We Set Out to Automate Customer Support in 2018. Here’s How Far We’ve Come
)
Where it started: our 2018 AI plan
We launched our first AI capabilities in eDesk back in 2018, when I was CTO. At the time we were developing our own machine learning models, using the customer conversations and eCommerce data flowing through the platform to understand incoming queries and help automate responses.
I came across our original AI R&D plan from 2018 recently. One line stood out:
“What we wish to do, using AI, is to provide either suggested automated responses or to actually respond, removing the requirement for human interaction with the customer exchange.”
It’s not the most elegantly written sentence, but it sums up what we were trying to achieve eight years ago. We were working towards automating customer support at a much greater scale.
The document goes into quite a bit of technical detail on the different approaches we were taking, including Naive Bayes classifiers and neural networks. Back then we had to build and train much of this ourselves. We had years of customer conversations to work with, which was a big advantage, but preparing that data and getting the models accurate enough was difficult. The computational costs were also a major consideration, particularly when training models for individual customers or different languages.
How LLMs changed what’s possible
Modern LLMs have changed what’s possible. We now have access to models that would have been extremely difficult and expensive to develop ourselves in 2018, and that’s allowed us to accelerate our development considerably. There’s still a lot of engineering involved though, particularly when you’re building AI specifically for eCommerce.
Built around the eCommerce transaction
We’ve always built eDesk around the eCommerce transaction. A customer message on its own only tells you so much. If someone contacts a seller about a damaged product, the order information and the seller’s policies are just as important as the message itself. And if the seller operates across multiple marketplaces, the correct process can vary depending on where the sale took place.
We’ve spent years developing integrations with hundreds of sales channels and bringing that commerce information directly into eDesk. Our Procedures technology allows sellers to control how their AI Agent handles different situations, including the actions it’s permitted to take. It’s a big part of why our AI Agent can actually resolve customer issues rather than simply respond to them.
Real results for sellers
Across eDesk, we’ve now powered tens of millions of AI interactions. Our customers have been using AI to automate their support for years, and we’ve seen the commercial impact first-hand.
Trade Chem is a good example. They’re a UK retailer processing around 3,000 orders a day across more than 20 sales channels, with a support team of four people. They’ve avoided the need to hire another full-time member of staff by using eDesk AI Agent, saving an estimated £35,000 a year. They’ve achieved that using a relatively small number of procedures, mainly around common queries such as order tracking and damaged parcels.
Introducing AI Agent 3.0
We’ve recently released AI Agent 3.0, and it’s our biggest technical improvement in quite some time. We’ve changed how the Agent retrieves information and works through more complicated customer queries. We’re now using different AI models for different stages of the process, which has allowed us to improve accuracy and the range of issues it can handle.
One of the challenges with this technology is that you can have an extremely capable underlying model, but if it retrieves the wrong order or applies the wrong policy, the result will still be wrong. We’ve put a lot of work into those areas with 3.0, including how the Agent handles conversations that change direction halfway through, or situations where the customer provides images or documents that need to be examined.
In testing across more than 11,000 real customer conversations, 3.0 resolved 45% more conversations than the previous version. It’s a substantial improvement on technology that was already handling millions of customer interactions.
Ready for Black Friday and Christmas
We’re heading into Black Friday and Christmas now, which is an important time of year for our customers. More orders mean more customer queries, and for businesses selling across multiple channels the workload can increase very quickly. We handle more than three million customer conversations a month through eDesk, so we’ve seen this pattern for years. It’s also when the ability to resolve more complicated queries automatically becomes particularly valuable, especially when support teams are already stretched.
Customers don’t particularly care whether it’s Black Friday or the busiest week of the year for a retailer. They still expect their orders to arrive on time, their questions answered and any problems resolved quickly. And rightly so.
What’s next
For me, the benefit of AI in customer support has always been straightforward. Sellers should be able to grow their business without having to keep adding people to deal with customer queries. We’ve already demonstrated that this is possible, and 3.0 allows us to automate more of the complicated work that previously needed a person to get involved.
The next challenge is getting AI to handle even more complex issues reliably, across the different channels and systems eCommerce sellers use. That’s where we’re putting our effort, and we’re already working towards our next major release early next year.
)