Decoding Bureaucracy with AI
00
How we used AI to automate complex legal classification to scale onboarding.
Problem
When entrepreneurs incorporate a company in Spain and France, they must select a specific activity code. These codes are rigid, legalistic, and confusing. Our users know what they do (e.g., "I drive an Uber"), but they don't know how to translate that into government syntax.
The User Pain: We noticed a 80% drop-off at this step in the registration flow.
The Business Cost: Because users provided vague descriptions, our internal operations team had to manually review and correct 100% of files (!!!), creating a massive bottleneck in the company creation process. This single screen was the cause of more than half of the contacts reasons.
Solution
From Rigid Forms to Fluid Conversation : We iterated through three distinct process models before abandoning static fields entirely. The breakthrough came with Candosa AI, a conversational interface that focuses on intent rather than syntax. By decoupling the user's input from the system's output, we allowed users to speak naturally. The AI analyzes their description, asks clarifying questions if the confidence score is low, and maps the intent to a precise legal code.
Impact:
Error Reduction: Drastically reduced the number of files requiring manual agent intervention.
Scalability: Allowed agents to support more files without additional workload.
User Sentiment: Turned the most intimidating step of incorporation into a magical "aha" moment.
Phase 1: The "Open-Ended" Approach (France)
We didn't arrive at the AI solution immediately. I led the team through three distinct iterations to understand user behavior.
Initially, we assumed users just needed a push. We used open text fields with automatic templates and strict character limits to force detailed descriptions. We also had struggles fitting all the important information inside the screen, between disclaimer, FAQ, description and text field, it was hard to define a clear priorization that would make this step look digestible.
The Insight: High friction doesn’t yield high quality. Users felt overwhelmed by the "blank canvas" and often pasted generic text just to bypass the validation.
Phase 2: The "Taxonomy" Method (France)
To reduce writing, we introduced a visual category selector, we structured the data into Categories and Sub-categories with icons to make it easily digestible.
The Insight: While visually cleaner, this introduced "Decision Paralysis." Users struggled to fit their unique jobs into generic buckets (e.g., is a personal trainer "Health" or "Services"?), leading to misclassification and users picking the "Other" category just to skip the step, bringing us back to the original problem..
Phase 3: The Conversational Pivot (The Solution)
To launch in Spain, we pivoted from a Form-based model to a Conversation-based model. We realized the problem wasn't the UI components, but the cognitive load. The solution was to remove the burden of choice entirely.
The Insight: "Users don't know the legal code, but they know how to describe their day-to-day work."
The Fix: We stopped trying to make them fill a form and built a conversation using LLMs (Candosa AI).
Natural Input: The user speaks naturally ("I want to be a VTC driver").
Contextual Loops: If the input is vague, the system doesn't guess. It triggers a clarification flow to ask for specifics.
Smart Suggestion: The AI maps the intent to the exact legal code, which the user simply confirms.
Designing for AI required mapping "Confidence Thresholds" rather than linear paths. We created flows where the system acts as a guide:
Final Design : Candosa AI
The final design prioritizes trust. We utilized a chat interface because it is a familiar pattern for "asking for help." By allowing users to verify the AI's suggestion rather than generating the data themselves, we turned the most intimidating step of incorporation into a magical "aha" moment, proving that complex bureaucracy can feel simple with the right abstraction layer.
see also
