AI Plan Check ROI: Signal vs Noise

For AEC teams piloting AI plan checkers. Puts a dollar value on the catches a reviewer wouldn't have caught manually, against the time spent triaging the noise to find them.

Total issues the tool surfaces on a typical drawing set.
70%
Share of findings worth acting on. Below ~70% reviewers quietly stop opening the report.
Read it, look it up in the drawings, decide.
Architect or engineer time. Loaded ~1.5–2× billable.
Average miss: RFI, rework, schedule slip. Tail misses (undersized septic, missed code citation) clear six figures plus a permit reset.
40%
Only the catches the team wouldn't have made on their own count as incremental value.
Across the firm. Scales the monthly numbers below.

Per month

Incremental catches surfaced—
False positives reviewers waded through—
Triage time spent—
Triage cost (reviewer hours)—
Value of incremental catches—
Net monthly ROI —

ROI multiple: —

How this is calculated
  • Incremental catches = findings × precision × manual-miss rate × projects/month. Only catches the team wouldn't have made manually count as net-new value.
  • Noise = findings × (1 − precision) × projects/month. Pure cost — reviewer time spent dismissing junk.
  • Triage time = findings × minutes per finding × projects/month, converted to hours. Applied to every finding, real or not, since the reviewer doesn't know which is which until they look.
  • Triage cost = triage hours × loaded reviewer rate.
  • Catch value = incremental catches × cost per missed issue. Anchors to the expensive-to-miss issues — coordination gaps, code citations, drawing-vs-spec conflicts — not RFI volume.
  • Net ROI = catch value − triage cost. Positive means the tool earns out the reviewer time it consumes.

The headline number is most sensitive to precision. A tool that finds 1,000 issues at 90% precision pays back fast. The same tool at 50% precision burns reviewer hours and erodes trust before it ever delivers a real catch. That's the signal-to-noise problem in one number.

If this was useful

I build small, focused tools like this in a day or two — usually for a specific founder making a specific argument. If this was useful, or you'd want a variant pointed at a different problem, follow along.

Follow on X  ·  Connect on LinkedIn