We think most of it shouldn't be built.
We started this because we kept watching the same thing happen. A company would get excited about AI. Someone would build a demo. The demo would work. Then twelve months later, the workflow was still running on the same spreadsheets as before.
We wanted to understand why. This is our attempt to do something about it.
Elements Agents is led by Vladimir de Ziegler. Vladimir is an AI engineer who's spent the last few years shipping production systems for companies ranging from 10-person startups to enterprises with 2,000+ employees. Before this, he ran an AI lead-gen business. Before that, he was solving operational problems at high-growth companies.
He works with a small bench of engineers, designers, and operators built with over years of projects. When we need more, we bring them in for specific work. The "we" on this site is real. It's just small on purpose.
We work with teams across Europe and the US.
Three years ago, we were building AI lead-gen workflows for B2B companies. Models were getting good. Tools like LangChain and the OpenAI APIs made building fast. If a client asked for an agent, we built them an agent.
It worked. And it didn't.
The demos worked. Some of the production systems worked. A lot of them got quietly abandoned six months in.
The pattern was almost always the same. The model could do the task. The integration was fine. But the team didn't trust the output, or didn't know how to fix it when it broke, or quietly kept doing the work manually because the cost of a wrong AI output was higher than the cost of the manual work.
The problem wasn't AI engineering. It was everything around it — evals, feedback loops, governance, change management, handoff. The stuff that doesn't demo well but decides whether a system lives or dies.
We stopped pitching "we'll build you an agent." We started pitching: we'll figure out whether an agent is even the right answer, build it if it is, and make sure your team actually owns it by the end.
That meant smaller projects sometimes. Occasionally telling a client "don't do this." Often building in places that weren't glamorous — internal tooling, ops workflows, back-office agents that nobody would post about on LinkedIn.
The projects started lasting longer in production. The case studies on this site are the ones that did.
There are a lot of people who can build an AI demo. Very few who can sit across from a mid-market CEO, tell them which AI projects are worth doing, ship the ones that are, and leave when the team can run them without us.
That's the work we're committing to. The podcast is an extension of it — we talk to executives and AI champions at real companies who've crossed the messy middle, because if we're right about this being the category that matters, we're not going to be the only ones doing it for long.
Conversations with executives and AI champions at companies pushing AI into production. What worked, what didn't, what to expect. Guests so far include Abraham Gomez (Google), Eliott Wertheimer (VanMoof CEO), David Arnoux (Growth Tribe), Tim Masek (Storetasker), Alexis d'Eudeville (Lemlist), Sean Griffith (Truffle), Charlotte Lucas (ScorePlay), and Max Emma (Building Bookkeeping).
Listen →Monthly AI workshops for owners, operators, and builders shipping AI in real companies. Free to attend.
Join the next one →Weekly YouTube videos on the tools and patterns we use in production. A year in. A few viral. Most not. Both useful.
Watch →One of our fintech agent projects was featured in Google Cloud's case study series.
Read →The ones that can are worth doing properly. The ones that can't are worth saying so.
The messy middle is. Evals, feedback loops, team adoption, governance. That's the work that decides whether a system survives.
We arrive as engineers. We leave as a team that made your team capable of running what we built. If we're still there in year three doing the same work, we failed.
If we can't explain what we're building to your COO in two sentences, we haven't thought about it enough.
The first step is always a conversation. 30 minutes. No deck.