AI agents: what do they really deliver? Our AI developer on agents and keeping a human in the loop

AI agents are everywhere. Every software vendor promises one, and every organisation wonders what to do with them. But what actually is an agent, and where does it really pay off?

We asked Thomas again, AI developer at Humanoids. He currently works at IVO Rechtspraak, the IT organisation of the Dutch judiciary. In part 1, he talked about sovereignty and local AI. This second part is about agents: what they can do, where they add value and where things go wrong.

A language model that takes action

Since ChatGPT arrived, much of the AI debate has been driven by marketing. The term AI agent comes from there too. Still, there is something real underneath.

"An agent is a language model with a layer built around it, a so-called harness," Thomas explains. "That gives it access to its environment: your email, your Slack messages, a search function."

An example makes the difference clear. Ask a standalone language model what a flight to Italy in November costs, and it will name a price it once came across in a brochure. An agent reasons differently: you want current information, I have a search function, so I'll look it up. "That's what allows it to take actions on its own."

Opportunities for smaller businesses

According to Thomas, the biggest gains lie in repetitive work that is still done by hand. Think of:

• copying data from invoices into Excel
• transferring spreadsheets into an ERP system
• recognising text in PDFs and processing it accurately

For smaller businesses in particular, this pays off quickly. Often a small team handles the admin alongside the real work, so every hour an agent takes over is felt straight away.

"You have to look critically at which processes an agent can take over. Setting it up still has to be done by a person, and it needs to be fine-tuned properly. But after that, it runs."

From programmer to architect

In Thomas's own field, programming, agents have already become indispensable. "I used to write every line of code myself. Now I send an agent off with clear instructions, and it directs other agents and puts together entire files."

The prerequisite is a good brief: which patterns, which conventions, which data to use and which not. "If you get that right, you no longer have to write syntax. You're less of a programmer and more of an architect."

Without boundaries, things go wrong

Precisely because an agent decides for itself which steps to take, clear boundaries are needed. "Otherwise the language model goes off track. If it can't get to the information, it tries another route, even if that route is undesirable. It thinks that's what you as the user want."

You set those boundaries in advance. You decide what an agent can access, what it may only read and what it may change. For actions with consequences, such as sending an email or preparing a payment, you have it ask for confirmation first.

Always keep a human in the loop

The biggest pitfall is blindly trusting the output. Say you have an AI produce the quarterly business report and base policy on it. If nobody checks the figures, you only find out too late that they're wrong.

"Is something business-critical? Then always have someone look at it," says Thomas. "You've still saved a lot of work on the input."

When is it worth it?

Ultimately, you measure the gain in time. If an agent takes over typing up, transferring and checking data, your people have time left for work that really needs their attention: helping customers, improving processes, coming up with new ideas.

What do you do yourself, and what does the agent do?

Which tasks you hand to an agent and which stay with people differs per organisation. One company gains most in invoice processing, another in reporting or customer queries. So start by asking where the time really goes today.

That's what we do at Humanoids. With UX research, we map out which processes lend themselves to an agent and which should stay with people. Our developers then build the agent, including the boundaries it works within. For smaller businesses that want to start with one process, and for government and corporates with more complex processes.

Curious which processes in your organisation an agent could take over? Book an exploratory call with us.