Agentic Operations (English)

How the mindset changes the solution

I want to design an architecture for an agentic AI system. While understanding the client requirements, it became clear to me that we need to begin a few steps earlier.

The article with the architecture design will come.

But first we have to meet people where they are.

In this article, we begin. The further articles then build on this one.

---

When I imagine what a truly helpful and meaningful AI system for a particular business area looks like, I think of something like this:

First of all, where are we?

I imagine several people who all have the same role. They are not techies. The people all work at the same company, but they are in a slightly competitive situation. Who accomplishes more and who does it better? That’s always in the background.

The people have many different work steps.

• Sometimes in one piece of software,

• sometimes in another piece of software,

• sometimes via email,

• sometimes they are out at external work meetings,

• sometimes they research something,

• sometimes they look at data in tables,

• sometimes they analyze data,

• sometimes they read work-relevant content on social media,

• sometimes they meet and discuss with each other,

• sometimes they create documents in the most varied forms

• and so on and so forth.

In addition, we know:

• The many work steps then culminate in decisions of consequence. If the decisions were good, success follows.

• There is one big goal that they pursue, and beneath it there are various sub-goals.

• They decide for themselves how to organize the work. There is a rough workflow that is prescribed. The details vary, however.

• The work steps often stretch over several days and weeks. In doing so, the medium is often switched.

• They often work alone, but there are also coordination efforts within the team.

A great deal of what they do is based on years or decades of experience. It is written down nowhere. The more experienced ones act on gut feeling, and do so successfully. Sure, mistakes happen too, but most of the time things go well.

When you ask the people, they find AI helpful, but nonetheless there are very large concerns about what the boss is now planning with the AI introduction…

All the hype on social media has unsettled one or another of them.

They would actually have ideas about where AI could help them, but they hold back with suggestions because they have no desire to be rationalized away by AI. People are already trying to shield themselves a little and help others less and not give away their knowledge under any circumstances.

And yet there would definitely be opportunities for relief:

• The people can hardly keep up with reading the emails, let alone processing them. Often there are also pointless emails that only cost time and are annoying.

• Research is very time-consuming, which is why that’s the first thing people skimp on.

• All the data they have to analyze is partly incomprehensible, which is why some have already built their own secret huge Excel as a workaround.

• There are also far too many meetings.

• Often one also forgets something or misses important opportunities.

• Then all the notes scattered everywhere, and one would most like to make voice messages when out and about and needing to jot something down.

• And these many documents where one has to manually type in data and information. Even when there are templates, it’s still enough work.

• One or another has ideas about how one could achieve much more, but they simply don’t get around to engaging more deeply with it.

But all of this is kept strictly under wraps. And of curse it is not communicated to the external AI consultants who are suddenly walking through the building.

Not all sunshine and roses

But there are also employees who take too many shortcuts and whose performance drops sharply compared to the others.

Who don’t follow the workflow precisely, who don’t proceed thoroughly enough.

With control and pressure you achieve nothing, but hushing up the topic is also wrong, because it depresses the morale of the others.

It has to be put on the table.

An ideal workflow was once defined, but hardly anyone sticks to it, mainly out of lack of time.

It stands and falls with how you see people

Since I was a teenager, I have worked. Vacation jobs, side jobs, during my studies and afterwards anyway. Always in large companies, always in software and data initiatives.

I have had a lot of opportunity to experience people in work situations. Since I never climbed the career ladder, but always worked either in the team or close to the team, my wealth of experience is truly immense. As a leader high up in the hierarchy, you don’t experience that so firsthand.

I can say one thing with certainty: the success of an initiative is connected to how you view people. How you see them.

If an employee is for you merely an executor whom you consider less competent, intelligent, or capable, then good luck with the implementation of larger projects.

It takes only one person who blocks or torpedoes everything. But you only notice that when it’s already too late.

The way you see people is the basis for being able to accomplish anything bigger together at all. And if you also want change and transformation, it’s the key to success.

Since AI unsettles many people, we first have to clarify how we see people, how we see AI, and how we see the combination of people and AI.

