Your Architect Checked 40 Pages of Code. The AI Checked 11 Million.

By Catherine Chen, August 28, 2026, Policy & Regulation
Architect reviewing building plans with AI code compliance overlay showing highlighted code sections

Pueblo County, Colorado, had a problem every rural building department recognizes, where staff manually reviewed house plans against the 2015 International Residential Code plus local amendments while initial turnaround ran two to three weeks and the backlog kept growing until planning director Rachel Gaffney tried something in October 2024 that a year earlier would have sounded like a vendor pitch.

She let an AI read the plans first, which cut that two-week wait to minutes for the triage that tells you what is missing, not approval but the first pass that used to consume a planner's entire morning.

Blitz is generative AI built for local governments that cross-references additions and remodel documents against local land use codes and the IRC, and county officials say it helps clear routine residential checks so experienced staff can spend time on complicated commercial projects along corridors from Union Avenue to the steel mill district, the municipal side of a market that just doubled in six months.

Two AIs, opposite sides of the counter

On June 3, 2026, UpCodes announced Plan Review, an AI-native QA/QC feature that does the inverse of Blitz and solves the same problem from opposite sides of the counter, where Blitz sits inside OpenGov asking whether an application is complete from the city's perspective while UpCodes sits inside the design firm's workflow asking whether the drawing set is correct before the city ever sees it.

UpCodes says 800,000 professionals trust its platform for code research, and its library now spans 11 million locally adopted sections across 6,000 jurisdictions, which explains why residential QA/QC is broken because the IBC is 700 pages, NEC 2026 is 1,100 pages with 30-plus residential changes, yet adopted code in Pueblo County is not adopted code in Denver and neither matches Boulder's amendments.

A builder working across three Colorado counties navigates three code families with local dialects that shift every cycle, and your electrician who reads maybe 40 pages that matter to him learns quickly that the code does not care about his schedule or yours.

Here is how UpCodes says Plan Review works in practice, where you create a project, set jurisdiction, code year, and building type once, upload drawings, choose discipline-specific analysis across architecture, structural, MEP, fire protection, life safety, accessibility, and energy, or a general review or checklist template, and the system returns issues organized by severity and category with each linked to drawing page and governing section while flagging discrepancies between plans and specs automatically so teams can filter, comment, assign, track, and call on Copilot for interpretation before re-uploading and exporting the report for officials, which is systematic in a way manual QA/QC rarely achieves when Denver approves just 37 percent on first try.

The 37 percent problem

Denver approved 37 percent of applications on first try before its current push according to Denver7 via PYMNTS in May 2026, and the city signed a five-year, $4.6 million contract with CivCheck in March, an AI-enabled plan review platform built by Clariti that flags missing documents, incomplete fields, and errors before reviewers see them, while Robert Peek, Denver's director of development systems performance, said the goal is fewer redundant cycles that have long bottlenecked housing.

Denver also set a 180-day shot clock with a $10,000 refund promise if missed, and it had already cut possession time for single-family and duplex projects by roughly 45 percent since 2023, yet staffing makes the case sharper because Community Planning cut 59 positions for 2026 down to 251 per the Denver Gazette, leaving fewer people to handle the same complexity that grows every code cycle.

Honolulu launched CivCheck in December for residential with commercial to follow by mid-2026, Seattle Mayor Bruce Harrell signed an executive order in June 2025 routing all development through an AI pilot led by a Permitting and Customer Trust team, Austin partnered with Archistar for zoning review, and California Governor Gavin Newsom announced April 30, 2025 an Archistar tool for LA wildfire rebuild that is free to local governments via LA Rises and Steadfast LA plus Autodesk and Amazon, using computer vision and ML to instantly check designs against zoning and building codes for pre-check before submission.

PYMNTS summarized the pattern well when it noted permitting delays stem less from technical review complexity than from application quality at intake, so catching errors before formal submission costs less than cycling incomplete applications through the queue multiple times, and Denver wants to push 37 percent to 80 percent which is completeness while UpCodes wants to catch substantive compliance before completeness gets checked, meaning residential needs both but currently gets neither reliably despite paying the highest price for failure.

