This is a story about the limits of AI—and why human intervention isn’t optional, but essential.
Leo and I had been thinking about building Votomatic for over a year. The idea was inspired by Wahl-O-Mat, a German app that helps voters make informed decisions by comparing their views with those of political candidates.
We studied the Wahl-O-Mat process, which is remarkably well documented. We learned that creating effective voter guidance tools requires more than just automation—it demands a team of experts to interpret political information and craft meaningful questions.
It didn’t take long for us to have a working prototype of Votomatic: the UI was ready, the CMS for managing questions and candidates was in place. The only missing piece was the questions themselves.
With less than a week before the elections, I decided to experiment—just to see how far AI could take us.
The AI process
I started by downloading the official plans of the 35 parties so Claude could process them.
Then I built a pipeline orchestrator to:
- Run sub-agents in parallel to process each party’s plan, extracting proposals and storing them in JSON—keeping original statements intact for traceability and verification
- Classify those proposals into 15 predefined topics and determine each party’s stance within them
- Identify points of controversy across parties within each topic, based on their proposals and positions
- Generate neutral, yes-or-no questions grounded in those controversies
- Validate the neutrality of the questions and the accuracy of the mapped party positions
It took several iterations to get usable results—questions that weren’t redundant, that captured real points of disagreement, and that clearly differentiated party positions.
The Human Touch
I showed the app to Nathaly, a political scientist. Her immediate reaction: “A few topics are missing.” She was right.
When I asked the AI to double-check, it responded: “I don’t see controversy there.” But after manually reviewing the documents, I realized the issue wasn’t the absence of disagreement—it was how that disagreement was expressed.
The language in party plans, combined with the pipeline’s focus on explicit proposals, made certain controversies difficult to detect. They were there—but only if you knew how to look.
For example, the initial AI-generated question was:
“Should the government invest more in military spending?”
Based on party positions, there was little apparent controversy: two parties in favor, one against, and thirty-two neutral.
But when the question was reframed as:
“The State should prioritize the modernization and strengthening of the Armed Forces”
the divide became much clearer.
Another early realization: party plans alone weren’t enough. They tend to be broad and non-committal. As Wahl-O-Mat itself notes, you need to complement them with other sources—debates, speeches, interviews—to build a fuller picture and surface meaningful disagreements.
This is where AI struggles. Finding the right sources, interpreting context, and verifying information all require significant human involvement.
The work demands distinct expertise:
- A political scientist understands the political landscape and ideological nuances
- A journalist knows how to find and validate the right sources
- The public helps surface what is truly controversial
- A software developer translates all of this into structured processes—defining the rules and constraints that allow AI to perform effectively
The Result
AI proved incredibly useful for processing large volumes of information quickly. It was instrumental in generating an initial draft of questions and structuring the data.
But without human input, the outcome would have been shallow and misleading. The questions would have missed key political divides, misrepresented party positions, and ultimately failed voters.
This experience reinforces a critical insight: AI is powerful—but it works best as a tool that amplifies human expertise, not as a replacement for it. The most effective systems aren’t purely automated. They’re hybrid, combining AI’s speed and scale with human judgment, domain knowledge, and contextual understanding.
In our case, AI handled the heavy lifting. Humans provided the wisdom.
Read more