By using AI to remove friction from customer support, the company created more capacity to tackle complex issues

Keeping a support desk running at full speed can be a challenge for any business. One of our businesses found a way to make theirs run more smoothly.

Guy Hiddleston is the Head of IT Systems and Compliance at Artifax, a Volaris-owned business that builds venue and event management software for theaters, museums, and other cultural and heritage organizations.

He recently shared his experience with the Volaris community about building a self-hosted support agent from the ground up, driven by the motivation to make the team’s existing knowledge more accessible to its own staff and customers.

Challenges with finding help desk information

Artifax has a relatively small support operation that handles about 2,500 tickets a year. Typically, the function has been handled by one full-time support specialist with help from professional services. But the knowledge that they worked with was scattered across a huge back catalogue of help center articles, ticket histories, and long, unstructured transcripts of professional services and customer success calls.

“Unless you knew what you were looking for and knew what to search, it was kind of challenging to actually access it,” Hiddleston said.

Their customers had a challenge as well. Artifax’s knowledge base wasn’t easily accessible to them, and that created a steady stream of repetitive FAQ tickets. Answering these queries was taking time away for staff to work on genuinely complex cases.

Meanwhile, the company needed a way to pass down knowledge about niche modules to new hires.

Moving away from tickets to a searchable brain

Artifax decided to build an AI agent to solve their challenges. Their build started with extraction. The team pulled their full ticket history and knowledge base archive, had an agent run through each ticket, and summarize it down to what the customer’s problem actually was, how it was communicated, and how it got resolved. Those streamlined summaries, not the raw threads, went into a database.

Artifax then built two separate retrievers — one for support tickets, one for knowledge-base articles — each kept in separate tables.

The first version of the agent lived in an open-source workflow tool where Hiddleston could chat with it directly. It worked, but it wasn’t ready for customers.

Getting the agent ready for customers

Rather than pointing the agent straight at customers, Artifax built in a safeguard. It did this by sending a new ticket to the agent the moment it was created, then posted the agent’s draft response back as a private, internal-only note. Support staff used a draft that was created by the agent to investigate faster, or borrow a proposed response. In some cases, for more straightforward FAQs, support staff could simply copy the draft into a customer-facing reply.

“The reason we didn’t go straight to the public reply is that from a governance point of view, that would be a high-risk activity,” Hiddleston said. “We wanted to be able to validate first, refine the model, and then move forward.”

That validation period ran roughly a month, with humans checking every response for tone, scope, and citations. This stage proved to be important to help prevent hallucinations from the AI agent.

The team changed the agent so that every answer it gives now carries a source at the bottom, so both Artifax and its customers can trace a claim back to where it came from. “We needed customers to build trust to be able to go to the source to validate that information,” he explained.

Building guardrails against hallucinations

Some guardrails that Artifax built into the agent were obvious, such as staying on topic and not discussing anything outside the product.

But other ideas for guardrails surfaced through internal testing.

Since the agent has access to support ticket data, it’s technically capable of answering questions that go beyond support but can contain information that is not meant for that purpose. To prevent misuse, the system prompt now spells out citation rules, explicit prohibitions, and how the agent should refuse and deflect consistently, with every output checked against those guardrails before it’s shown to anyone. And Artifax built in protection for personal data, which got a two-layer defense.

To stress test all of it, Hiddleston challenged anyone within the company to break the guardrails.

Not everything went smoothly. Early on, the team experimented with lower-reasoning, faster models to chase faster response times. Those models hallucinated far more and invented help center URLs instead of citing the real ones.

“It was frustrating during our period of experimentation. The information was still from live articles, but the model would create its own URL,” Hiddleston said.

Artifax has since worked through their learning period. The team learned that changing the model they were using can shift tone and behavior enough that everything needs to be retested. In practice, that means building a parallel copy of the agent on the candidate model, validating it independently, and only then flipping production over.

Once the team was comfortable, the agent moved out of testing and into the Artifax product itself, where the agent was soft-launched to customers through the help button.

How the support team has streamlined workflows

The clearest win has been with onboarding. A new hire in the support department, who was unfamiliar with Artifax’s more niche modules, was able to answer tickets in her first week after using the agent’s private ticket summaries.

Overall ticket volume hasn’t dropped dramatically, but Artifax is seeing very few chat sessions escalate into a logged ticket at all. Hiddleston says this doesn’t reflect a deflection of support issues, but the success of the new and genuinely self-serve support channel that removes friction for customers.

Internally, the team uses the agent daily, and the flood of repetitive FAQ tickets has melted away, freeing staff’s time to improve help centre articles, which then feed straight back into the agent’s own knowledge.

What’s next for Artifax

Artifax is now recording professional services and customer success calls to build a second, internal-only knowledge layer. This way, account managers will eventually be able to ask the agent for a summary of where a specific customer stands.

The team is also weighing a move from one platform to another, largely for a more mature audit trail as the stakes of automation rise through further AI adoption.

“Having a really strong audit trail and decision log is critical to adding any additional agents,” he said.

His advice to other businesses who may want to pursue a similar project? “Your data is already there. Go and reach out for it.”