What Is AI Plan Review? The Definition, the Data, and the Limits
What Is AI Plan Review? The Definition, the Data, and the Limits
AI plan review is easiest to understand by starting with a number no vendor puts on their homepage. In New York City, the Department of Buildings completes a first plan review in 3.7 business days on average. Getting from filing to approval takes 20.3 days. The reviewer is not the bottleneck. The correction cycle is - the loop of objections, revisions, and resubmissions that runs between the first look and the stamp, largely on the applicant's clock.
AI plan review is the use of AI systems to read construction documents - drawings, specifications, and schedules - and flag errors, conflicts, and code issues before a human reviewer or a building department finds them. Where traditional review means people paging through sheets and clash detection means comparing 3D model geometry, AI plan review reads the documents themselves: it can hold a structural drawing, an architectural sheet, and a spec section in view at once and mark the places where they disagree. This guide covers what the term actually means, what the public data says about the problem it exists to solve, what two decades of research found, what the tools reliably do today, and the questions worth asking any vendor before trusting one with your set.
What is AI plan review?
The term covers two different rooms, and it pays to know which one you are standing in.
- Submitter-side review. Design firms, GCs, and owner's reps run AI across their own document sets before bid or filing - catching cross-discipline conflicts, spec-to-drawing contradictions, missing schedules, and code red flags while they are still cheap edits.
- Jurisdiction-side plan check. Building departments run AI on incoming permit submissions to screen for completeness and prescriptive code compliance before or alongside a human plans examiner.
Same underlying technology, very different stakes. A submitter-side miss costs you a correction cycle. A jurisdiction-side miss is an enforcement question with legal weight behind it.
AI plan review is also not the same thing as the tools most firms already run. Clash detection compares model geometry and needs a coordinated BIM model to exist. Code Q&A tools answer questions about code text. AI plan review works on the documents you actually submit - the PDF set - which is why it applies to the majority of projects where a fully coordinated model never exists.
Why plan review became the bottleneck
The delay problem is public record, published by the departments themselves.
- Seattle posts its permit performance live: middle housing reviews are running 117 days against a 60-day goal, and large multifamily 374 days against a 180-day goal. SDCI itself notes the total time an applicant experiences is roughly twice the days in city control, because most permits go through several correction cycles.
- New York approves in 20.3 days against a 3.7-day first review, per the state comptroller's June 2026 report - which also notes the DOB does not publish rejection rates or the average number of resubmissions at all.
- San Francisco housing projects approved in 2022 saw median departmental review of 289 days at Planning and 259 days at Building Inspection.
- Los Angeles officially tells applicants a regular plan check takes 15 to 30 business days - per cycle.
Seattle SDCI, on its own dashboard: the time an applicant experiences is about twice the time the city controls, because most permits take several correction rounds.
Two structural facts sit underneath those numbers. First, capacity: a national ICC survey a decade ago projected 80% of code officials would retire within 15 years - a window now closing - with only about 3% of the profession under 35, and the Bureau of Labor Statistics projects inspector employment to be flat-to-declining through 2034. Reviewer supply is not coming back. Second, plumbing: a 2025 Washington State survey of 181 jurisdictions found only 42.5% run digital-only permitting - 12.7% still work entirely on paper - and replacing those systems costs anywhere from $65,000 for a small jurisdiction to $4 million for a large one.
The money side is just as documented. NAHB's 2026 study puts regulation at $131,734 of the cost of an average new single-family home - 26.4% of the price - and in its developer surveys, 95.9% report regulatory compliance adds a typical delay of about six months. Time in the correction loop is not administrative friction. It is carrying cost, crew idle time, and financing burn.
What twenty years of research actually says
AI plan review did not appear with ChatGPT. The academic field is called automated code compliance checking, and its founding paper - Eastman and colleagues, 2009, in Automation in Construction - laid out the four stages every tool still follows: interpret the rules, prepare the building data, execute the checks, report the findings.
What happened next is the interesting part. A 2020 study in the same journal concluded that despite decades of research and industry enthusiasm there had been "no meaningful adoption" of automated compliance checking - and that the barriers were commercial and political rather than technical. The rules were hard to encode, jurisdictions adopt amended code editions on their own timelines, and nobody owned the problem of keeping rule sets current.
Large language models changed the economics of that first stage. Research in 2024 showed LLMs can translate building regulations into machine-checkable rules from just a few examples - with the caveat that their output needs careful contextualization to be reliable. A 2026 study from the National University of Singapore was blunter: LLM-based compliance systems hallucinate machine-readable rules outright, and even with reinforcement-learning fine-tuning, error rates drop by about a quarter rather than to zero.
That is the honest state of the science: the rule-translation bottleneck that stalled the field for a decade is opening, and the systems still hallucinate enough that unsupervised use is not defensible. Any vendor claiming otherwise is ahead of the literature.
What AI plan review is good at today - and where it fails
Practitioners who have actually tested these tools draw a consistent line.
Reliably strong:
- Completeness and intake screening. Is every referenced sheet, schedule, and detail actually in the set? This is checklist work, and machines do not fatigue on sheet 400.
