Auditors at Brazil's Federal Court of Accounts were spending months writing Audit and Merit Instructions: important legal documents sent to the Public Ministry. The Court wanted AI to speed this up.

But there was a bigger issue than speed: the auditors did not trust the AI.

These were senior experts who had done this work manually for decades. They were constitutionally skeptical. They would not sign their name to something they could not fully verify.

The challenge was not making the AI smarter. It was making the output trustworthy enough for guarded experts to actually use it.

I spent time observing how the auditors actually worked. I broke their process into small mental steps: what they looked at first, what made them pause, and what made them trust or reject a draft.

That is when I realized the biggest trust problem was not the AI model. It was the messy, inconsistent inputs the system was receiving. Different teams used different templates, so the AI produced unpredictable results.

My key decision: instead of trying to fix the output, I went upstream and reshaped the inputs.

We consolidated dozens of legacy templates into one clean, consistent pattern. The AI immediately started producing more predictable and accurate drafts. That single change became the biggest trust lever in the entire project.

Once the inputs were cleaner, I focused on giving auditors control and visibility. Every major decision made the AI easier to inspect, edit, and correct.

During a usability test, a notoriously picky senior Secretariat Director read an AI-generated draft. He paused, re-read it twice, then looked up and said:

This could turn months of work into hours.

That moment, watching a deeply skeptical expert choose to trust the output, became the foundation of how I think about AI design.

63% reduction in instruction-writing time. Work that took months could be completed in hours.

The system shipped into production and was adopted across multiple teams.

The work was praised by ministries and won Best technology case in the public sector in Brazil.

This project taught me that trust is an architecture, not a feature. You have to design it into the inputs, the outputs, and the correction surfaces from the beginning.

The same core pattern I used here, reshape inputs and give users the right to correct, is exactly what I later applied to LLM and GenUI work on Ask Markerr and TaleWeaver.