How AI reads a mechanical specification

A technical but accessible look at how document extraction works on real engineering documents. What the AI can do reliably (count equipment, match tags, classify types), what it cannot do (make judgement calls on scope), and why the human estimator is still essential.

What the AI does well

Modern AI models can read a 200-page mechanical specification and extract structured data from it. They can identify equipment tags (AHU-01, LTHW-P01, VRF-ODU-01), extract quantities from schedules, classify equipment by type, and cross-reference information across multiple documents.

This is not keyword matching. The AI understands context. It knows that "The AHU shall be provided with a BMS interface" means the AHU is in scope, while "The AHU BMS interface shall be by the mechanical contractor" means it is not.

Where it needs human judgement

Scope decisions on ambiguous items require experience the AI does not have. When a specification says "BMS monitoring of the domestic hot water system to be agreed at detailed design stage", a senior estimator knows whether to include it based on the type of project, the consulting engineer's typical approach, and what their company usually offers.

The AI flags these items for review rather than guessing. This is a deliberate design choice. An incorrect assumption that looks confident is worse than a flagged item that prompts a conversation.

How the pipeline works

The extraction pipeline runs in stages:

Classification — each document is identified as a specification, schedule, schematic, floor plan, or non-relevant document.
Extraction — equipment is extracted from each document type using approaches tuned to the format (structured tables vs narrative text vs visual schematics).
Correlation — information from multiple documents is merged into a unified equipment list, resolving duplicates and filling gaps.
Points generation — validated IO templates are applied to each piece of equipment to produce the points schedule.
Pricing — rates and labour calculations produce a priced proposal.

Each stage is deterministic where possible. The same documents produce the same output every time. Where AI judgement is involved, the results are constrained by templates and post-processing rules that enforce consistency.

The role of the estimator

The AI produces the first draft. The estimator's job shifts from building the schedule from scratch to reviewing a complete output, adjusting items that need their judgement, and making scope decisions on flagged items. Their expertise shapes every bid. Their time is no longer consumed by the mechanical work of reading and counting.