AI Deal Create starts with a straightforward idea: user uploads a document, AI extracts the deal data, user reviews and submits. The natural design model is "complete a form." It held up long enough to drive the initial layout, the field-level decisions, the first developer walkthrough.

Then the edge cases started compounding. Locked fields. Partial tables where users could remove a row but not add one. Tabs that weren't truly required but weren't clearly optional either. Each was solvable in isolation — a tooltip here, a confirmation modal there. But the fixes kept producing new friction. After enough of them, the pattern became clear: we weren't solving edge cases. We were patching a mental model that didn't fit how AI extraction actually behaves.

AI extraction isn't neat — it's confident about some things, uncertain about others, and sometimes wrong. A form model assumes completeness. It penalizes partial confidence and forces users to commit to data they don't fully trust. That's the wrong contract for this kind of workflow.

That reframe changed the entire interaction architecture. Instead of mandatory tabs requiring resolution, each tab became an optional package — extracted if the AI found the data, included by default, but skippable if the user doesn't trust the extraction or wants to handle it in the deal system directly. The Deal tab stays required. Everything else is a judgment call based on confidence in what the AI found. The compounding edge cases didn't need individual fixes. They dissolved.

I sent the product lead a message framing this before touching a single screen. He validated it in one reply. That's how foundational the reframe was — it didn't require a prototype to land.

AI played a real but bounded role in getting there. A Replit conversation helped surface the structural problem with remove-only tables — it articulated why partial editability is worse than no editability, which sharpened the argument for the reframe. For duplicate detection, I used Replit to rapidly explore 8 UI patterns — then rejected its top recommendation and synthesized a solution it hadn't suggested. Figma stayed the artifact developers built from. Replit was where the reasoning happened.

The Scale tab showed the limit. Two separate validation problems, two weeks of AI back-and-forth — both Replit and Claude kept hallucinating. One sync with product and devs resolved it. Knowing when to stop prompting and go talk to people is part of using AI well.

While Munis was working through this, a strategic divergence was growing between asset classes. Equities and Fixed Income had a grander North Star vision; Munis had devs building in a week. Rather than wait for alignment that wasn't coming, I proposed a phased v1-to-v3 framing — Munis builds now, architected to scale toward the shared vision rather than box it out. The product lead bought in. Munis moved.

Because the work was good and shipped when the others were still aligning, Fixed Income leadership asked to apply the same approach across all three asset classes. Munis went from the lagging asset class to setting the direction — not through a coordinated strategy, but by making good local decisions and moving.

The most consequential design decision on this project happened before any screen was finalized — recognizing that weeks of compounding problems were a diagnostic signal, not a to-do list. The mental model reframe didn't show up in the Figma file. But it's what made the rest of the work possible.