Back to Blog.

Too Good to Be True Costs the Whole Industry

Every builder has seen the number that's too good.

The sub who's way under everyone else. The schedule that only works if nothing goes wrong. You know how it ends. That's why the next low number gets a harder look.

Software is doing the same thing to construction right now. And the bill doesn't land on one buyer. It lands on everyone who comes after.

One bad pilot poisons the well

Construction runs on word of mouth. A project executive who gets burned by a tool doesn't write a review. They tell the next ten people at the next ten jobs. "We tried AI on drawings. Didn't work."

That sentence doesn't name a vendor. It names a whole category.

Our snake-oil post covered the pitch: upload the set, we catch everything, trust us. This one is about what that pitch costs after the demo. Over-promising doesn't just lose a deal. It trains an industry that already doesn't trust easily to stop listening. The honest tools pay for it too.

There's a simple line that fixes most of this. What is real today, and what is coming. Say which is which, out loud, with dates. Most of the damage happens when a vendor blurs the two.

The man behind the curtain

Silicon Valley has a name for one version of the blur. Designers call it a "Wizard of Oz" test: you fake the machine with a person behind the curtain to see if anyone wants the product before you build it. As a research trick, it's fine. Sold to paying customers as finished technology, it's something else.

There are public examples, and they aren't small.

  • Nate. In April 2025, federal prosecutors and the SEC charged the founder of a shopping app called Nate. They allege he raised more than $40 million by telling investors the app used AI to complete purchases, when contract workers in call centers in the Philippines and Romania were doing much of that work by hand. Those are charges, not a verdict. He has pleaded not guilty.
  • Presto Automation. In January 2025, the SEC found that this drive-thru company claimed its AI ordering product eliminated the need for human order-taking. Per the SEC, workers abroad were handling the vast majority of those orders. Presto settled without admitting or denying the findings.
  • Expensify. In 2017, customer receipts, names and addresses included, turned up as tasks for crowd workers on Amazon's Mechanical Turk. The company had long used human transcribers when its "SmartScan" couldn't read a receipt.

Now put your drawing set in that story.

A human-powered "AI" can look great on a ten-sheet demo. People are good at ten sheets. Then the 600-sheet set hits the week the bid is due, along with everyone else's, and the clock stops making sense. And somebody you never met may be looking at an owner's drawings you signed an NDA to protect.

Humans in the loop aren't the problem. Hiding them is.

To be fair to this category: some drawing-review tools say plainly that experts check the findings and a full review takes days. That's honest. If you know a person is in the loop, you can price it, schedule around it, and ask who they are.

The problem is never the human. It's the curtain.

What we've been finding

For the past two weeks we've been doing something simple. We signed up for the drawing-review tools in this category the way a buyer would, using research accounts, with no sales calls. We ran the same public sample set, seven residential sheets a city building department publishes, and kept a clock running. We ran MarkedUp the same way.

A few things are already clear.

Most of them won't let you try them. We tried 15 tools. Three gave us results without talking to anyone. The rest stopped us at a sales call, an invite, an approval queue, a payment screen, a work-email check or a desktop install. Some of that is reasonable. It also means most buyers see the demo long before the product.

Speed claims and real speed are different animals. Published turnaround claims in this category run from under a minute to two weeks for work that sounds the same on the website. On our small set, real results took anywhere from about four minutes to over an hour. Ours took about 17 minutes, and our progress bar sat at zero long enough to look broken. We're fixing that.

The page and the product don't always match. We saw a preview advertised as free ask for payment before it would start. We saw a report call the title blocks "consistent across 6 sheets" on a seven-sheet set.

What we have not found is proof that anyone in this space is hiding people behind the curtain. We're not claiming it. We're keeping the clock running, and we'll publish the full write-up, method included, when it's done.

Our line: real vs coming

If we're asking others to draw the line, here's ours.

