How Much Should an AI Project Cost? Margins Are Razor-Thin Right Now

One agent: the client sends over a development spec document and asks the AI to automatically extract component information, specs, and parameters from it, and output everything into Excel to feed the internal quoting process downstream.

How the project team described it: hard, and the margin is thin enough to be “gut-wrenching.”

A few technical sticking points, specifically:

1. The input is anything but standardized

File formats are never consistent: PDFs, Word docs, scans, annotated drawings, mixed Chinese and English, inconsistent units of measure — that’s just the baseline reality. The same parameter can go by completely different names across different files — “withstand voltage,” “rated voltage,” “Rated Voltage” are all the same thing. That means you can’t rely on fixed parsing rules; you need a model with real semantic understanding to handle every variant. The engineering effort this takes already exceeds what most requirement descriptions imply.

2. There’s almost zero room for error, and validation ends up costing more engineering than extraction itself

Once engineering spec data flows into procurement, the BOM, or the production line, a single parameter error can scrap materials or shut down a production run.

Extraction might just need a well-tuned prompt and a working model. Validation is a different animal entirely: you need a standalone review interface, a constantly updated rulebook of “which fields are prone to error,” and a design for how human confirmations feed back into the system. Put simply, extraction is a one-time build. Validation is a system you maintain indefinitely — and that’s where the real engineering load lives.

The result looks deceptively simple to build. Making that result reproducible at production grade is the hard part.