- Text, notes, and schedules. Reading door schedules, spec sections, and general notes across hundreds of pages and cross-referencing them is where document AI already outperforms a skimming human.
- Contradiction detection. Two sheets that disagree, a spec section that conflicts with a drawing note, a referenced detail that does not exist - measurable, checkable, catchable.
- Code research with citations. Finding where a provision lives and quoting it accurately.
Still unreliable:
- Linework and geometry. Field tests by reviewers report tools that read schedules well but misread drawn conditions - one architect found a purpose-built tool got three of eight door swings wrong and invented room numbers that were not on the plans.
- Multi-step code paths. Real compliance runs scope, then exception, then referenced standard, then local amendment. Tools that treat the base code as universal get jurisdictions wrong, because the local amendments are the code.
- Judgment. Alternative means and methods, design intent, constructability - the reviewer's actual profession.
That split explains the trust data. Autodesk's 2025 industry survey found trust in AI fell eleven points year over year even as adoption grew, and the AIA found a third of firms use AI day-to-day while only 8% have implemented it into practice. The industry is experimenting with one hand and hedging with the other, which is exactly what the capability profile above deserves.
The questions vendor pages do not answer
Spend an evening on the practitioner forums where plans examiners and building officials actually discuss these tools and you will find the same questions asked again and again - and rarely answered.
- Who is liable when the AI misses something? As of mid-2026, no published court decision has tested liability for a deficiency an AI review failed to catch, and procurement contracts rarely address indemnification or audit rights. "A qualified human makes the final call" is a disclaimer, not an allocation of responsibility.
- Can AI-reviewed plans be stamped? No. A plan review that carries legal weight requires a licensed human. Plans examiners now report applicants citing chatbot output to dispute review comments - including fabricated code sections that do not exist in any adopted edition. AI does not hold a license, and reviewers know it.
- What are the measured error rates? Almost no vendor publishes false-positive or miss rates. Asked directly, most answer with credentials rather than data. Until per-check accuracy numbers are normal in this category, run the practitioner's evaluation: replay a project you already reviewed and diff the tool's findings against your own comments.
- Which code edition, and whose amendments? A tool that checks the base IBC while your jurisdiction enforces a state-amended edition is checking the wrong book.
None of this is a reason to avoid the category. It is the due-diligence list for entering it.
What happens when review is automated well
There is one clean natural experiment. SolarAPP+, the federally developed automated permitting platform for residential solar, published its results: median permit review dropped from 7 business days to 0 - same-day - and full submission-to-final-inspection timelines fell by roughly two weeks, eliminating over 150,000 business days of delay in 2023 alone. The domain is narrow and prescriptive, which is exactly the point: automation works where the checks are measurable.
Government is moving in the same direction at larger scope. California launched a statewide AI plan-check tool for local governments in April 2025 to speed wildfire rebuilding, and a federal memorandum the same month ordered a permitting technology action plan across agencies. Austin cut initial site-plan review from 87 days to 32 not with AI but by fixing the process itself - proof that the correction cycle, not the technology, is the real variable.
The pattern across all of it: the clock moves when fewer cycles happen, and cycles happen because sets go in with problems still inside them.
How Lintel fits
Lintel works the submitter side of that equation. It reads the full document set - drawings, specifications, and schedules together - and surfaces the contradictions between them before the set goes out the door: the spec section that disagrees with the drawing note, the schedule that references a detail that is not there, the document contradictions that get expensive precisely when they surface late.
The public data above is the reason that matters. You cannot control the reviewer's queue, the department's staffing, or the code cycle. You control the quality of the set you file and the number of objection rounds it triggers. Find the conflict early and it is an edit. Find it late and it is a correction cycle - or a claim.
And the evaluation standard this guide recommends applies to Lintel first: run it on a document set you have already reviewed, and judge it by the diff. That is what a demo is for. And if you are comparing tools, the plan review software for contractors guide maps the field.
FAQ
Is AI plan review the same as automated code compliance checking?
No. Automated code compliance checking is one component - verifying documented conditions against prescriptive code requirements. AI plan review is broader: completeness, cross-discipline coordination, spec-to-drawing consistency, and code flags together. The research field behind the code-checking component dates to 2009; the document-wide review capability is what LLM-era tools added.
Can AI approve or stamp construction plans?
No. Permit approval and professional stamping legally require licensed humans in every US jurisdiction. Today's tools operate as a first-pass filter that hands findings to a qualified reviewer - and given the documented hallucination rates in the research, that is the only defensible configuration.
How accurate is AI plan review?
Vendors rarely publish error rates, so treat accuracy as domain-specific: strong on text, schedules, completeness, and cross-referencing; unreliable on drawn geometry, local amendments, and multi-step code logic. The 2026 academic literature still reports hallucinated rules as a core failure mode. Test any tool against a project you have already reviewed and measure the diff yourself.
Plan review software reads your documents against the code. Lintel reads them against each other - and the contradictions it finds are the correction cycles you never enter.
See what Lintel finds in your document set - Book a demo