What a missed note costs

Navigant and the Construction Industry Institute put average RFI cost at $1,080 and 7.4 days to resolve, commercial projects average 9.9 RFIs, and custom homes run 6 to 9 RFIs on a $1.2M to $2M build when you count structural clarification and MEP coordination where margin disappears quietly between trades.

Here is original math for residential QA/QC that no vendor publishes because it exposes how thin margins are, where catching 40 to 60 percent of RFIs pre-submission, conservative given UpCodes claims of systematic identification, equals 3 to 5 RFIs avoided, which at $1,080 each is $3,240 to $5,400 direct and at 7.4 days each is 22 to 37 days schedule avoided, while carrying cost on a $650,000 construction loan at 8.5 percent is about $151 per day so 22 days is $3,322 and 37 days is $5,587 in interest alone that the builder absorbs.

Combined value per project lands between $6,562 and $10,987, and on a $1.5M home with 12 percent margin where profit is $180,000, that protection represents 3.6 to 6.1 percent of margin saved by a check that takes minutes rather than the hour a principal would spend redlining at 4 PM on a Friday when the third set of the day all starts to look the same.

Methodology matters, so inputs are Navigant RFI cost and cycle time, residential RFI range from three custom builder interviews in Northern California in March 2026 for this publication's supply chain series, and Freddie Mac average construction loan rate Q1 2026 at 8.25 to 8.75 percent, with assumption that UpCodes catches 40 to 60 percent, unproven for residential because vendor has published no third-party audit, meaning if catch rate is 20 percent value drops to $3,281 to $5,493 and if 70 percent rises to $7,500 to $12,500, while break-even on a hypothetical $300 per month subscription is one avoided RFI every 3.7 months, which any active residential firm hits easily.

Second calculation addresses jurisdiction complexity where 6,000 jurisdictions times 15 local amendments equals 90,000 variation points and 11 million sections divided by 6,000 equals 1,833 sections per jurisdiction average, which no small residential firm with one designer covering three jurisdictions tracks manually while a large commercial firm has a dedicated code specialist and a three-person studio has a principal who knows home jurisdiction well and guesses at the other two, and guessing creates the $35,000 rework that wipes 19.4 percent of margin on that $1.5M home and teaches the lesson that code knowledge is not evenly distributed.

Where this falls apart

UpCodes talks about multi-million dollar errors shrinking margins when not caught pre-construction, yet all examples are commercial and no residential case study exists yet, while residential rejection appears higher than commercial at 63 percent in Denver versus an estimated 35 to 45 percent commercial, but residential has less tooling and less cushion to absorb the hit which is exactly the gap a residential-focused QA/QC tool should fill if vendors bothered to look.

Eleven million sections sounds comprehensive, but much is duplication where IBC 2021 Section 1011.1 means the same thing in Fort Collins and Colorado Springs until a local amendment changes it, so effective unique logic is smaller yet still large enough to be unmanageable, and headline overstates novelty while understating the real problem which is local interpretation, not sheer volume, because an inspector's reading matters more than the text itself.

Code is the floor, not the ceiling, because a house can pass every code section and still leak when water management, flashing sequence, vapor drive, and grading do not live in UpCodes library, and a junior designer who gets zero flags and skips senior review builds a different risk that Pueblo County avoids by retaining human authority explicitly while UpCodes phrasing like systematically identify nudges toward autonomy that residential cannot afford when the failure mode is a wet wall cavity.

Liability is unresolved because ICC has not issued guidance on AI-assisted QA/QC, no state licensing board mentions AI in QA/QC requirements, UpCodes terms likely disclaim liability for missed issues, and if Plan Review misses a 2015 IRC R311.7.5.1 egress violation that becomes a failed final, the designer who stamped it carries the cost because tools do not stamp and never will, which is worth remembering when marketing suggests otherwise.

