The Errors That Cost the Most Are the Boring Ones

Nobody brags about catching a door schedule mismatch.

Clash detection gets the demo. A duct through a beam, lit up red in a model, is the kind of thing people screenshot. A door tag on A-201 pointing at a schedule row that doesn't exist is the kind of thing people skip.

The skipped one keeps showing up in the data.

What 40,401 issues tell us

In September, InspectMind published its 2026 Construction Drawing Error Index. It's useful vendor research, and it's built with care: 521 published, anonymized plan reviews, 90,768 project-level findings, and 40,401 individual issue records structured well enough to analyze in public. They publish the aggregate data as CSV and JSON. They explain how they counted. That's more than most numbers in this industry get.

One caveat up front, and InspectMind states it themselves: this is a sample of projects reviewed with InspectMind, not a random sample of the construction industry. Project mix, set size, design stage and review depth all vary, and the sample leans commercial. Read it as a large, honest look at real sets. Not an industry census.

With that said, three things stand out.

Most of what turns up isn't minor. Of the 40,401 records, 15.5% were tagged Critical and 52.1% High. That's 67.6% combined. InspectMind is careful here too: severity means review priority, not that every finding was accepted unchanged by the project team. Still, two out of three flagged items were rated worth chasing first.

The top recurring pattern is references. InspectMind grouped issue titles by matching terms. The largest group was "Specifications, schedules & references": 6,918 titles, 17.1%. Those are titles that mention specs, schedules, callouts, drawing indexes, title blocks, or broken sheet and detail references. "Dimensions & geometry" came in a hair behind at 6,888 (17%). Then "Missing information" at 8.7%, "MEP systems" at 6.9%, "Codes & standards" at 5.4%, "Fire & life safety" at 4.6%, and "ADA & accessibility" at 1.5%. One title can match more than one group, so these don't add up to anything.

The biggest bucket belongs to no single discipline. Sorted by source category, "General & coordination" was 42.6% of records. That's more than architectural (20.8%) and structural (16.7%) combined. InspectMind kept it separate on purpose instead of guessing a discipline for cross-sheet findings. Put plainly: the largest share of problems lives between the sheets, not on any one of them.

Boring isn't the same as cheap

None of those figures are dollars. The index counts findings and severity, not cost. So here's the argument, and where each piece comes from.

A broken reference is cheap to fix on paper and expensive to discover anywhere else. When a callout points to a detail that isn't there, or the spec says one thing and the schedule says another, the field doesn't guess. It writes an RFI.

In August, Flikt published a root-cause study of 827 RFIs from three multifamily projects. 60.2% traced back to the documents themselves: 32.2% missing information, 18.5% documents that conflicted with each other (sheet against sheet, plan against schedule, drawing against spec), and 9.3% ambiguous. It's three projects of one building type, and it's vendor research. Flikt says so. But the shape matches.

Every one of those RFIs has a floor cost. Navigant's 2013 study put the average cost to receive, review and answer a single RFI at about $1,080. That number came from interviews with three firms, in 2013 dollars, and it covers the paperwork only. Not the lost day, the change order, or the rework. It's the cheapest layer of the bill, not the bill.

Now put that next to volume. The median project in InspectMind's data carried 88 findings. Not every finding becomes an RFI. Some do. The boring ones do it quietly, a few at a time, all job long.

That's how a dull error gets expensive. Not one big hit. A steady drip nobody plans for, because nobody saw it coming.

Why coordination doesn't catch it

3D coordination did real work. Clash detection catches geometry that used to walk onto site. That's why it gets the attention.

But clash checks shapes against shapes. It doesn't check whether door tag 214 has a row in the schedule. It doesn't check whether every sheet in the index is actually in the set. It doesn't check whether detail 5 on A-501 exists, or whether keynote 12 on the plan matches keynote 12 in the legend. Those links live in the 2D set people actually build from.

Checking them is a census, not a sample. Every tag, every callout, every index line, every sheet. On a large set that's thousands of cross-references. Nobody has those hours, so it gets spot-checked. Spot-check a census and you find some of the problems and ship the rest.

That's not a talent problem. It's arithmetic.

Where markedup.ai sits

markedup.ai runs that census.

Sheet index against sheets. Callouts against details. Door tags against schedules. Keynotes against the notes on the plan. MEP checks are expanding. We don't do clash detection, code review, or design. Those jobs are real, and other tools do them.

It works inside the tools you already use. Point us at a folder in Trimble Connect, Trimble ProjectSight, Autodesk Forma, Google Drive or SharePoint, and we check sets as they land. You get a marked-up PDF back that opens in Bluebeam, Procore, Adobe, or wherever your team already marks up drawings. Nothing new to learn. No new place to keep drawings.

Think of it as Grammarly for construction drawings. It stays quiet while the set holds. It speaks up only when it finds something that will cost money, time, or a claim.

When it does speak up, you aren't just handed more information. We find the error, show what it puts at risk, and move it to the party who has to fix it with a One-Click RFI. For a GC, that matters. You don't own the design. But you can end up owning the risk of a design error you saw and didn't raise. A dated, documented notice moves it back where it belongs.

That's the dam. Catch it upstream, while it's still a markup, before it turns into rework, a lost week, or a claim.

Try it on a live set

You don't need an integration to see this. Drop a PDF set into ConstructionProof and get a marked-up PDF back. Pick a live job, ideally one you think is clean.

If we find nothing, good. The set held. If we find something, you found it on paper, which is the cheapest place it will ever be.

Sources

1. InspectMind, 2026 Construction Drawing Error Index (published 2026-09-16; accessed 2026-10-06). 521 published case-study records, 90,768 project-level findings, 40,401 structured public issue records; severity, source-category and title-cluster breakdowns; limitations stated on page (InspectMind-reviewed projects, not a random industry sample). https://www.inspectmind.ai/research/construction-drawing-error-index · Aggregate data (incl. median 88 findings per project): https://www.inspectmind.ai/research/construction-drawing-error-index-2026.json · https://www.inspectmind.ai/research/construction-drawing-error-index-2026.csv

2. Flikt.AI, "Where RFIs Actually Come From: A Root-Cause Study of 827 RFIs" (2026-08-30). 60.2% document-caused (32.2% omission, 18.5% conflict, 9.3% ambiguity); three multifamily projects, first-party, LLM classification with a 100-RFI human spot check. https://flikt.ai/rfi-root-cause-study/

3. Navigant Construction Forum, Impact & Control of RFIs on Construction Projects (April 2013). ~$1,080 average cost to review and respond to an RFI, built from interviews with three firms; 2012–13 dollars; processing cost only. https://www.cmaanet.org/sites/default/files/resource/Impact%20%26%20Control%20of%20RFIs%20on%20Construction%20Projects.pdf