Guardian360
Start free trialBecome a partner
← All posts

Opinion

AI agents need no social maintenance. Your organisation does: how to safeguard organisational culture in the age of agentic AI

By Jan Martijn Broekhof ·

This year, our team at Guardian360 rebuilt its platform with AI agents in roughly three months. The previous version took us more than three years. We are a small team. A shortage of hands is no longer what holds us back. What holds us back is keeping our partners, and each other, on the same page.

That observation stayed with me while reading Storylistening by Annet Scheringa and Simone Beemster, and Ethan Mollick’s reflections behind his new book Co-Existence. Scheringa and Beemster write about the two currents that run through every organisation’s culture: the one you measure and the one you feel. Together, they made me ask a question I suspect many leaders will face soon: if people hand more and more of their work to AI agents, who looks after the undercurrent?

The current you measure and the current you feel

In Dutch management literature, organisations have a bovenstroom and an onderstroom. The upper current is what we can see and steer: facts, plans, targets and KPIs. The undercurrent is what we feel: emotions, implicit processes, unwritten rules, and everything that grows when people work together.

Scheringa and Beemster argue that seducing people, presenting alarming figures and emailing change plans around simply does not work, while listening to stories from the undercurrent does. They also observe that many organisations leave that world of experiences, perceptions and emotions untouched. In my experience, that is exactly where an organisation either holds together or quietly falls apart.

What happens to the undercurrent when agents do the work?

A field experiment by Harang Ju of Johns Hopkins and Sinan Aral of MIT, with more than 2,200 participants, offers an uncomfortable clue. Teams working with AI agents sent 18 percent fewer interpersonal messages than human teams. The researchers explain why: agents follow instructions reliably and need no social maintenance, so human and AI teams bypassed much of the social overhead.

The researchers present this as efficiency, and in a narrow sense it is. Yet they also acknowledge that these interpersonal processes are not merely overhead; they build trust and psychological safety. Read it again with the undercurrent in mind. The overhead that disappeared is the clarifying question, the small talk before a task, the “can you look at this with me?” That overhead is where trust is built and where a junior learns how a senior thinks.

An article from UNSW Business School, featuring Professor Mary-Anne Williams, describes where this leads: experienced staff stop mentoring because they assume the agent handles it. Nobody decides that. It simply happens.

A company brain only remembers what was said out loud

At Guardian360 we work with Hermes, our own AI layer. It runs on open-source models in our own server cluster in the Netherlands and acts as a company brain for the agents that support our work. It is remarkably good at everything we have written down.

It does not know why we turned down a partner request two years ago. It does not know which partner needs a phone call rather than an email, or that a colleague has gone quiet in meetings lately. The philosopher Michael Polanyi put it simply: we can know more than we can tell. A company brain makes the upper current faster and more complete. The undercurrent it cannot hold.

Doesn’t AI make us better at sharing knowledge?

The strongest counterargument is that AI improves how knowledge flows through an organisation, and it deserves a fair hearing. A 2026 study in Acta Psychologica found that collaboration with AI agents strongly predicted knowledge sharing, job thriving and well-being, while stressing that organisational climate matters.

A study by Liu, Gao and Yu in Frontiers in Psychology adds nuance. AI usage increased both knowledge sharing and knowledge hiding, with sharing the stronger effect, and a competitive climate strengthened both.

My reading is that AI amplifies the undercurrent you already have. A healthy undercurrent leads to more sharing; a competitive one leads to more hiding. That does not make the undercurrent less important. It makes it decisive.

When should you refuse AI’s help?

Ethan Mollick, whose Co-Intelligence sold particularly well in the Netherlands, has just published Co-Existence. Where his first book described humans at the centre with chatbots as helpers, the new one is about agents that are sometimes better than us and sometimes worse.

Mollick wrote every chapter himself, because of what he calls an “underlying contract” with his readers. He is also candid about the price: some intellectual struggles may have been resolved too quickly. One of the questions he now asks is when you should refuse AI’s help, even when it is on offer.

My answer: refuse it wherever the friction of thinking, struggling or working something out together is the point.

What is an engineer for, when agents write the code?

If agents write most of our code, what are our engineers for? Our answer is that we want to give the time agents save back to meaningful conversations. Conversations with partners about what they actually need. Conversations with each other about why we build something, not only how. Conversations about what is bothering someone.

This is harder than it sounds. Many engineers chose their profession because they love to build. Code is their craft and part of their identity. Telling them that coding is no longer their most important task can feel like a demotion.

It is not. Understanding the problem, exercising judgement and reviewing what agents produce become more important, not less; at Guardian360, AI-written code is reviewed before it reaches production. But the shift has to be learned, and it needs leaders who value a good conversation as highly as a merged pull request.

There is a second trap. Time saved by agents does not flow to conversations by itself. Without a deliberate choice, it flows into more output, because agents can always do more.

How do you safeguard the undercurrent in an agentic organisation?

Start by listening to stories, not only to dashboards. Storylistening is a discipline: ask people about a specific moment, then listen without steering. The patterns you hear will tell you more about your culture than any engagement score.

Next, protect conversation time as real work. If it is not in the planning, it will lose to the next feature every time.

Finally, keep certain moments deliberately human: onboarding, mentoring and the conversation after something has gone wrong. That last one matters for resilience too. When an incident happens, the undercurrent decides whether someone dares to say “I think I clicked that link” in time. No agent can create that kind of safety.

The question that remains

AI agents need no social maintenance. Your organisation does. The real question is not how much work you can hand to agents, but what you do with the time you get back.

When did you last hear a story in your organisation that no dashboard could have told you?

Sources