Youforce automates internal support with its own AI agent

One support portal, endless confusion

Youforce, formerly known as Visma | Raet, is a Dutch cloud-based HR software platform that digitises HR and payroll administration for organisations across the Netherlands. Internally, they already seemed to have the right setup: one central support portal where employees could go with questions about IT, finance, facilities, HR and networking.

But behind that single portal sat countless departments, each with their own responsibilities. Employees regularly ended up in the wrong queue. To make matters worse, the portal contained seven different templates per category: one for a broken coffee machine, another for a leaking toilet and yet another for catering requests.

Click the wrong template and you had to start over from scratch. And if employees did not know where to begin, they simply sent a Slack message to the colleague who had helped them last time.

Stick to what you’re good at

Steven Koppenol, Head of IT at Youforce, saw two options. Option one: hire a first-line support employee to manually sort and route all incoming requests. Option two: automate the process intelligently. The choice was easy.

“We’re a tech company, so we chose smart technology,” Steven explains. “It simply doesn’t fit our identity to solve this with manual work.”

At the same time, Youforce was in the middle of reorganisations, mergers and rebranding efforts. Systems had been consolidated, people were still finding their place and the organisation was heavily focused on itself. AI readiness did not come naturally under those circumstances.

But with renewed focus and a fresh set of principles, the moment had arrived to take AI seriously. The use case for Incentro: internal support.

YARA: the colleague that's available 24/7

For Youforce, we built the Youforce Automated Request Assistant, or: YARA, an AI agent living inside Slack, the communication platform employees already used every day. No new portal to learn. No extra step in another system.

Here’s how it works: an employee types a question or issue into Slack, whether that’s just two words or a full description. YARA analyses the request, retrieves the user’s context and checks whether a similar ticket already exists. If not, the agent asks follow-up questions until enough information has been collected to complete the ticket properly.

Once everything is clear, YARA creates the ticket in the correct location within the service portal, confirms this to the employee with a ticket number and ensures everything is traceable in Jira as well.

Prompt engineering

Maran Leen, consultant at Incentro and closely involved in building YARA, is very honest about the challenges.

“You can click together a working AI agent surprisingly fast, and at first glance it already looks useful. But there’s so much more work involved before an agent is truly production-ready.”

One of the biggest challenges was routing: how does the agent know exactly where a ticket should go across dozens of departments and categories? And how do you ensure those decisions stay consistent?

“As a developer, you’re used to code being true or false. An LLM works with probabilities, not if-statements. That requires a completely different mindset,” says Maran.

The solution came from a combination of extensive prompt engineering, knowledge files packed with examples and a clear separation between what is controlled through prompts and what is deterministically fixed.

Guard rails: raising the agent properly

One thing users never see when an AI agent works well — but where most of the work actually goes — is the guard rails.

YARA was not allowed to expose configuration files, create tickets on behalf of other users or blindly fulfil every request. Because LLMs naturally want to help. Sometimes a little too eagerly.

During testing, the agent created tickets for completely irrelevant requests, suddenly opened a ticket under the name “Manuela” — someone nobody at Youforce had ever heard of — and could even be tricked into creating tickets on behalf of colleagues.

Every single one of those scenarios had to be shut down individually. On top of that, YARA is legally required to inform employees during the very first interaction that they are speaking to an AI agent, not a human being. That disclaimer is now built directly into the opening flow.

Step by step, test by test, YARA matured into a production-ready colleague.

Live in production....and still evolving

YARA is now live. Employees can create tickets through Slack, request status updates and add comments to existing tickets. Routing works well for most categories, and the team continues to improve the system every day.

Steven remains realistic about where they are today: “It’s not completely flawless yet, but we have to embrace the fact that it’s not perfect. AI agents are not magic. We’re constantly programming, refining and shaping behaviour until it truly works.”

The next step is expanding YARA’s skill set, always with human oversight as a core principle: no action without user confirmation, and no autonomous decisions unless the agent has been properly trained for the task at hand.

* This blog is based on the presentation by Steven Koppenol, Head of IT at Youforce, during IncentroCon Agentic ’26.