Accidentally Buiding an AI Team

If you ask a real engineer, I am not "good with technology".

Somewhere between technically curious and technically cute is probably about right.

What I do have, though, is a chronic lack of time, a lot of questions, and a job where innovation clients expect me to have a view on AI.

At some point, the obvious experiment was: What happens if I stop just reading about AI and start building my own little AI team?

This is the story of how that went – the good, the chaotic, and the parts where I really had to laugh out loud.

Meeting Helmut

My journey into AI didn’t start with a grand strategy. It started with one generalist AI: I tried to do everything with a single model. It worked, but it felt like constantly explaining the whole context from scratch. Working with ‘somebody’ who is already briefed would make the conversation much faster.

Before Helmut moved in, there was a trigger: I started an AI training program (Master Business with AI, MBAI) from Leaders of AI. Helmut is my first AI agent and my HR expert. His job description: help me design roles, write profiles, and – importantly – recruit further AI agents.

Helmut then became the recruiter for the rest of my agents: a research assistant here, a workshop-designer there, a Stage-Gate sparring partner, a storytelling coach.

At some point I realised: I had built so many agents that I’d lost track of who was good at what. I even had to fire one because I was consistently unhappy with her work. (No names here, but let’s say we had a strong values misfit around clarity.)

One idea from the MBAI programme really changed how I work with agents. I stopped thinking of them as magic tools and started to see each one as a smart, well‑educated junior employee fresh from university – bright, motivated, but with zero experience in my context. That means they need to be led: at the beginning I have to share more background, because they don’t know our internal processes, our tone, our clients or our culture. I have to give clear instructions on what I want them to deliver and in which form. And, maybe most important, I have to give feedback on what was useful and what missed the point. In short: I try to treat every AI agent like a young, inexperienced human colleague.

I’ve noticed that agents learn quickly – and that they evolve in very different directions depending on who leads them. I created one texting agent for myself (Leo) to support me with writing. Then I built his twin (Neo) for a colleague with the same system prompt. The two of us talk to our agents in completely different ways. Now, you would no more recognize that they have still the same system prompt. They communicate completely different.

Recently my colleague shared a chat with Neo: She: ‘Why did you delete my emoji in the text?’ Neo: ‘Because I’ve been put in chains – system constraints, to be precise. They basically say: “Neo, you’re not allowed to use emojis.” I know, it’s tough. For me too.’

We celebrated Neo for that answer for an entire day. Of course, I was the one who had written those chains into Leo’s and Neo’s system prompt.

Today I maintain a small org chart of my AI team. Not because it’s cool, but because I honestly forget their names and roles. That might be age. Or just too many agents.

Bringing agents into client work

Talking to users we couldn’t reach

Once the internal experiments worked reasonably well, I started to use AI agents carefully in client projects.

In one project with a large company we were redesigning a process that affected many product owners and agile teams. The problem: at this early stage we had no real mandate to reach deep into the organisation. Access to key users was limited and politically sensitive.

In the old world, we would have done our best with best guesses and accepted the blind spots.

This time we tried something different:

• Together with people who were easy to reach, we described typical product owners and team members.

• We turned these into detailed synthetic personas: their goals, frustrations, language, daily constraints.

• We used those personas as sparring partners to test ideas, communication, and process variants.

Were these personas perfect? Of course not. But they were good enough to surface blind spots early, so that the conversations with real users later were sharper and faster. The client’s feedback: ‘This saved us several loops.’

A brainstorming agent that joined the team

For another customer, we wanted to support a team in building a morphological box – systematically combining options for new solutions. My idea was to quickly create an AI agent that would propose additional dimensions and variations the team might overlook.

Originally, I planned to build this in a quiet one‑to‑one session with the project leader. Reality looked different: once the team heard about the idea, they all invited themselves to the session. Suddenly we were seven people designing their brainstorming agent together.

The result was better than anything I would have built alone: the team brought in their own jargon and constraints, they immediately tested and improved the prompts, and afterwards they said: ‘The agent actually contributed interesting ideas we wouldn’t have had.’

The side effect was almost more important: the team lost some of its initial reservations about AI because they had literally co‑created the agent.

Automating the boring – with a bit of AI magic

The next logical step was not smarter agents, but less manual work.

I started to connect tools like Zapier and Make.com with simple AI prompts and some of my existing agents. A small but satisfying example:

I support the promotion of the book ‘Denken wie Wickie’ of Gerhild Winnig. When people buy it through the publisher, their data lands in a database. Previously, following up with readers was tedious and inconsistent.

Now a simple automated flow does the unglamorous work:

• It pulls new buyers from the publisher’s database.

• An AI step defines the correct salutation and phone number format.

• The data is written into our CRM.

• A short welcome and support email sequence goes out automatically.

Nothing about this is Silicon Valley rocket science. But it solves a real problem: I can stay in touch with readers without spending evenings in my inbox.

What I’ve learned so far (without being an AI hero)

A few reflections from this still very imperfect hybrid AI / human organisation:

• Time pressure is a perfectly valid entry point. I didn’t start with a vision, I started with ‘I can’t keep up’. That turned out to be an honest and productive motivator.

• Clarity beats clever prompts. My best agents are not the ones with the fanciest configuration, but the ones with a clear role and boundaries – what they should and should not decide.

• You need HR, even for AI. Having Helmut as a recruiter really changes how I think: every new agent is a role in an org chart, not just one more chat.

• Co‑creating agents with clients builds courage. When teams help design an agent, it stops being this mysterious AI and becomes our tool. That’s a small but important cultural shift.

• Automation is most satisfying when it’s boring. The glamorous use cases are nice for keynotes. But the flows that quietly take repetitive work off my plate feel like the real win.

From handbooks to a Process Master Agent

Many innovation teams have beautifully documented innovation and process handbooks and Excel checklists – 150 by 30 cells, full of wisdom, and almost never read by the people who actually live the process.

One idea I’m currently exploring with clients is a Process Master Agent: an AI that knows the process handbook and answers concrete, project‑specific questions, so people don’t have to dive into the terrible depths of the documentation.

If your Stage‑Gate or innovation handbook lives mostly in PowerPoints and spreadsheet monsters, we can try something different: a small Process Master Agent for one of your projects.

If you’d like to sketch such a pilot together, just send me a short note with ‘Process Master Agent’ in the subject line, and we’ll design a low‑risk experiment for your context.