Key Takeaways

  • Reducto CEO Adit argues white-collar work has not reached its agent milestone because tools stop at data extraction rather than editing live files.
  • Frontier language models have mastered long-horizon reasoning, making raw cognitive intelligence less of an obstacle than visual layout geometry.
  • Translating documents across formats requires spatial awareness, such as handling text overflow when switching from left-to-right English to right-to-left Arabic.
  • Reducto moved from general frontier models to specialized in-house systems like R1 to enable contract redlining and automated loan application filling.

The White-Collar Agent Gap

Software engineering had its breakout agent moments when models began modifying full codebases directly. Desk work has not arrived at that point yet. As Adit puts it: “My general view here is I don't think white collar work has quite had the equivalence of like um a quad code moment.” (play from 39:10)

Most enterprise pipelines run one way. An optical character recognition tool or vision model ingests a PDF, strips the text out, and dumps the raw content into a chat box. That solves retrieval, but real office tasks rarely end with reading.

Adit points out the missing half of the workflow: “At the end of a lot of these processes, you write into a spreadsheet or you write into a word doc, whatever it is.” (play from 39:37) A paralegal's final deliverable is a marked-up Word file with track changes. A loan officer finishes an evaluation only when the application contains verified data. Useful agents must write back into the source formats humans actually use.

The Bottleneck Shifted From Reasoning to Geometry

For two years, founders assumed white-collar automation waited on smarter models. Better reasoning, larger context windows, and fewer hallucinations were expected to unlock the knowledge economy.

Adit sees a different reality: “Models have gotten so good at long horizon reasoning that it feels less and less likely that intelligence is the biggest bottleneck.” (play from 44:54)

The primary point of failure is spatial. Consider document localization. If you feed an English contract into a frontier model, the linguistic translation into Arabic is fast and accurate. The breakdown happens when placing that Arabic back onto the physical page.

“Models are great at the translation,” (play from 40:33) Adit explains. “But applying that edit to a document is really hard because you're going to need to take a document that was generated in English and change the layout. Like when you change it to a right to left language, now you need to consider like maybe the text flows over onto the next page.” (play from 40:36)

A human designer knows how to adjust margins or carry an orphan line over to page four without obliterating a signature block. An LLM operating purely on text tokens cannot see page geometry. If an agent shifts a paragraph and pushes a signature box off the page, the output fails.

Moving Beyond Ingestion

Reducto started by parsing documents for finance and healthcare teams, where general frontier models silently missed tiny footnotes or misread cell coordinates. The company built specialized vision architectures like its R1 model to solve that extraction flaw.

Now, Reducto is directing those systems toward active document editing: “We want to make it possible for models to redline contracts as reliably as possible. Um, we want them to be able to fill out loan applications, do whatever a human would have had to do, right?” (play from 39:51)

Building an agent that edits contracts requires treating a document as a visual canvas governed by typography, alignment, and hierarchy. The companies that win in document automation will master layout manipulation down to the millimeter instead of relying on thin wrappers.

What to Do With This

Audit your team's document pipeline this week. Open your last five completed document workflows and note where manual human effort occurs. If your software extracts text from an agreement but forces an analyst to paste values into a spreadsheet or reformat a Word document, you have only automated the intake step. Pick one high-frequency form or contract template, map its visual layout constraints, and build an automated writeback pipeline that produces a finished, fully formatted file directly.