Local interpretation matters more than text since an inspector in unincorporated San Mateo County reads WUI Chapter 7A differently than one in Los Angeles County even though both cite the same code, and no AI captures that tacit knowledge that only the builder who has failed with that inspector before possesses, which is why field experience still commands a premium that software cannot discount.

Strongest counterargument

The best case against AI-native QA/QC is not that it fails but that it succeeds just enough to make you lazy, because principals already stretch across sales, site visits, and permit chasing, and a clean AI report makes skipping the hour of redlining tempting even though that hour is where non-code failures get caught, like window head flashing that relies on tape that fails in five years, HVAC closets with no service clearance, or stairs with 7-7/8 inch risers that meet code but trip a 68-year-old client who will remember your name for the wrong reason.

Denver cutting 59 positions while adding a $4.6M AI contract is the municipal version of the same trade where fewer humans and more software create a gap between complete and correct, and if software says complete does anyone verify that complete means correct when CivCheck helps applicants identify issues earlier and preserves professional judgment per Clariti VP Julia Richman, that last clause does heavy lifting that residential builders cannot outsource to a checkbox.

Residential needs more judgment, not less, because the library grew to 11 million sections precisely because we kept adding rules to compensate for judgment we stopped exercising, and another rule-checking tool does not fix that structural deficit even if it catches a missing egress window on the way, which has value but is not a substitute for looking at how water moves through an assembly that code describes only in abstract terms.

Limitations

This analysis uses UpCodes claim of 11M sections and 6,000 jurisdictions without independent verification, and we could not obtain pricing for Plan Review so ROI uses hypothetical $200 to $500 per month based on existing tiers where Copilot is $49 and full platform is $199, though actual price may be higher for multi-jurisdiction firms that need the most coverage and would benefit most.

No third-party accuracy audit exists for UpCodes Plan Review, Blitz AI, or CivCheck, all performance claims are vendor or municipal press releases, Pueblo County's minutes versus weeks is for initial triage not full review, and Denver's 37 to 80 percent goal is aspirational not achieved as of May 2026, which means any ROI math is provisional until independent testing exists that measures actual catch rates on residential drawings.

RFI cost $1,080 is commercial and may overstate residential, carrying cost assumes full loan drawn overstating early-phase cost, and residential RFI count 6 to 9 is based on small-n builder interviews not a published dataset, so treat as directional rather than definitive, which is still better than no estimate at all when you are deciding whether to pilot a tool.

We did not test UpCodes Plan Review on residential drawings and could not verify whether it flags IRC egress, stair geometry, or energy code correctly for custom homes, meaning performance may differ between residential and commercial drawings where training examples likely concentrate, and that uncertainty should be priced into any adoption decision.

What to do with this

If you design 8 to 12 custom homes per year across 2 to 4 jurisdictions, UpCodes is worth piloting if priced under $400 per month because break-even is one avoided RFI per month or one avoided rework per year, and workflow that works is setting jurisdiction, code year, and building type once per project, uploading PDFs, running architecture plus structural plus MEP plus energy, filtering by severity, assigning to team, re-running after revision, and exporting the report to include in your permit package as internal QA/QC documentation rather than certification.

If your building department uses Blitz, CivCheck, or Archistar, ask because these municipal tools are free to you as applicant, with Pueblo County using OpenGov, Denver using Clariti, and Honolulu using CivCheck for residential now, and if your jurisdiction uses one the pre-check logic is public enough to self-audit where incomplete fields, missing documents, and basic code compliance top rejection lists, so fixing those before submission already puts you in Denver's targeted 80 percent.

Do not skip manual QA/QC because AI catches code while it does not catch constructability, so keep senior review for water management, flashing, and site-specific issues, document AI use in your log as first-pass not final, and that note protects you if AI misses something and you need to show standard of care to a client or inspector who asks what you did to verify compliance before you stamped the set.

The code does not care about your timeline, but it is now readable by machines that do not get tired at 4 PM on a Friday when you are reviewing your third set that day, so use them for the tedious part and save judgment for the parts that leak, burn, or fall down, which no library will ever capture completely no matter how many million sections it claims to index.

Sources