Scroll through any tech feed and you'll find someone who built a working SaaS product over lunch, powered entirely by AI. It's inspiring, until it isn't.
For every viral success story, there's a much larger, much quieter group of developers who tried something similar, hit a wall, and never posted about it. That silence has a cost: it sets unnatural expectations, both for developers who assume they're doing something wrong, and for clients who now walk into scoping calls expecting a Facebook clone by Friday.
That's the gap Dropsolid's Dieter Blomme and Valery Lourie are addressing head-on in their upcoming DrupalCon Rotterdam session. Rather than another highlight reel, they're bringing real projects, real mistakes, and the lessons that came out of them. These are the kind of failure that actually teaches you something. Here's what's on the agenda.
Why talk about failure at all?
Expect the session to open by naming the elephant in the room: people are hesitant to talk about what didn't work. A strong bias toward success stories creates real friction. It breeds a "why even start" mentality for developers comparing their first rough attempt to someone else's polished demo, and it hands clients a brief shaped by hype rather than by what AI tooling can actually deliver today.
Add in the extra layer that comes with operating in the EU, including questions about who processes user data, where it's hosted, who can see shipped code and configuration, and how vendors fit into the supply chain, and you've got the real backdrop this talk is set against.
What's on the agenda
The session digs into four real projects, each with its own hard-won lesson. There's Valery's personal trading application, built on Opencode, a local Ollama model, and an agent-suggested stack of Laravel, Laravel Reverb, and React, where two deceptively good weeks of all-green tests turned out to be false confidence rather than real progress, a sharp reminder that AI accelerates writing code and, with it, the rate at which technical debt can quietly pile up.
There's Dieter's fireflyiii-mcp, an MCP server built to query the Firefly III finance manager in plain language, where AI nailed the speed of getting there but quietly missed a test suite that tested nothing, an architecture that swung from one giant file to over-engineered chaos, and, the scariest find, a shared access token baked directly into the container that the agent never flagged as a risk.
There's an honest Drupal 7 to 11 migration experiment, where AI got a deliberately simple site to roughly 80–85% done on its own, while declaring victory early, staying blind to its own ugly visual output, and needing constant correction, with a twist showing how much stronger design conventions can improve the same tooling's results.
And there's a look at what it takes to run AI tooling responsibly inside an ISO-certified, compliance-conscious organization, from data sandboxing and local models to data processing agreements with vendors.
What you'll leave with
Expect a set of hard-won, practical takeaways, including:
- Start small: not a mammoth overnight, at least not on the first night
- Experience matters more than the tool
- The harness matters
- A failure from four months ago won't necessarily repeat itself today
- Models can get worse with a version increase
- Don't treat an LLM as a person, understand what it actually is
- Plan, then build
- Expect to spend money
New tools and models ship every week, and whatever was best six months ago has usually already been replaced. If you want the unfiltered version of what AI-assisted development is really like right now, mistakes included, this is the session to be in the room for.