecommerce sales under COVID-19, there's no doubt that huge growth in the use of artificial intelligence for ecommerce companies and online retailers is coming – but it may not be in the sectors that were anticipated just 12 months ago.

So what can we depend upon, going forward? Here are four baseline approaches to AI in ecommerce that are likely to prove resistant to the current volatile market.

1: Using AI To Generate Relevant Customer Recommendations

A better understanding of product relationships and customer intent can transform turnover and conversion, and a really effective AI-driven inventory search system will stand out in a sector that's developed a bad reputation in recent years.

For instance, it's a common complaint that if you buy — for instance — a TV, the site you bought it from might follow up with product recommendations for non-sequitur purchases such as:

  • Another TV you could buy (on the basis that Customer likes TVs = recommend another TV!).

  • A six-pack of oatmeal cookies (on the basis of 3.7% of TV buyers in my CRM database also bought oatmeal cookies at some point + this is the optimum hour to push grocery sales).
  • An air conditioning system (same category [electronics] as TV + same price as a TV + it's summer = recommend air conditioning system!).

In an ideal world, you're going to need better search logic (and possibly better data) than that.

In Search Of Meaningful Data/Event Relationships

As the above examples demonstrate, the key to a successful recommender system lies not in merely finding relationships between data points and events, but in identifying associations, hidden in different strands of the data, that are both meaningful and incentivizing for the customer.

Chinese researchers have developed a next-item recommendation system that accounts for both known user behavior and current session behavior.

The process of developing an effective recommender system with ai technology requires a lot of initial curation, an open mind, and a willingness to experiment. A/B testing can help establish recommended weights, while offline A/B testing is another useful tool for developing optimal parameters and settings for an ecommerce product recommender algorithm.

Dealing With Outliers

If a single customer buys a TV and a box of chocolates in the same purchase, most supervised learning approaches will dismiss the event as an 'outlier' anomaly. Perhaps the customer was, by coincidence, buying a gift at the same time as a typical domestic fixture. Perhaps they were feeling a bit down. Who knows?

If two or more apparently unrelated customers were to place the same two items — or item types — in an order, a typical artificial intelligence framework might identify an implicit association between TV/Box Of Chocolates, and later present either as an after-market recommendation to customers who purchased the 'correlated' item.

The Sweet Spot

But the system might fail to notice that the original orders were placed in the second week of February — the second-highest seasonal driver for chocolate sales.

Will the algorithm continue to associate the items throughout the summer and fall? Or will it recognize that the association is specious outside of the Easter and Valentine's Day surges, and that there is no intrinsic 'relationship' between the two items at all?

As we can see, some of the relevant data and insights necessary for an effective recommendation system will not be present in a typical customer relationship management (CRM) database, or in conventional sales analytics systems.

It's here that the deployment of machine learning analysis can really help in filtering out the noise of coincidence from the discovery of useful and exploitable data relationships.

The Advantages Of Developing In-House CRM

  • Customer Churn (Customer Attention)
  • Analyzing historical data can reveal the usage and communications patterns that lead to a customer abandoning a brand. Factors might include excessive subscription marketing, a problematic change in the user experience, or poorly-designed conversion funnels. Sectors addressing churn with AI include banking,, telecoms, and retail.

  • Customer Segmentation
  • Identifying high value customers enables a business to concentrate resources on the best prospects. Amongst other artificial intelligence approaches, recurrent neural networks (RNNs) are proving more valuable than traditional Pareto analysis in this respect.

  • Additionally, the use of unsupervised machine learning techniques allows the discovery of new customer patterns and insights in a way that can't occur when 'unexpected' insights are dismissed because they don't conform to a traditional analytical model,.

  • Customer Lifetime Value (CLV)
  • Through machine learning analysis it's possible to predict the long-term value of new customers by studying historical transaction data from existing customers. AI analytics can also help to estimate the current value of clients that already have some history with your business.

  • This field of ML, now becoming commoditized by larger tech players, is used in commercial aviation, retail, and the insurance industry, among others.

An assessment of regular videogame customers vs. 'whales' (high spenders for micro-transactions), in an international study investigating machine learning techniques to estimate customer lifetime value.

Personalized Search Results Through Machine Learning

In cases where the customer is viewing your commercial environment from a dedicated mobile app in a device with client-side neural network capabilities, it's possible to deliver truly customer-specific results, by updating the local search algorithm, and assigning dedicated weights to the customer.

In these circumstances, the algorithm can operate with almost zero latency, even for time-critical functions such as text prediction.

In a less accommodating web environment, customer inference and intent can still be fueled by targeted AI-driven analytics, depending on the server-side resources that are available for updating models for individual clients.

Overcoming The 'Cold Start' CRM Issue

What do you do if you have no historical data yet, or if the customer is so new that a 'generalized' search algorithm might leave them with sub-standard matches?

One US startup grew nearly 400% in a year by using artificial intelligence to analyze its own first-party data for product engagement and platform usage metrics. In this way it was possible to develop a 'fictional' historical sales database that could be improved and refined by real-world transactions.