Real today. MarkedUp checks the boring cross-references that turn into RFIs: sheet index against the sheets in the set, callouts against the details they point to, door tags against the door schedule, keynotes against the legend. It connects to Trimble Connect, Trimble ProjectSight, Autodesk Forma, Google Drive and SharePoint, and sends a marked-up PDF back into Bluebeam, Procore or Adobe. It's Grammarly for construction drawings: quiet until something will cost money, time or a claim.

Expanding. MEP checks. Some are running. More are coming. We won't call it full MEP coverage until it is.

Coming. One-Click RFI launches October 23. It turns a finding into a Procore RFI aimed at the party who has to fix it. It is not live today. Until it is, you get the marked-up PDF and your team writes the RFI.

Not us. We don't do clash detection, code review or design. Those are real jobs. Other tools do them.

That's the dam: catch the error upstream while it's still a markup. We'd rather undersell it and have you find out it's better than oversell it and have you tell ten people it wasn't.

How to tell before you sign

You don't need our research to protect yourself. Five questions:

  1. Run your set, not theirs. A live job you know well, with a clock running.
  2. Who touches my drawings, and where? People included. Get it in writing.
  3. What's live today, and what's on the roadmap? Dates, not "soon."
  4. Does it scale like software or like people? Time a 50-sheet set, then a 500-sheet set.
  5. What did it miss? If they only show hits, keep walking.

See it on a live set

Book a meeting. Bring a live set, ideally one you think is clean, and we'll run MarkedUp on it in front of you. If it holds up, the next step is a 90-day pilot on your real projects.

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

Sources

1. U.S. Attorney's Office, S.D.N.Y., indictment, United States v. Albert Saniger, 25 Cr. 157 (unsealed April 9, 2025). Allegations: Nate marketed as AI-powered; purchases at times manually completed by contractors in the Philippines and Romania. https://www.justice.gov/usao-sdny/media/1396131/dl?inline=

2. U.S. SEC, Litigation Release No. 26282, SEC v. Albert Saniger (April 9, 2025): alleged >$42M raised on misleading AI claims. https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26282

3. TechCrunch, "Fintech founder charged with fraud after 'AI' shopping app found to be powered by humans in the Philippines" (April 10, 2025). https://techcrunch.com/2025/04/10/fintech-founder-charged-with-fraud-after-ai-shopping-app-found-to-be-powered-by-humans-in-the-philippines/

4. MLex, "Former Nate CEO pleads not guilty to US fraud charges over AI claims" (Dec. 17, 2025). https://www.mlex.com/mlex/articles/2423455/former-nate-ceo-pleads-not-guilty-to-us-fraud-charges-over-ai-claims

5. U.S. SEC, "SEC Charges Restaurant-Technology Company Presto Automation for Misleading Statements About AI Product" (Jan. 14, 2025), and order, Release No. 33-11352. https://www.sec.gov/enforcement-litigation/administrative-proceedings/33-11352-s · https://www.sec.gov/files/litigation/admin/2025/33-11352.pdf

6. The Verge, "Expensify says receipts found on Amazon Turk were from testing a new feature" (Nov. 28, 2017). https://www.theverge.com/2017/11/28/16703962/expensify-receipts-amazon-turk-privacy-controversy · WIRED, "Expensify's Use of Amazon Mechanical Turk Reveals Privacy Risks Behind AI" (2017). https://www.wired.com/story/not-always-ai-that-sifts-through-sensitive-info-crowdsourced-labor/

7. City of Kirkland, WA, sample construction plan set (public, 7 sheets; the test set for our timed runs). https://www.kirklandwa.gov/files/sharedassets/public/v/1/development-services/pdfs/building-pdfs/sample-construction-plan-set.pdf

8. MarkedUp internal research, Sept. 26 – Oct. 8, 2026: public marketing pages of drawing-review tools (turnaround claims) and timed runs under research accounts on source 7. Unpublished; full write-up to follow.