Despite the endless headlines of all the ways companies are exploring, investing in, or are concerned about generative AI, there are few stories about companies doing something with the novel technology.
That’s not to say real projects aren’t happening; New Zealand-headquartered and Canada-connected Xero is a prime example. The small business accounting software company boasts nearly four million users globally and recently launched a gen AI solution for its Xero Central self-service helpdesk, built in partnership with Québec-based AI software platform Coveo.
Speaking with BetaKit, Nigel Piper, executive general manager at Xero, explained how the project came to be and what it took to build and implement a solution.
Identifying problem-solution fit
When a new small business owner signs up to Xero, Piper said the organization’s key goal is getting that person set up and to a value moment as quickly as possible. This north star, said Piper, came from years of customer analytics.
“We had invested in a strategy of content-led support for a number of years. And so I think new organizations would have to work out: do they have the content?”Nigel Piper,
“In that formative stage of their first few weeks or first few months at Xero, it’s very much the “how do I?” type questions,” said Piper. “And so, for us, it’s all about getting people set up well, trying to reduce the effort that they have to get set up.”
For years, Xero tackled this problem with a library of how-to content on the company’s Xero Central helpdesk site. There, users could search for an answer or follow segmented guides like getting started as a business owner or accountant. The company also collected data when a piece of content didn’t work and someone needed to submit a support ticket—that insight was used to continually update the library, editing or adding articles as needed.
“We have a digital record of the questions that customers have asked Xero for quite a number of years,” said Piper. “We have a lot of probability history and data of where customers may get stuck, the type of questions that we would ask.”
Building a gen AI solution
As Xero continued to grow over the years, Piper said the team added significant amounts of content across Xero Central, as well as the company’s blog, marketing pages, downloadable ebooks, etc. This meant discovery was leading to an increase in support tickets: if a user couldn’t quickly find the content they needed, they were more likely to need additional support.
Helping users discover answers was the first prompt for Xero to consider using AI as a solution, and the company partnered with Coveo on a discovery tool for the Xero Central platform. With this AI-powered capability, support articles would surface as a customer typed their question into the search bar on Xero Central. This saved users a lot of time, as the AI could often guess intent based on previously addressed queries and auto-fill an answer. If not, it served as a custom search engine through Xero’s vast content library, doing much of the heavy lifting to save customers time.
This approach had its limits, however: discovery was only helpful for straightforward problems with an article already written providing the solution. But sometimes a user might need to consult a variety of content to get the complete solution to a more complex issue.
This issue kickstarted the need for a new gen AI solution: a platform that can understand a customer’s search intent, search all available Xero articles on the subject from any source, parse which sections of each article are relevant to a customer’s specific problem, and generate a net-new, custom answer detailing the exact steps someone needs to take to solve their problem. The solution would also surface the articles that informed the gen AI response, so users could source and read the entire guide if they were curious.
“In our context, we could take all of the places that we had content and produce that in a list form to make it pretty easy for a customer to see the steps that they needed to take to do a particular action,” said Piper.
The team worked with Coveo to leverage their Relevance Generative Answering capability, with Nigel noting that it was a fairly easy process due to both the existing partnership and Xero’s significant content assets built up over the years. This content became the infrastructure behind Xero’s AI implementation for search and discovery, with gen AI used to develop new, customer-specific content.
“The unique opportunity that Xero had is… we had invested in a strategy of content-led support for a number of years,” said Piper. “And so I think new organizations would have to work out: do they have the content?”
While Coveo CEO and chairman Louis Têtu jokingly refers to this kind of content-based infrastructure as the “boring plumbing” of artificial intelligence, it’s an absolutely critical component to a successful gen AI implementation that cannot be ignored.
“The science and the challenge is in the data infrastructure, the security, and the unified index and relevance as well as in the prompt engineering,” Têtu said. “It is not in the LLM.”
Test, implement, and iterate
Before launching the solution, Piper said a key challenge was balancing “speed versus accuracy” to ensure customers got the right response. To accomplish this, Piper said testing included internal teams at both Xero and Coveo running through prompts, validating and training the AI models “to make sure that the answers we gave were accurate and delivered value.”
While the program is in its early days, a recently published Xero investor deck suggests things are going well: customer search time is down 40 percent and search sessions requiring additional support are down 20 percent.
Piper added that the support team also tracks traditional metrics like customer satisfaction (CSAT) and net promoter (NPS), both of which have stayed consistent through the experiment so far—suggesting that people are actually finding the answers they want rather than simply trying the gen AI tool then giving up.
Looking ahead, Piper said the name of the game is continued improvement for Xero customers. It’s also the advice he shares to other leaders considering gen AI solutions: get clear on what you’re hoping it will do for your organization and for your customers, since that will guide how you build and implement a solution.
“In this world, you can do some amazing things for customers if you get your technology and your strategy right,” said Piper. “I read lots of articles around the bad side of AI and how it could, perhaps in lots of countries and organizations, be used probably for the wrong reasons. I’m pretty clear that the right purpose of technology is to aid and help customers, never to replace the value that you can give them.”
Go here to learn more about how Xero’s gen AI project with Coveo came together and grew case deflection by taking customer support from reactive to proactive.