Research out of Harvard also recently proposed a 'First Impression Model' (FIM) capable of forecasting a customer's underlying traits and lifetime value using only the minimal information from their earliest transactions and interactions. The framework uses a probabilistic machine learning approach to feed forecast accuracy reports about new customers into the improvement of the technique over time.

The different acquisition characteristics of customers with varying CLV, analyzed in a study from Columbia University, NYC.

In a retail setting, insights from the FIM model revealed:

  • The top 10% of heavy spenders are acquired online, rather than through legacy media.

  • Customers that are most susceptible to email marketing will buy less expensive products, and in small amounts.

  • The top 10% of heavy spenders will initially buy high-value discounted products.

Beyond Legacy Analytics

The automation of useful insights like this through artificial intelligence is a recent phenomenon, due in part to the advent of GPU-based AI acceleration since 2010.

However, the analytics and forecasting marketplace remains dominated by legacy systems that apply 'static' theoretical models to real-time data, rather than the new wave of artificial intelligence approaches that are capable of suggesting entirely new sales flow models, architectures and core principles.

Sometimes the data will not bend to the theory, and only machine learning can dynamically address this. Even fairly recent analytics luminaries such as Hadoop and Apache Spark are now ceding ground to TensorFlow, and to the rise in more flexible neural networks.

Reasons To Adopt AI In Ecommerce

This is the ideal time to integrate ai solutions into product search, because:

  • The technologies are mature enough to develop genuinely new and profitable insights into customer behavior.

  • The best machine learning technologies are almost entirely open source.

  • There is wide developer availability across the most important domains, such as NLP.

  • Customer targeting through cross-tracking, central to many legacy analytics systems, is set for a rapid decline (see #4).

2: Implementing AI Chatbots

The advent of the COVID-19 pandemic greatly accelerated, uptake of Ai-enabled chatbots across all industries and in the public sector.  Consequently the chatbot sector is currently estimated to rise from its current value of $2.6 billion USD to $9.4 billion by 2024, a compound annual growth rate (CAGR) of 29.7%.

Key Sectors For Chatbots

Some of the many sectors that have increased use cases for AI-centered chatbots since the outbreak include:

  • Financial services, with the banking sector in particular deepening its commitment to AI-based agents and services in response to the pandemic.  

  • Call center services.

  • Health information services,.

Chatbots In Action

Examples of innovative chatbots include:

  • The Bank Of America agent known as 'Erica', piloted in 2017, which quickly gathered over a million users after rollout, soaring to 14 million users under the pressure of COVID-19 lockdowns.

  • Food retailer Lidl's highly conversational wine chatbot Margot, which won wide plaudits from the press shortly after launch, offering a natural conversational style in a number of languages.

  • The Whole Foods Facebook messenger bot, which provides personalized recipe recommendations based on user response to filters.

GPT-3 Shows The True Potential Of The Commercial Chatbot

In 2020 Microsoft acquired exclusive licensing rights for OpenAI's revolutionary GPT-3 text-generation system, which had already astonished the tech press, with its insight and capability earlier in the year, and which can operate as a chatbot if required.

It's been argued that GPT-3 finally demonstrates that a generalized knowledge system can conduct useful and informative discourse without needing to be trained with domain-specific datasets, or years of historical customer interactions.  

Will The Chatbot Become Just One Facet Of Wider Business AI Solutions?

Though it isn't yet a perfect conversational partner, GPT-3 does demonstrate the rise in public confidence and expectations for AI discourse agents over the next five years.

Companies and individuals licensing GPT-3 are using it for purposes as diverse as Python code generation, resume building, content generation and chat response automation. This indicates that chatbot deployments may, in the future, represent only one utilization of a central business-focused AI system.

The Future Of Chatbots

Facebook's own research into 'open domain' chatbots such as GPT-3 indicates that the near future will remain dominated by domain-specific AI agents, powered by insights from a company's own customer data.

Therefore the immediate value of chatbot development lies in leveraging historical (proprietary) data to ensure apposite responses in areas such as:

  • Customer support, reducing the burden on limited support staff – an approach that Amazon.com implemented in early 2020.

  • Gathering customer feedback, as IKEA has done with the ORC chatbot.

  • Multi-language customer relations, where integration of advanced third-party translation systems can combine with AI-focused analysis of previous (foreign language) interactions to improve dialogue with clients outside of the organization's native language.

  • Guiding users through a buying process, like Marriott International's ChatBotlr agent.
The architecture used by Amazon to evaluate response rankings in its chatbot interactions.

3: Making Use Of Image Recognition

The shopping experience is becoming noticeably more driven by data and ai enabled algorithms, with computer vision a crucial element in a new ecommerce & online retailers revolution.

The Rise Of Image Recognition In Retail And Ecommerce

The general global image recognition market is set to rise from $30.28 Billion USD in 2020 to $98.47 Billion by 2027, while the retail sector alone is estimated to rise from $1.4 billion USD to $3.7 billion by 2025.

Additionally, Gartner has forecast that by 2022 70% of retail interactions will involve ai solutions and their ancillary technologies.

