There is no AI in this tool

Prephase does not use AI to fill in a specification, guess a model number, read a quote, or suggest a product. Every value in the record was entered by a person or matched against a list that a person maintains.

This is not a position about the technology. It is a consequence of what a specification record has to do.

The record has to be exactly right

A model number with one digit wrong is a different product. Not an approximation of the right product. A different one, that ships, gets installed, and comes back out.

Deterministic entry means the same input produces the same output every time, and when something is wrong there is a reason that can be traced. A brand came from a list. A finish came from the finishes that brand offers in that category. A model number was typed by somebody looking at the vendor page.

Inference produces a value that is usually right. Usually right is a reasonable standard for a draft. It is not a reasonable standard for the number an installer orders from.

Wrong output does not look wrong

Construction already runs on estimates. Contingency, allowances, weather buffers, productivity factors. Nobody treats those as certain, and nobody has to be told, because each one announces its own uncertainty in its name.

Generated output does not do that. It arrives in the same register as the measured facts, in the same font, with no marking. Checking it requires already knowing the answer, which means the people most exposed are the ones with the least ability to catch it.

Once an inferred value is in the record, there is no reliable way to keep it labeled. It becomes indistinguishable from what was verified.

There is no seat in the chain

Construction has an established structure for who is responsible when something is specified wrong. An architect stamps drawings. An engineer stamps calculations. A general contractor carries the work. A subcontractor carries the install. Each one has a duty and insurance behind it.

Generated output has no seat in that structure. When the wrong thing is installed because a tool suggested it, the question of who carries it is unsettled, and unsettled liability in an industry that runs on indemnification is a problem worth avoiding entirely.

There is nothing upstream to read

Extraction works where a machine-readable document already exists as a byproduct of something else. A measurement report, a price feed, a loss summary. The document was going to be created regardless, so reading it is free.

A specification is not that. Somebody chooses the faucet. There is no upstream artifact to harvest, and the decision that produced the entry happened in a person's head.

Where it would fit

There is one shape where inference is a reasonable tool: reading a document and flagging language for a person to look at. The output there is a prompt to check something, not a value in the record. Being wrong costs a few seconds of attention.

That is a different job from specifying what gets installed. Being wrong there costs a demo, a reorder, and a delay, and somebody writes the check.

The distinction is not whether the technology is capable. It is whether being wrong is detectable and cheap.