The real bottleneck in an Excel + LLM system is preserving business knowledge
Tired40s · reddit · 2026-07-28
The real bottleneck in an Excel + LLM query system is lost business knowledge
A procurement professional with no coding background spent six months building an Excel + LLM query system and found that the hard part was not natural-language understanding. The system repeatedly failed on a seemingly unambiguous question like “what’s the current price for this part,” even burning through 1.5 million tokens before the account ran out of balance.
The root cause turned out to be much more basic: the spreadsheet contained a cell comment explaining abbreviations in the “Order Type” column, but the pipeline compressed that comment into a boolean flag (hascomment: true). The model never saw the actual meaning, so it kept guessing from training data and violating the “don’t make things up without evidence” rule.
The author argues that most failures in agent/RAG systems come from the same place: human knowledge already exists, but there is no reliable, lossless path for it to reach the model. The real engineering work is preserving tacit business knowledge as it moves through the pipeline.
More from coding & agent
- Claude Opus 5 Generates 3D Neurovascular Simulator With a Single Prompt — BraydonDymm · 2026-07-28
- A PR daemon turns reviewer comments into fix PRs so humans only do the final review — MikkoH · 2026-07-28
- Resetting Claude Context: Developers Debate Memory vs. Clean Slates — mobileraj · 2026-07-28
- SAP’s TRACE preserves tool knowledge and reaches 86% recall with greedy decoding — SAP · 2026-07-28
- Reddit debate asks why coding agents still run planning and review on the same expensive model — Neat_Initiative_7780 · 2026-07-28
- GlobalGPT pitches a $10 AI workspace with 100+ models and MCP inside Codex — hey_abusiddik · 2026-07-28