Back to Blog.

The Drawing-Check Gold Rush Has a Snake-Oil Problem

Everyone is checking drawings now.

markedup.ai, Callout, Helonic, Structured AI, Flikt, InspectMind, Bluebeam Max, Groundbook, PlanCheckPro, Archidian, Buildcheck, Aginera, Nomic, Permit Analyzer, CodeComply, CivCheck — and Autodesk Forma Drawing Compliance Review, planned for early 2027, not shipped.

The rush is real. Document defects drive a huge share of RFIs. Senior reviewers are scarce. OCR got cheap. Bluebeam shipped Max. Autodesk put compliance on the roadmap. Capital followed.

None of that means the pitch is true.

The pitch

Upload the set. We catch everything. Code, coordination, constructability. Seconds. Double-digit ROI. Trust us.

Snake oil with a progress bar.

Who actually names the brain

Trust marketing beats stack honesty in this category.

Named (from primary security/trust pages): Callout Anthropic. Flikt → Anthropic only. InspectMind → OpenAI / Anthropic / Google. Nomic → Anthropic + Vertex (ZDR claims) + Modal. Bluebeam Max → CV + LLM + Claude via MCP; Smart Review still Preview, PDFs on AWS UK. Aginera → OpenRouter. Procore AI → Azure OpenAI.

Fog / silent: Groundbook; PlanCheckPro (“third-party AI”); Archidian (unnamed “zero-retention” providers); Helonic (“vetted third parties”); Structured (proprietary pipeline, no frontier names); Buildcheck (vision-first marketing, no LLM routing on the FAQ).

Working hypothesis: a large share of the field is OCR + multimodal LLM + retrieval + prompts + accept/reject. Some teams have real detectors. Many have a harness and a homepage.

A harness is not a moat. An agent loop is not a moat. When the model improves — or Forma ships the button — the loop is a menu item.

Your drawings leave the building

Named Anthropic/OpenAI/Google/OpenRouter paths mean the sheet hits someone else’s GPU. “Zero retention” and “no training” are contracts, not physics — and they are uneven.

Harder no-train claims: Structured, Archidian, Callout/Nomic ZDR language, parts of Helonic policy. Vendor may improve on your docs or anonymized feedback: InspectMind (own models), Flikt internals, Buildcheck anonymized improvement. Explicit train-on-customer with opt-out buried in terms: Aginera.

If they will not answer model, region, train, retain, subprocessors — in writing — do not send the bid set.

Accuracy theater

This is where snake oil lives.

InspectMind is rare: roughly 70–80% actionable / 2030% false positives on their own trust pages. That is honesty. Helonic “90%+” and CivCheck “98%+” — treat as marketing until methodology shows up. PlanCheckPro terms: informational, not a stamp, not AHJ approval. Callout: PE must verify. Bluebeam Smart Review: Preview, documented false alerts. Independent constraint-check research still shows false-pass rates that make autonomous approval reckless.

Every serious product eventually says: not a PE, not a certification, human must verify. Correct. Also a tell. If a person re-checks everything, you bought a noisy highlighter.

Ask for the miss list. If the demo only shows hits, keep walking.

Cost and latency

A four-sheet demo is not a 400-sheet IFC. Token pipelines get expensive when they retry. Per-sheet fees look cheap until you multiply by revisions and hours dismissing junk.

Who authors vs who bleeds

Architects and engineers draw under compressed fees. GCs and owners eat RFIs and claims. That handoff is why the category exists — and why “we catch everything” sells to exhausted buyers.

VCs are sorting

Buildcheck: $5.9M → $12M A, 110+ customers, named GCs. Structured: $4.2M, design-side. Nomic: Aurecon/Arcadis strategic (dollars undisclosed). Bluebeam: distribution. Most peers: early or quiet.

Money still moves. It is getting pickier about unnamed “AI” that promises the world.

Demand this before you trust anyone

  1. Named model path and subprocessors — in writing.
  2. 2. False-pass / false-fail on a set like yours — not a highlight reel.
  3. 3. What they cannot check.
  4. 4. Full-package cost and time, including human cleanup.
  5. 5. Proof vs vibes — citations, sheet refs, residual risk a human owns.

If the answer is “trust us, we do everything,” keep your drawings.

markedup.ai

We will not pretend a model approved your set.

Grammarly for construction drawings — an audit layer. ConstructionProof returns markups into tools you already use. Typed checks on boring RFI factories: sheet index, callouts, door schedules, keynotes, expanding MEP. ML/CV where a chat harness is the wrong hammer. Proof of what was checked. Human still on the hook.

If frontier LLMs eat thin cockpits, good. If we cannot beat a rented harness on the checks we claim, we should lose too.

The handoff is real. The gold rush is real. The “we catch everything, trust us” line is the part that needs a redline.