Across business, the question of whether to build or buy new IT solutions has traditionally focused on such factors as availability, costs, time to value, and maintenance, to name a few. However, with a new generation of AI-powered tools now available, there’s another factor to consider: can the required solution be built in-house?
For CarParts.com, which sells after-market car parts and accessories online, the complexity of its business model meant many off-the-shelf solutions simply didn’t fit the needs of the business. When the time came to renew a contract on its chatbot, the decision was taken to bring the capabilities in-house.
“Buying enterprise software is like buying cable TV,” says Aurellia Pollet, director of CX for CarParts.com. “You only want two channels, but you can only buy a bundle with sports. This is kind of similar. There were things we wanted to add – like agents for the chatbot and a ticketing system – but we have a lot of expertise inhouse and, financially, it wasn’t making sense anymore.”
Pollet adds: “We approached it as a focused pilot. If it didn’t meet our standards, we were prepared to reassess. But it exceeded expectations!”
Levelling up with AI
CX partnered with data science, technology, and marketing to build the solution, leveraging enterprise-grade AI models and cloud infrastructure integrated through APIs to fit our existing systems. By involving these teams, it was possible to avoid the bottleneck that so often forms around the development team by shifting the workload away from a dev and toward data science.
Pollet says: “We’re not a traditional software company, but with the right data science expertise and AI tools, we were able to build what we needed.”
By developing the capabilities in-house Pollet says the business had an opportunity to “rethink how work happens and improve inefficient processes”. It also allowed the business to move with agility and build exactly what it needed – no more, no less.
“No features we’d use once a year. No workflows designed for someone else’s business,” she says.
When developing the chatbot – nicknamed Spark – the teams had to work through dozens of possible service journeys to decide which could be automated and which would need to be routed to a human agent. “We started simple with tracking, then added order cancellations and address changes,” she explains.
When deciding where and when to route to a human agent, Pollet and the team created workflows for numerous scenarios, for example, when a customer queries a payment, the rare occasion when there is an issue with the system, or something is amiss a checkout. “You don't want to leave AI to handle those contacts, because you have to dig, investigate, and ask people to look into this,” Pollet explains.
“The system is proactive. For example, sometimes USPS flags a customer’s address and in these cases an email is sent to the customer to check the address is correct. All of this was done by a human before, but now we have automated. That’s how we looked at it: what is the information that people need to have access 24-7.”
Alongside the chatbot, CarParts.com also needed a ticketing system and catalog, both of which were built in-house.
“Because of the complexity of our business, it was really hard to buy solutions like a catalog, for example. So, we built our own. Before AI, every time we wanted to buy new software to do something, we had to consider how it would plug it to all our custom systems. Finally, this has been leveled up with AI and we don't have as many constraints anymore. It’s so freeing,” says Pollet.
In a world where everybody is talking about how they’re using AI, Pollet says it’s easy to feel like you’re behind the curve. “And we were trying to catch up to be like the big players. For once, we’re ahead of where we thought we’d be,” she says.
“We still have a lot of work to do, yes. But in three months, we were able to build Spark and a ticketing system that has three AI agents and can answer customers directly. They currently handle a meaningful portion of our ticket volume independently,” she adds.
Nurturing the AI culture
When debating the buy vs build question, many focus on in-house technical capabilities, but as Pollet explains, there are cultural capabilities to consider, too. After all, three-month internal tech builds are not for the faint hearted.
“I would not recommend a small company to go into a venture like this,” Pollet says.
“First of all, you need a team of data specialists. I have three data scientists working full-time on this project. You cannot teach yourself these skills on your computer. Also, sometimes an off-the-shelf solution will work perfectly for if business is not complicated. Maybe you sell t-shirts, for example. You may have a lot of SKUs, maybe lots of different colors and sizes. But for us, a single model of car can have multiple types of parts that differ by the year the car was made,” she explains.
Businesses also need to assign people to be responsible for AI as its own entity.
“Someone has to manage the platform, you have to manage the training, the tool itself, and so we're seeing that we need new roles that don't exist today,” Pollet explains.
“You cannot just say it's on and then go do something else because AI has a mind of its own and requires active monitoring and governance,” she adds, predicting this is one of many reasons she believes AI will create multiple new types of jobs in the coming years.
However, even a business like CarParts.com – which is highly complex but staffed with a team of data scientists – there are certain tech build projects that are off limits. “I wouldn't try to develop my own website structure in-house and we're not developing our own payment processor,” Pollet adds.
Lessons in process optimization
The builds didn’t just give CarParts.com a new set of tools to use – through necessity, they also afforded the business the time and space needed to get under the hood of its own processes and improve the things that needed attention.
“With AI, you have no choice,” Pollet says. “You have to go back and look at all your processes again. Because broken processes might work for a human because they find workaround. But if you put AI on top of a broken process, it’s like, you told me it's pink. I see blue. I'm stuck.”
Calling it the “biggest learning and maybe the thing that I'm most grateful about”,
Pollet says: “It was like a mirror that we held up on things that need to be improved on the back end so the front end can actually work. And we continue to do that.”
Among the many learnings Pollet reports, is the realization that with AI “you can’t just copy paste something”, not even LLM prompts. It’s just part of why selecting a suitable LLM was a learning curve.
“It was more trial and error than a straightforward selection,” Pollet says, recounting how the teams have to test multiple LLMs, each requiring new prompts and programming. Some failed on technicalities, others were found to be too inflexible or too expensive. But even with Meta’s solution selected, it is not a set and forget tool, instead calling for frequent modifications and maintenance to ensure optimal performance.
The next steps
While it could be construed as such, Pollet is quick to clarify that this “isn’t an anti-SaaS message. It’s a pro-CX courage one.”
“Sometimes with a contract, it's like putting a square peg in a round hole and it isn’t going to fit,” she says.
By having the courage to build their own solutions, CarParts.com has freedom to continually customize and refine, and add to the capabilities as and when needed.
The company is currently busy ensuring it is ready for agentic commerce, but there’s also plenty on the horizon in terms of the next CX and service builds. Pollet says there is still work to do on Spark – “we can already do customer service questions, soon we'll do sales too” – and in future SMS will also be added, as well as voice AI. “If you're looking for your tracking information and you would rather have a call than doing it on your phone or your computer, totally fine. You don't have to wait. You can call at midnight if you want to. So that's where we're going,” Pollet says.
Quick links
- How Lidl built a 92% accurate AI model to transform customer service
- Generative AI used by more than one third of people across OECD
- Michael Kors applies AI to improve KPIs