Five pitfalls to avoid when deploying AI in retail

Five pitfalls to avoid when deploying AI in retail

By Valter Andersson, Technical Solutions Manager, Nosto

Everyone is talking about how artificial intelligence (AI) will transform retail, from chatbots and robot-driven warehouses to increasing personalisation. But it is by no means a magic bullet and there are a variety of decisions that have to be carefully considered. 

First it’s important to appreciate that when people discuss AI in retail, they’re normally referring to machine learning - a subset of AI in which algorithms are trained to learn and get better at performing specific tasks based on the patterns they identify in data. In our experience there are five common pitfalls that stop this kind of AI delivering fully on its promise:

1            Deploying AI in the wrong areas

Like any technology, you need to focus on achieving concrete business goals, such as increasing revenues or margins. Not being clear enough on this, or directing AI at the wrong targets, could leave you disappointed. 

Begin with the areas where you are likely to drive the biggest impact, while minimising effort and disruption. Once you see returns there, move on to more marginal applications. If you’re looking to use AI to improve aspects of the online user experience, for example, start with the home or checkout page, where any improvements can drive the greatest potential benefit. Then move on to less visible or active parts of your site.

2            Having unrealistic expectations

AI isn’t an instant solution. Machine learning algorithms take time to learn about your customers or different aspects of your business. The results will get better with exposure to more data. So ask your AI supplier to be clear and realistic on the timescales before you’re likely to start seeing positive results.

Equally, rather than getting caught up in the hype, be realistic about whether AI is truly the best option for realising a particular goal – or whether a simpler solution might work better. In ecommerce, rather than using machine learning to analyse and target individual customers based on complex predictions of customer lifetime value, you might actually get a bigger revenue uplift by offering a straight-forward discount or offer whenever a visitor views a product page a set number of times.  

3            Removing the human touch

AI doesn’t mean the end of human involvement – it needs to be guided by people to ensure it’s applied correctly and trained on the right data to reach your goals. 

For example, no amount of chatbots or machine learning can drive up online conversions from shoppers coming to your site, if a large proportion of them are not interested in your products or aren’t ultimately in your target market. It takes human retail experts to recognise from the data that you’re attracting the wrong or unengaged visitors to the site.     

4            Not using enough or all the data you could

Obviously data is key for machine learning – so make sure it’s comprehensive, up-to-date and high quality. For example, when gathering customer data, don’t just focus on transactional data relating to customer purchase histories - which typically accounts for less than 2% of online shopper data. Make sure you are creating a complete picture by also capturing and analysing the remaining 98%, which is normally made up of people’s behavioural signals.

Similarly, make sure your AI platform can break down data silos between different applications and departments. You should, for example, be able to bring together data from CRM, email tools, Facebook marketing and search as well as integrating data from the web business, instore and warehousing systems. If valuable data is shut away in separate tools it is an opportunity cost.

5            Getting too caught up on ‘deep learning’ 

There’s a fascination around deep learning which uses neural networks to imitate the workings of neurons in the brain, processing large volumes of data layered on top of each other. But in ecommerce there are times when you need to make smart decisions with whatever limited data is available. An online customer interaction may be restricted to just a few clicks for example, or a product may be in stock for a very short time, meaning you simply don’t have those multi-layered signals. But you can still use ‘shallow learning’ to make intelligent decisions based on the limited data you do have access to.

As with any new technology, you need to make sure that you have built a strong business case for AI that is based on the metrics that matter to your company, and work with a supplier that not only understands the technology, but also has deep experience of retail. With this approach and by being highly rigourous about how and where you direct the technology, you’ll be less likely to be disappointed.

When is Apple going to take payments seriously?

When is Apple going to take payments seriously?

Removing friction from retail: four technologies retailers should consider

Removing friction from retail: four technologies retailers should consider