Natural Language Querying

Why it matters in construction

The questions a GC’s leadership asks every week are simple to say and annoying to answer. What is our hit rate on healthcare in the last two years? How much backlog burns off by June? Who is unassigned in Q2? Each one means opening a report builder and exporting to a spreadsheet, and by the time it comes back the meeting has moved on.

Natural language querying removes the report-building step. The person with the question asks it directly, and the system translates it into a query. It also lets people who would not open a query tool ask a question in a sentence.

How it works

1. Understand the question. The language model reads the plain-English question along with a description of the available data: what tables or objects exist, what fields they have, and what business terms like “weighted pipeline” or “utilization” mean in this firm.
2. Generate a query. The model writes a structured query (SQL or an equivalent) that expresses the question. It does not compute the answer itself.
3. Run it against the real data. The database executes the query and returns actual records and totals.
4. Present the result. The model formats the answer as a sentence, a table, or a chart, and states any assumptions it made, such as which date field it used or whether it filtered out lost opportunities.
5. Show the work. The user can see the query and the underlying rows, so they can check a surprising number.

The failure mode to watch for is silent misinterpretation. “Pipeline” can mean total, weighted, or active. “Q4” can mean calendar or fiscal. A good tool either asks for clarification or states which definition it used. A tool that does neither may be guessing.

Example in practice

In one hypothetical scenario, a commercial GC’s CFO is prepping for a Monday leadership meeting. Sunday evening she asks the firm’s precon platform three questions.

“What is weighted pipeline expected to award in Q4?” The system returns $38.2M across 11 opportunities, notes that it used the estimated award date and the stage-based probability, and lists the 11 with their values.

“How does that compare to the same point last year?” It returns $29.6M, so up about 29 percent.

“Do we have superintendents available for the two largest ones if both hit?” It shows both projects, their estimated start dates in November and January, and the three superintendents whose current assignments end before then, one of whom is already penciled to a backlog job.

Each answer links to the records behind it. The meeting can start with the staffing decision instead of with someone reading a spreadsheet aloud.