Live production proof

Real answers, timed and cited — including the one it refused to guess

Marketing pages tell you what a tool is supposed to do. This page shows what ours actually did, on the live system, timestamps and all — the cited answers and the honest "the documents don't say."

Production examples from June 2026 · lightly redacted for project identity · answers unedited

56 secquestion emailed → cited answer delivered (live timing)
Same minutenew PDF attached → indexed → question answered
100%of answers cite their source — or decline
0guesses. Cite-or-escalate, every time

Example 1 — a dimension question, answered and cited

Emailed to the project intake address:
"What is the loading dock leveler pit depth?"
✓ Answered, cited
"The loading dock leveler pit depth is 11 inches, measured from finished floor [RFI-2026-042, Sheet T-1, Rev A]."

One sentence, one number, one checkable source — sheet and revision included. Anyone on the job can verify it in fifteen seconds. That's the standard every answer is held to: no citation, no claim.

Example 2 — the one it refused to guess

Emailed with an addendum PDF attached:
"Per the attached Addendum A3, what is the fire rating of mechanical room door M-101? Please cite the source."
⚠ Honest decline
"The sources don't specify the fire rating of door M-101…" — followed by which document would resolve it.

This is the answer we're proudest of. A generic chatbot would have produced a confident, plausible, invented fire rating — on a life-safety question. IntelMS read the document, found the rating genuinely absent, and said so. A missing-document answer is a useful answer; a guessed one is a hazard. More on that: what's safe to use AI for on construction documents.

Example 3 — revision-aware by default

Emailed to the project intake address:
"What is the latest RFI log revision and its date? Please cite the source."
✓ Answered, cited — 56 seconds
"The latest RFI log with an explicit revision and date is the one marked Revision A, dated 2026-06-08…" — with the source document named.

Asked at 2:41pm, answered at 2:42pm. The system also distinguishes a document's stated revision from its file metadata — the kind of trap that catches humans working fast. Why that matters: the real cost of building from a superseded revision.

Why we publish declines, not just wins: any AI demo can show a right answer. The trustworthy ones are defined by what they do when the documents are silent. Cite-or-escalate isn't a tagline — it's the production behavior, shown above, unedited.

The guardrails behind these answers

Every answer above passed the same gate: grounded only in the project's uploaded documents, latest revision, exact source cited, conflicts flagged instead of silently resolved — and anything contractual, cost, code, or safety-critical routed to a human. A person approves replies; the system just makes them fast and defensible. The full record of every Q&A lands in a per-project timeline for closeout and dispute defense.

Run this test on your own project

14-day pilot on one real project. Email your documents in, ask the questions you argued about last week, and judge the citations yourself.

Start a free pilot