My stance on people, AI, and people and AI

My stance is: we treat everyone as equals.

• Nobody knows everything these days. Nobody can do everything. And for larger initiatives you need the knowledge and experience of many people.

• For something to succeed, everyone has to pull together.

• If I therefore need everyone, I also have to treat everyone equally. Even those who, in the opinion of the majority, perform less, I have to treat with respect. I achieve nothing with pressure and blame. Then they block all the more.

I have to create an atmosphere in which that is also felt and experienced. Saying it is not enough. The behavior and actions of leaders and AI consultants must reflect exactly that.

What can AI do really well?

How do I see AI? I think: start with less hype for now. Afterwards one can see whether more is possible after all.

I find AI very helpful with:

• Reading texts of every kind in large quantities

• Summarizing texts

• Researching in vast amounts of material

• Creating material

• Creating software

• Identifying questions or topics that I myself didn’t have on my radar

There are many more use cases, I’ve now limited myself to those relevant in this article.

My ideal vision of the interplay of AI and human

• For me, humans are the ones who make decisions of consequence and take responsibility for decisions.

• For me, humans are the ones who understand the meaning of something.

• For me, humans are the ones who can meaningfully combine all that is present and, based on their own experience, choose the right approach.

So this means:

• The AI researches, reads, fetches data, perhaps gives tips, but the decision still lies with the human. This means the human also needs time to review the things, to understand them, in order to even be capable of deciding. We’re not talking about trivialities now, but about decisions that have major impacts.

• The AI can read and interpret material, but the AI doesn’t notice what else is happening in the outside world, or what things were like in the past, or what the interpersonal interactions are, which is why an AI cannot fully understand what something means.

• The AI can understand and interpret prescribed structures, but one cannot possibly codify all possible constellations and special cases, which is why the human is still needed. One cannot anticipate everything in advance, and that isn’t even useful.

Different mindset, different AI solution

Did you notice while reading that with this mindset something different is being built than with the stance that the human should just work faster and ideally the AI does everything on its own?

The mindset determines what the solution looks like in the end.

What do we do about the conflicts?

So far we’ve identified two sources of conflict:

• the latent competitive situation

• the performance difference between the employees

This must of course be addressed at the same time. Introducing the AI solution and ignoring the conflicts causes unnecessary friction and costs. Why?

One person, for example, has many insider tips that have arisen over the years. Ones that really bring value. Ones that would also bring value to the others.

As a company, one has an interest in this knowledge being distributed to the others.

But the person has no interest at all in passing it on. On the one hand because of the competitive situation, on the other hand because one feels unfairly treated by the fact that the underperformers get off so easily.

This is a topic for humans, not for AI

A leader would have to deal with it, but perhaps one would rather have the AI handle it. Or the external consultant.

But it doesn’t work that way.

It’s a topic for humans, for leaders.

That’s why being an AI consultant is not a purely technical role. The AI consultant will sooner or later reach the point where they coach leaders.

Experience in conflict management is very helpful here.

Conflict management

There are many different ways to approach this. Here I describe one that, in my experience, works well.

Work with the group

Of course there must also be one-on-one conversations between employee and leader. But that alone is not enough.

All people with the same role must come into a joint meeting with the leadership team. Best in person or with cameras on.

Everything must be put openly on the table!

Getting along well doesn’t mean there’s no performance expectation

Speaking with one another as equals doesn’t mean there are no expectations. Of course there are requirements of the employees, but also of the leaders. One can speak respectfully with one another and at the same time address what one expects from each other. That isn’t a contradiction.

Be honest

In my experience, it comes across best when one is honest and authentic. You don’t need to worry so much that one isn’t allowed to say anything. Often everyone is even relieved that someone finally opens their mouth.

Speak the way a normal person speaks.

Say as a leader what you really think.

You’ve surely heard this saying before. The best AI strategy is worthless if you can’t implement it. The company culture determines how quickly you arrive at results.

In our example, you could start with what you generally appreciate about the employees. How you see and classify AI and what the vision for the future is.

Recall the roots of the company, describe how you envision the company’s future and how AI is meant to support people in this.