Stock Photos

In the current 'masked' age, the locus of AI research and investment has shifted from in-store facial recognition to image-based analysis of stock items in a retail context:

  • Two years after collaborating with SnapChat to equate mobile users' real world images to commercially available products, Amazon launched AI-driven service StyleSnap (see image below), which allows users to upload a photo and find clothing items for sale that are similar to the clothing identified in the picture.

  • Retail shelf monitoring through image recognition is emerging as a solution to a number of challenges, including automated inventory management, with a 20% improvement in accuracy over manual stocktaking, according to one study.

  • The use of computer vision to compare product availability and positioning against sales records for the same items can help develop better store planograms,.

  • Warehouse automation can be facilitated through artificial intelligence-driven systems whose recognition models are constantly updated, so that changes in label templates and other periodic anomalies don't break the entire system.
Amazon uses image recognition and style transfer to identify clothing inventory similar to a customer's uploaded image.

Other Use Cases For Image Recognition Ai In Ecommerce & Retail

Image recognition can also help in evaluating how paid influencer campaigns are disseminating images of your product (through image recognition), and in gauging the type of responses that the product is creating (through NLP and associated techniques).

Though the need for customers to wear masks has impacted the usefulness of facial recognition in retail, computer vision can be deployed in various ways in a store without requiring facial identification:

  • Visual search for ecommerce allows potential customers to find products they want quickly through just a picture on their phone.

  • A company in the US deployed an ai powered image recognition system in a store car park. When any abnormal situation is recognized by the AI (from violent behavior through to injury and signaling for help), the system alerts staff.

  • Amazon's cashier/register-free shops use image recognition and IoT technologies to allow customers to bag items without any check-out process, with all orders finalized automatically upon leaving the store.
AI-powered vigilance against suspicious behavior in stores.

4: Building And Growing A Proprietary Data Domain

You might be looking forward to running your historical customer data or site traffic statistics through a neural network in order to:

  • Fill your 'similar items' sidebar with uncannily apt personalized suggestions, none of which are generated from your own data about the client.

  • Haunt a non-purchasing site visitor with an AI-driven ad that seems to have read their mind or overheard a recent conversation of theirs, and which follows them across the ads sections of multiple unrelated websites.

  • Follow up on customer website activity with individually tailored AI-generated email newsletter content, triggered by customer events that you're monitoring through third parties.

However, nearly all of these activities are made possible by tracking methods that are increasingly coming up for legislative review around the world, or else being suppressed by the makers of the most popular web-browsers and manufacturers of mobile devices.

Network map of a website using a supercookie for cross-domain analysis and data-gathering.

The Demise Of Cross-Domain Data

The major technology in decline is cross-domain tracking. The past five years of scandals around AI-enabled political social engineering have brought the use of supercookies and other 'omniscient' web tracking techniques far enough into the public consciousness to effect real change:

  • In late January of 2021 it was announced that the next version of the Firefox web browser would 'silo' local storage for each domain visited, making so-called 'supercookies' difficult or impossible to implement.

  • Google has promised to phase out third-party cookies in Chrome by 2022, cutting off cross-tracking solutions for millions of advertisers eager to exploit the browser's 66% market share.

  • In early 2021 Facebook ramped up its angry response to Apple making targeted advertising opt-in instead of opt-out in its mobile devices.

  • It's believed that Google's Android OS will eventually be forced to follow Apple's lead.

  • Facebook-owned WhatsApp backtracked, on an attempted cross-domain land-grab after millions of users migrated to more privacy-conscious alternatives.

Preparing For A Privacy-Focused Data Environment

The message here is that the 'wild west' era of populating ecommerce datasets with cross-tracked data is coming to an end,.

In this transitional period, it's unclear whether the new generation of privacy-focused frameworks — such as Google's Privacy Sandbox initiative, LiveRamp's Authenticated Traffic Solution, and offerings from Neustar and Project Rearc, — may eventually fall foul of later privacy initiatives.

Therefore it's prudent to:

  • Integrate new AI-driven analytics systems on an SaaS basis while the political and social arena plays the topic out, in order to limit your business model's exposure to legislative change.

  • Avoid accruing technical debt with dedicated engineering solutions that make sweeping presumptions about what will be an 'acceptable' data/privacy trade-off in 2-5 years.

  • And, most importantly, to develop unassailable first-party AI-driven customer insight environments.


The restrictions that come with a growing emphasis on privacy, and with the privations of the pandemic, are also unique opportunities for using artificial intelligence in ecommerce and online retail space to continue to grow.

The massive growth of consumer data under coronavirus restrictions has led to a corresponding boom in the Big Data Analytics (BDA) sector, leading in turn to an unexpected acceleration in machine-driven analytics that most believe is here to stay.

In the meantime, the prospect of well-defined privacy laws promises, at last, a stable legal environment and a relatively level playing field for ecommerce analytics and artificial intelligence systems, furnishing the business confidence necessary to invest in a new generation of compliant and performant AI-driven analysis frameworks.

Article Sources

Interested in seeing what Ai models can do for your ecommerce business? Want to start collecting data that gives you ROI your competition would beg to have? Let's talk about how we can use Ai to increase conversions and find potential high value customers hiding in plain site.

Contact us to learn more

Don't let your niche competitors find these Ai models first.