What 3,000 AI projects taught us about implementation

Today, Alexander Thamm is recognized as a European leader in the number of agentic projects underway. Their recent white paper, ‘The Agentic Shift’, outlines what this future could mean for companies seeking to remain competitive.

We sat down with Marc to discuss what he’s learned, why Europe’s regulatory approach could be an advantage and how to inspire those organizations who still have the “AI is not for us” mindset.

How do you describe your work to those outside the world of AI?

We make things easier and faster using math, statistics and computers to help organizations run their businesses.

You help companies transition to getting the most out of their data. I noted on your CV that you also did some transitioning.

That’s true! While I was the kid who took apart computers when I should have been outside playing and have always worked with data, I only dove deep into the inference – or “AI” – side during my master’s degree in business. I wanted to do something more technical. Instead of writing the 100,000th thesis on SWOT analysis, I built a recommender system for films and jokes. That was my first hands-on AI project.

Ha! That’s great. And you are now at Alexander Thamm. What makes your company unique?

We were one of the first service providers in Europe focused solely on data and AI. We’ve been doing this for over 10 years now – around 3,000 projects. That's a unique differentiator. Whenever an organization approaches us, we’ve done something similar in 99% of cases. We can pull up working examples immediately, not just slides.

We also went all-in on agentic AI three years ago, before there was even a job market for it. We hired people from neuroscience, software architecture and physics – people who understand how to model systems that make autonomous decisions. They are the agentic experts of today. They’re building the agentic systems companies need, not just talking about them.

What are you most proud of?

My focus is chemicals and pharma, primarily in Germany but also across Europe. We’ve built a strong footprint in this industry, which is under massive pressure – oversupply of basic chemicals from China, bureaucracy, energy costs. At the same time, people with 30 years’ experience are retiring.

That’s where AI comes in: how do you optimize energy consumption in plants while also capturing knowledge from people’s heads into a model others can query? We’ve shaped this connection over the years – from global enterprises down to a family-owned adhesive shop in rural Bavaria that heard about our work and asked if we could build them an AI customer assistant. That made us realize: what we've been doing for a decade is now really becoming mainstream, even for small firms.

There’s a strong message from Alexander Thamm about doing it for Europe. What’s your view?

We need to build on our high-quality education system and R&D strength – traditional strengths in engineering, chemistry and pharma. Significant inventions have been made here. But we can’t continue being dependent on US and Chinese models and data centers. That’s a massive geopolitical dependency.

Here’s the counterintuitive part: European regulation can be an advantage. Yes, there’s bureaucracy, but because we were among the first to regulate AI with the European AI Act, we’ll learn early how to fine-tune it. This creates a relatively safe investment environment. If you invest in AI solutions in the Netherlands, Germany or Italy, it's much less likely than elsewhere that someone will come along later and scrap your investment because your model is suddenly prohibited.

Regulations can also help build trust – bringing the general public and future end users into the fold. Do you see trust as essential for adoption?

Absolutely. You can build technologically excellent systems, but in the end, there’s a person – a patient, a doctor, a lab technician – who has to be open to adopting it. You can’t force them. They need to trust that their data is protected, or that they won’t lose their jobs. These are fair and legitimate concerns.

Do you see these cultural aspects as one of the main challenges ahead?

Technology isn’t usually the limiting factor anymore. It’s cultural. People worry the agent in the laboratory will replace them. We need to address these concerns, make them try the technology, build it and learn.

For AI, the classical IT strategy worldview doesn't work – plan two or three years ahead, write down requirements, build, test, roll out, done. With agents, since language models play a huge role, there’s a non-deterministic reasoning component. It's not trivial to get the exact same outcome up to the last decimal every time. If you need that, you need to give your agent the right tool, say, a regression rather than a language model. Not everything needs agents – that’s nonsense. Right tool for the right job.

What’s your big message for the industry?

Let’s get started. It sounds trivial, but I see strong interest from CEOs and heads of AI at chemical and pharma firms, increasingly also at smaller companies. At the same time, they often think AI isn’t for them – it’s exotic, something only big players like Covestro do. That’s not true.

Can we afford to be that risk-averse? We can learn in secure sandbox environments. Others who started early are now building a competitive advantage. You’re already using IT systems, you already have data and you already have ways to secure it. Just get started.

Concrete advice for making the cultural change?

Start small and simple with one use case. Ask people on the ground doing the actual work – not only through some top-down strategy – what problems they face today. Bring in someone with technical AI expertise and work it through together. Involve them. Don’t build something and throw it over the fence.

Run an ideation workshop with these people, then start building the first version. Take it from there. This is how successful companies start.

When do you see people’s eyes light up and think “Ah, AI is for me!”.

Two moments. First, in these ideation workshops, when we show 10 similar projects we’ve built – not Googled, but actually built. When people see someone else using it, they realize it’s possible. You can tell when they start asking curious questions. The ideas start flowing.

The second moment, even bigger, comes when they have the first prototype, usually two to four weeks later. It’s not perfect, but it gets AI off PowerPoint. It’s tangible, on the screen, and concretely improvable. Seeing is believing. That’s when they really believe: we can build this for us.

What personally excites you most about AI?

Remember Clippy?

Only that it was annoying and universally mocked.

[Laughs] Well, it was supposed to be your desktop assistant, but it never quite worked. Now, 20 or so years later, with language models, we’re closer than ever to achieving that personal helper we can train and customize.

Another thing that excites me is world or physical models – bringing AI into the physical world. In pharma and chemicals, that’s robotic arms handling samples and using physical tools in the lab. Imagine the convergence of language models you can talk to, with robotic platforms that act in the physical world.

Yes, amazing, but also scary.

Absolutely. I worry about safety, both physical and autonomy. I wouldn’t recommend deploying agents in safety-critical areas without first doing serious homework.

Is there anything else that makes you nervous?

The blurring of what’s real. Fake news, misinformation – there’s huge potential to be misled. We know political actors use this in cyber warfare. That worries me.

If you had to distill your advice into a few key principles, what would they be?

Let’s see how far I get...

Number one: just start. AI is for you, too. It's not the cutting-edge stuff that only big players do anymore. I’ve seen it multiple times – those moments when people’s eyes light up and they say, “We can do that today?!? I didn’t know.”

Number two: Don't be too risk averse. Not taking risks is a risk in itself. Whether you build AI skills or not, your global competitors will. It’s really about risk appetite, especially in the EU. The technology is ready. The risks are manageable if you start small and learn as you go.

Number three: find the right partners. Work with people who’ve actually built these systems, not just talked about them. Many mistakes have already been made; you don’t need to repeat them.

Number four: think about geopolitical dependency. It’s crucial that we reduce dependency, as we are in fact already reliant on overseas data centers and models. We need to mitigate that. It’s not just about competition, it’s about having control over your own tools. Is four enough?

More than enough. Especially if it means Clippy finally gets his moment.

[Laughter] Exactly. The vision around Clippy was never wrong – we just needed a couple of decades of better math and technology. And that’s really the big story. What’s considered science fiction is a lot closer than many imagine. Stop waiting. Start building.

Watch Marc and experts from Chevron Phillips Chemical Company and Bristol Myers Squibb in the panel discussion ‘AI in R&D: Driving innovation with agentic AI’.

Related Articles:

AI agents in action: Uncovering our inner mutants

The secret of successful AI pilots in R&D

LLMs as a Jury: Bringing Quality to Quantity in GenAI-Aided R&D

From quicksand to bedrock: How data quality shapes AI