But then the conflicts must be addressed openly and transparently. Think of team sports. There, the games and weaknesses are also analyzed and discussed.

Establish in your company a way for this to happen regularly such that no one finds it embarrassing.

Explain that each individual is called upon. Explain that one finds individuality good, but that a quality standard must be upheld.

Explicit written standards

Make the quality standard explicit, put it in writing. Create various career levels and, per level, make measurable criteria that demonstrate the quality of performance.

And work together with the HR department to design the paths that must be taken when a person undermines the standards. And follow through with that consistently.

Also include in the standards, per level, what deployment and use of AI is expected in each case.

Distributing knowledge

Certainly one could distribute the knowledge via the AI. One analyzes what one person did and distributes the info to others.

But I would find another way better: the one via the group.

Ask the group what suggestions all the employees have for distributing the knowledge evenly to everyone. Don’t let up when everyone stays silent. Everyone should help think about how this works. Passing on knowledge should become part of the career path.

Afterwards, the knowledge can also be codified and thus again made available to everyone’s AI assistants. But I would always start with the human.

Transformation

If advancement within the company is coupled to value creation, you defuse many conflicts from the outset. That’s easily said, but not so trivial to implement. That’s why I always speak of transformation.

A transformation fundamentally changes the company. It’s a huge opportunity to tackle this. One can, in the course of this, also kick off many things that have been left undone.

So that the company doesn’t end up in chaos or in a huge change project, you need AI consultants who know how to proceed step by step here.

But now back to our AI solution. What could it look like?

First ideas

Okay, let’s brainstorm. What could one build? What would be helpful?

• The employee does their job as always, and the AI assistant is along for everything: in the interface of the software as a chat, on the phone via a messenger, connected to emails, meeting notes, and so on.

• The AI assistant is a real assistant. When the person asks: Where is such-and-such again? The AI assistant finds it. When the person says, okay, I need the data from xyz, ta-da!, there they are.

• The AI assistant researches, writes texts, summarizes texts, and searches for things.

• At the same time, the AI assistant keeps in mind that the employee can also improve their own work process. There is an ideal workflow, and the AI notices when a step was forgotten and reminds them of it.

• Or the AI assistant proactively makes helpful suggestions.

The employee feels relieved and can concentrate fully on where they deliver real added value: namely in assessing the situation, in decisions, in managing situations involving people, in weighing risks, all that where something important is at stake and which only a human can do really well, because the human has everything in view.

The AI assistant meanwhile diligently writes documents according to specifications, researches, informs, or works through to-dos.

I think that sounds great.

Whether one can build it exactly like that is another matter. Maybe we’ll have to make compromises. We’ll see that when the time comes.

Can we get started already?

No, not by a long shot. We first have to identify the right questions.

If we now jump straight into the solution, we risk that the user ends up saying, nah, this only annoys and doesn’t help at all, and the solution doesn’t get adopted.

We need a list of the right questions, and we bring these into various categories.

That will be the next article.

If questions already occurred to you while reading, then leave a comment. I’ll then take them along.

How do we proceed?

After that, we have to get to know and understand general concepts in building agentic AI systems. So to speak, learn what others build, how, and why.

In that course, we also have to define all the new terms and explain what they mean.

Maybe one or another has already come across them; people speak of context, memory, knowledge, semantics, in themselves not new words, but what exactly is meant by them here? Then there’s state, guardrails, evals, orchestrator, observability, MCP, prompt injection, etc., and so on.

Before we plunge into the design of the architecture, we first have to establish a common understanding.

Very exciting too will be the question of where one begins to build. Which use cases do we want to look at? Which do we select and why?

And then, of course, the undertaking itself: how one can tell whether it was successful?

Where do we start, how do we continue? And with which project management method? Agile anyway, but what does that look like concretely?

Finally, things continue with the organization itself. What, then, is the target vision of Agentic Operations? How do we get there? And how does it change our company?

As you can see, a lot of work lies ahead of us. All of this and much more we will also consider in upcoming articles.

I hope you enjoyed the article. Until next time!