Would you give an AI agent access to your smartphone, email, calendar and bank account? The answer, for now, is probably no. Yet, agentic AI is eager to enter our lives. This raises questions, for example, about how to automate (back-office) processes using AI without going off track. 

 

In customer contact, "AI agents" or agentic AI are common terms, but at QuandaGo, they prefer to call them "digital agents." This term indicates that virtual agents don't just work with AI, but also collaborate with traditional automation.

This distinction becomes even more relevant now that the practical application of agentic AI has come a step closer. Much has been written in the trade media in recent weeks about ClawBot, among other things. This is an AI agent that not only converses but also performs actions independently. You could consider it the consumer version of the AI ​​agent offered by a growing number of vendors. It's high time to revisit Raoul Fasel and Jeroen Kromme of QuandaGo.

Can I have a kilo of AI?

"When handling a customer query or solving a problem, employees go through various steps, sometimes with the help of digital agents and traditional automation solutions. Organizations that are starting to apply AI in these kinds of processes tend to open up new paths for AI to follow, in addition to standard automation. They're simply opening up a whole new world of AI functionality," says Raoul Fasel, product manager at QuandaGo.

We don't support that. After all, if digital or AI agents get stuck in a process or if the customer requests it, they should always be able to hand over to a human employee. If that person then takes over the process, they'll have to rely on well-thought-out, structured steps within existing applications. An example is submitting meter readings. If an AI agent can't complete that process, it's passed on to an employee. This employee follows the same steps as the digital agent. An employee shouldn't simply make changes to a MySQL database independently—interfaces and controlled steps are designed for that.

Avoid a second process layer

Jeroen Kromme, head of product & data at QuandaGo: “If you want to keep your processes clear and secure, it is wise to ensure that both your human employees and digital agents are connected to the same processes. If you don't do that, you create a new, second layer of processes and automation specifically for the application of AI. And then any handovers more complicated than necessary.”

That handover remains important, Jeroen emphasizes. "When resolving customer queries, you always have three options: you have an employee handle everything; you use AI to assist the human employee—for example, by preparing or finishing something—or you have a digital agent handle everything. In all cases, you want to be able to assign a query to an employee because the customer requests it, or because the digital agent can't figure it out."

 

There's a lot of buzz about applications like OpenClaw right now. Is this the next step in agentic's development?

Kromme: “A clear shift has been visible for the past six months. Some AI suppliers are explicitly opting for agentic, but others have switched to deterministic softwareI think that's driven by the growing importance of governance. Companies are searching for the right direction. OpenClaw isn't new, as the major players in the agentic field already have this capability. In a sense, OpenClaw is the consumer version of agentic AI: you give the agent access to everything. Of course, there are risks associated with that.

 

With these kinds of solutions, can we say that consumers will soon be deploying software on contact centers instead of the other way around, as with AI chatbots intended for customer contact?

Kromme: "There will indeed be an additional channel for customer service. I expect that serious agentic AI products for consumers will be tried out first by 'prosumers.'

Fasel: "I expect that in the coming period, companies will be faced with increasing amounts of messages, interactions, and traffic from consumer AI agents. Organizations will have to deal with this overload of incoming content."

 

We're actually talking about the machine customer that's truly emerging. What can you do as a contact center to accommodate this?

Fasel: "Bots will start talking to bots. That sounds like a roundabout way, but it's also necessary to keep interactions traceable and verifiable—for example, because you want to be able to intervene. You also need to be able to identify a bot. Customer verification in the contact center isn't always well-organized—think of postal code and house number. Moreover, your customer can only call once at a time, or they're in a queue; a bot can scale itself and fire a whole series of interactions at you one after the other. For example, focused on submitting a request. You can, of course, recognize that you're facing a kind of DDoS attack from a customer's AI agent."

Kromme: "These kinds of scenarios clearly demonstrate why QuandaGo chooses deterministic processes as a starting point. This allows you to handle many matters watertight, with clear parameters. If a customer request falls outside the parameters, it won't be honored. Or if a meter reading doesn't make sense, it won't be recorded. As an AI agent, you can then repeat your request in endless ways or have it search for a workaround, but the process software checks requests using scripts with strict conditions."

The ultimate security is to not allow the core processes to be compromised. large action models (LAMs) and AI agents to be managed, but by traditional software, as we've always done. "They can be started by an AI agent, but not executed," Kromme emphasizes.

A digital agent for every use case

QuandaGo puts this combination of a deterministic approach and the power of generative AI into so-called value packages: specialized solutions for very specific processes, recognizable for specific industries or sectors.

Fasel: "Think about recording and submitting meter readings. For example, before this conversation, we've set up an agent you can call to request the Ziptone newsletter. We're looking at all sorts of sectors, of course: what are suitable processes to incorporate into a value package?"

Digital agents embrace workflow automation

Kromme: "We've now made our in-house workflow automation solutions available to digital agents as well. QuandaGo has focused on industries with complex processes and strict governance regulations, such as the energy sector and financial services. Think of a switching agent, an agent for outages, or an agent for adjusting installment amounts. Multiple digital agents can also be deployed alongside each other, call each other, or collaborate with each other."

“For example: when you 'call' a digital agent, another person is called LLM Then use it with the digital agent that initiates a transaction or process and checks its output—think of the aforementioned meter reading. You can then assign a conversational agent the task of informing the customer that the submitted meter reading has been checked and recorded. This allows you to set up more complex systems. It's obvious that you should start with high-volume, low-impact processes. But the digital agents will never perform actions themselves, only initiate and complete them.

ClawBot as a consumer version

Fasel: "Applications like ClawBot or OpenClaw may attract attention now, but their use in a business environment is prohibited under the European AI Act, simply because the process is inexplicable."

The bottom line is that you primarily use agentic AI to reap the benefits of conversational AI, that you use deterministic automation to keep processes compliant, and that when combining the two, you assume small process steps with multiple agents. This brings the principle of orchestration into view. "That's where we're headed, but it's just the beginning. So we're choosing a different approach," says Kromme, "namely, only using AI when it's safe to do so, and letting classic automation handle everything else."

(Ziptone/Erik Bouwer)

Featured, Knowledge base, Knowledge partners, Technology QandaGo