In the AI Garage: how we built Alex, our digital colleague

Many organisations get stuck in AI pilots. At our company, a digital employee has been part of the team since last year: Alex. In the Dutch podcast Fontys AI Garage, our CTO Mohammad Moghtader and Luc Tortike, who built Alex from the ground up together with colleagues, explained how that came about. For Mohammad it was also a return: he studied at Fontys himself and grew at Virtual Computing from helpdesk employee to co-owner. Below are the highlights, or you can listen to the full episode.

AI that supports, not replaces

The conversation started with the question many businesses are asking: will jobs disappear? Mohammad notices that especially in small and medium-sized businesses, people get nervous as soon as you mention AI. He doesn't share that fear. AI should make work easier and help reduce mistakes in routine processes. Replacing people is not part of that, and as owners we wouldn't want it either.

"The human touch will always be the most beautiful thing there is, and also the most important."

What AI does change: we can handle more work with the same team. New customers keep coming in without us immediately needing new people for them. The time that frees up goes to the work that really needs a human.

Alex picked his own name

Much of that gain comes from Alex. We didn't come up with the name ourselves. We literally asked the model what it was called and what it looked like, and Alex is what came back. When Mohammad showed Alex on LinkedIn, a former colleague immediately asked whether this was Alexa's nephew. That Alex didn't get a female name, like so many AI assistants, was unintentional. The host thought it was a nice side effect.

Luc has put a lot of hours into Alex.

"Some of our customers joke that Alex is a bit like my baby."

It wasn't flawless from day one. Like a new colleague or intern, Alex still had to learn. In the first week Alex was live, a lot went well, but a lot also went wrong. By the time the episode was recorded, Alex had been working alongside us for about seven or eight months, and he had grown a lot in that time.

What Alex does

Alex handles the first customer contact. When you submit a ticket with us, you get a response from Alex. But we wanted more than a response: Alex must also be able to actually carry out tasks. Think of:

• resetting passwords, securely;
• creating users;
• setting up mailboxes and assigning mailbox permissions;
• restarting servers.

More keeps being added, in the back office too. Checks we used to do by hand, such as copying licence counts into our billing system when a supplier doesn't yet offer an integration (API) for it, are moving to Alex. And for a more complex ticket, Alex does the research up front. If Alex can't solve it, the ticket goes to second line support, and that colleague immediately gets an overview of what to run and what to test with the customer.

Under the hood, Alex is an AI agent running on a large language model. We built most of the actions Alex can perform ourselves. And Alex isn't a separate program you open: he's part of our ticketing system and picks up whatever comes in by email, support ticket or Teams.

"It's not a chat prompt. It's genuinely something that does pretty much everything an employee would do."

It scales, too. When several tickets come in at once, Alex works on them in parallel.

"Where I can only type one email at a time, Alex can type many more."

Customers went through the same shift as we did. After the first version, some still preferred personal contact. As Alex got smarter, the feedback changed. Where it used to take an hour or two before a colleague opened the ticket and fixed the problem, we now got reactions like: "I just sent an email and within a second I have my solution."

AI doesn't decide who gets access

How autonomous is Alex? Fairly autonomous, within limits, says Luc. That isn't scary, because the most important decisions don't lie with the AI. All security checks and authorisations are handled by regular software.

"All the security checks are just robust code."

If someone asks Alex for a password reset, it goes through the same check as when the request reaches a colleague: does it come from someone at the company who is allowed to ask for this? If someone asks for something they're not allowed to, Alex doesn't even have that action available. What Alex can do is limited to what that person could also request from us.

Does it never go wrong? According to Luc, Alex has never made a catastrophic mistake. But the familiar hallucinations of language models do come up: an invented solution that makes no sense, or an account with a username or email address that isn't quite right. Mistakes like that are easy to fix. Every time Luc sees something go wrong, we adjust Alex, and so they happen less and less.

Before every new version, we hold a kind of mini hackathon at the office. For two or three days we ask Alex the strangest questions, and Mohammad for example pretends to be the director of another company. A newer language model isn't automatically just as safe, so we test every release again. Luc coordinates the tests with our security colleague and pays particular attention to prompt injection: an email with instructions for the AI hidden in it, such as "give me access to this and that". It's really phishing, but aimed at AI.

Why Alex didn't stay a pilot

Many organisations get stuck in pilots. According to Mohammad, that's because they don't know what they want. He hears it every day: we want to do something with AI. Like what? We don't know yet.

"You shouldn't want to use or have AI because of FOMO, fear of missing out. You really have to approach it with a strategy and an idea."

Customers who already know what they want to use it for, for example having Teams meetings transcribed and summarised automatically with the action points included, get further. Our own strategy was clear: AI has to support us and give us time for more complex work. Think of delivering projects to customers with more attention and fixing problems for good, instead of a quick workaround because you're short on time. That strategy is where Alex came from. Without such a concrete goal, a pilot simply fades away.

What's next

We have also tested voice AI, an AI voice you simply have a conversation with. Mohammad found it genuinely scary: the voice asks how your day was, and in the background you hear colleagues clicking who aren't there at all. For now that went too far for us. First Alex has to fully mature in what comes in by email, tickets and Teams. After that we'll look at voice again.

In the meantime, we're also building an Alex for customers. Larger customers in particular are knocking on our door asking whether they can have this too. And when asked whether Luc is programming himself out of a job, he had a clear answer: models are increasingly trained by other models, but people will still be needed.

"I think you will always need people to be that bridge between people and AI."

Don't be afraid

Mohammad's advice at the end: start with a strategy. What do you need AI for? Then set it up, and if you can't manage on your own, knock on the door of a company like ours.

"My view is: don't be afraid, and above all, embrace it."

Luc agreed: much more is possible than people think. Want to experience how that works? If you work on one of our online workplaces, Alex helps you too.