UpCodes published a data release on 13 October saying its AI Plan Review has reviewed more than 1,300 construction plan sets since launching in June and identified more than 180,000 potential issues, according to the announcement, which carries an Austin, Texas dateline.
Vendor usage milestones are usually the least informative genre in this sector — a count of something the vendor controls, presented as evidence of something it does not. This one is worth reading because the release does not stop at the totals. It breaks the findings down by severity, by cross-discipline conflict and by code category, and in doing so publishes a profile of what is actually wrong with construction documents at the point an AI first looks at them. That profile is the story, and it is more interesting than the headline number.
What the Release Says It Found
The figures, all as stated in the release:
- 1,300+ plan sets reviewed since the June launch
- 180,000+ potential issues identified
- Coverage spanning 73 jurisdictions and 14 code editions, from single-family homes to commercial and medical facilities
- An average of 21 critical issues per plan set
- Missing exit signage as the single most common critical finding
- Roughly one in ten critical issues being cross-discipline coordination conflicts
- 46% of critical issues involving life-safety systems — specifically means of egress, fire protection and fire-resistance-rated construction
One calculation is ours rather than the company’s, and it is worth doing because the release does not: 1,300 plan sets at an average of 21 critical issues each puts critical findings on the order of 27,000, against a total of more than 180,000 potential issues. Both inputs are approximate — “1,300+” and an average — so the product is approximate too. But the shape it implies is that the large majority of what the tool flags is not critical, and that a reviewer’s problem is triage rather than volume.
That is a more useful way to read an AI document-review tool than the headline total on its own, and the per-set shape is what a design principal would want before buying. A tool that surfaces roughly 140 findings on an average set — 180,000 spread across 1,300, again our arithmetic — and tells you which 21 will fail an inspection is useful. One that surfaces 140 and ranks none of them is a new source of work.
Missing Exit Signage Is a More Revealing Answer Than It Sounds
On the company’s own figures, the most common critical finding across those 1,300 sets is not a structural miscalculation or an energy-code failure. It is a missing exit sign.
Taken with the 46% life-safety share, that points at something specific about how drawing sets fail. Egress, fire protection and rated construction are among the most heavily prescribed parts of the building code and among the most systematically checked at plan review by the authority having jurisdiction. They are also distributed across sheets — a life-safety plan, a reflected ceiling plan, an electrical sheet, a door schedule — which is exactly the kind of claim that is easy to satisfy in one place and omit in another.
The cross-discipline figure points the same way from the other side: roughly a tenth of critical issues are conflicts between disciplines rather than errors within one. The release does not say which discipline the missing exit signs belonged to, and the inference that follows is ours rather than the company’s — that a requirement of this kind is more often lost between sheets than specified wrongly on any one of them.
This is the structural argument for machine review that has been circling the category, and it is the same argument made from opposite directions by two QA/QC launches this newsroom covered earlier this month. Structured AI moved from reviewing drawings to reviewing the model the drawings came from precisely because a drawing-only check cannot see the authoring error behind an inconsistency. Pirros pointed its Revit agent at a whole set rather than a sheet for the same reason. A code requirement that is satisfied by the coordinated state of several sheets is not a reading problem; it is a cross-referencing problem, and a human doing it at 2am before an issue date is the control that fails.
The $1 Million Figure Is a Model, and the Footnote Says So
The release’s third headline number is that Plan Review “surfaced more than $1 million in estimated potential construction error costs on average per project.” It carries an asterisk, and the footnote is the useful part: “Based on average issue counts from UpCodes Plan Review run plan sets and estimated costs to resolve comparable issues during design, permitting, and construction phases.”
In other words it is a modelled exposure — issue counts observed multiplied by cost estimates for comparable issues — and not money any customer has reported saving. RuntimeWire’s write-up of the release makes the same point, calling the figure “a modeled cost, not verified savings.” Readers should treat it as a sizing argument rather than a result. Disclosing the derivation in the document is better practice than a good deal of what this sector publishes, where the multiple appears and the method does not.
The release pairs that with a scope limit which is unusual to see in a vendor announcement at all: UpCodes states that Plan Review is a first-pass analysis and does not replace review by a licensed design professional or approval by the authority having jurisdiction. On a product whose output is code-compliance findings, that sentence is doing legal work, and it is also an accurate description of what the tool is for.
Chief executive and co-founder Scott Reynolds frames the shift the release is arguing for: “The first chapter of AI at work was about helping people create faster.” Later in the same quote comes the pitch — “AI gives teams another set of eyes to surface potential issues earlier, when they’re easier and less expensive to address.”
The Customers Named, and the Base Behind Them
Three adopters are named: Dallas-Fort Worth Airport, BDG Architects and Hereford Dooley Architects. An aviation owner and two architecture practices is a mixed set, and the release does not say what any of the three is using the tool on or at what scale.
The product itself launched on 3 June, and that release is where the platform numbers sit: UpCodes puts its library at 11 million locally adopted building code sections and amendments across 6,000+ jurisdictions, and says 800,000 architecture, engineering and construction professionals use the platform. The June release also describes the mechanics — a project is configured once with its jurisdiction, code year and building type, then analysed by discipline across architecture, structural, MEP, fire protection, life safety, accessibility and energy, with issues sorted by severity, linked to the drawing page and the code section, and discrepancies between plans and specifications flagged automatically.
Set the two releases against each other and the strategic move is visible. UpCodes built its position on code research — the place a designer goes to read what the adopted code says. Plan Review turns that corpus into a check that runs against the drawings. The asset is the same; the product is now in the project workflow rather than beside it. That is a materially stronger place to sell from, and a materially harder one to be wrong in.
What This Does Not Settle
Analysis, labelled as this publication’s own.
Four months of usage data tells you what a tool flags. It does not tell you how often it is right. The release reports potential issues identified; it does not report a false-positive rate, a precision figure, or how many flagged items the design teams agreed with. On a code-compliance product that is the number that decides whether the tool saves a reviewer time or costs it, and vendor announcements in this category — this one included — rarely carry it.
The second open question is competitive. AI drawing and code review is a crowded segment and getting more so: Buildcheck raised $12 million nine days after taking in Plangrep, a sequence this newsroom read at the time as two venture-stage companies in the same category combining rather than an incumbent buying capability it lacked, and the model-side and set-side entrants above were both published here within the past fortnight. What distinguishes UpCodes is not the review engine. It is the 11 million code sections and the 800,000 professionals who already open the site to read them. In a segment where everyone’s demo looks similar, the defensible asset is the corpus and the distribution — and publishing a findings profile from 1,300 sets is, read commercially, a demonstration that it has both.
