Grounding and Citations
Grounding ties an AI answer to specific source material rather than the model's general knowledge. Citations show where each claim came from. In construction, that might mean pointing to Section 09 29 00, paragraph 3.4, for a level 5 finish requirement.
Why it matters in construction
Construction documents are contractual. If an estimator carries a scope item because an AI said the spec required it, and the spec did not, that is real money. If a project engineer tells an owner the warranty on the RTUs is five years because the model said so, and the O&M manual says one, that is a warranty claim the GC eats.
Grounding limits the model to the documents provided for the task. Citations make its work checkable. Together, they give an estimator a way to review an answer before relying on it.
How it works
- Retrieve the relevant source. Before the model answers, the system pulls the passages most likely to contain the answer from the spec, the RFP, the sub proposals, or the closeout package. This is the retrieval step in RAG.
- Instruct the model to answer only from those passages. The prompt tells the model that if the answer is not in the provided material, it should say so rather than guess.
- Attach references. For each factual claim, the output identifies the document and location, and may quote the source text.
- Link to the citation. The interface should let the reader open the relevant page or section.
The citation needs enough detail to be useful. “The spec” is not enough on a 400-page book; a section and paragraph let an estimator find the relevant text. The system also needs to handle gaps honestly. “Not found in documents” is a valid answer when the source material does not contain the answer.
Grounding does not eliminate errors. A model can still misread a passage, but a reviewer can see the source and catch the mistake.
Example in practice
Let’s say a commercial GC is bidding a $31M K-12 addition. The estimator asks the firm’s AI plan and spec tool whether the project requires seismic bracing on suspended ceilings. The tool answers yes, cites Section 09 51 13, paragraph 2.6, and quotes the two sentences requiring bracing per the referenced ASCE standard. It also notes that Structural sheet S-001 general note 14 references the same requirement.
The estimator opens both references, confirms the requirement, and adds bracing to the ACT scope in the bid package. When three subs exclude it in their proposals, the team can point to the cited requirement and resolve the gap before award.
Frequently asked questions
Does a citation guarantee the answer is correct?
No. A model can cite a passage and still misread it. A citation narrows the review to the relevant passage instead of the whole document. Treat it as a review aid, not proof.
What should a good citation look like in a construction tool?
It should include the document name, page or section number, and quoted passage. Ideally, a link opens the source at that location. A document name alone is not enough for a 400-page spec book.
Can the AI answer without a source?
A well-designed grounded system should say it could not find the answer in the provided documents rather than fill the gap from general knowledge. If a tool never says 'not found,' be suspicious.
Go deeper
- From the blog AI for Estimators: 5 Practical Uses that Work Today (Not Someday) Construction's drowning in AI hype, but here are five use cases AI can actually do for estimators today, if it has access to your preconstruction data instead of just whatever's on the internet.
- From the blog AI and Construction: The GC's No BS Guide to What Works in 2026 A founder's reality check on where AI actually delivers value in preconstruction, and where it still falls short.