What a hallucination is
A language model writes the text that is most likely to come next. When it does not have a fact, it can write a likely one. It does not warn you. The made-up answer has the same tone as a real one.
What it looks like in a takeoff
- A count for a sheet the model could not read.
- A model number that fits the manufacturer but is not in the spec.
- A requirement that is common in specs of this type, but is not in this spec.
- A bid date or a bond requirement with no page behind it.
- A row in a fixture schedule that does not exist.
Why it matters more in estimating
In a quantity list, an invented number looks the same as a real one. No one questions line 14 of 40. You find the error after you win the job. Then an invented item costs you margin, and a missed item costs you more.
How software can reduce it
- Tie every line to a quote or a location on a sheet.
- Let the model answer not sure, and do not force a number.
- Check that each quote is really in the document.
- Hold the job when a key document is missing.
How an estimator can catch it
- Open the source for each line that matters.
- Treat a line with no source as not verified.
- Spot-check a few lines that look sure against the sheets.
- Read bid dates, bonds and wage rules on the page yourself.
- Be more careful when the documents are poor scans.
Takeoff Tool links every line to the sheet or the spec text it came from. A line it cannot stand behind is graded Not sure.
AI and takeoff software · Updated 11 Nov 2026