Structured AI has launched a quality-assurance platform that brings a BIM model and the drawing set produced from it into a single review. The platform loads BIM models as IFC or native Revit files alongside PDF drawing sets, specifications, codes and project manuals, and its agents cross-reference the lot, according to AEC Magazine, which reported the launch on 6 October. The launch framing and its timing are the magazine’s reporting, and this piece carries them as that; the underlying capability pages are live on Structured AI’s own site, and the detail below is drawn from them.
The mechanical consequence is the point. A problem found on a drawing can now be traced to the element in the Revit model that caused it, and a Revit add-in surfaces findings in a sidebar so an engineer can navigate to the affected elements. Proposed fixes can be applied to the model once an engineer approves them. AEC Magazine puts the library at more than 900 checks, tailored to each firm’s standards, covering architectural, structural, MEP, fire-protection and civil disciplines, with integrations to Autodesk Construction Cloud and Bluebeam Studio; that figure is the magazine’s reporting — the product pages this publication reviewed on Structured AI’s own site carry no check count, describing the library instead as one that “keeps growing, built check by check as your firm defines what matters.”
“A problem on a drawing is usually a problem in the model as well, and engineers have had to connect the two by hand,” co-founder and chief executive Raymond Zhao told AEC Magazine.
Why the Export Is the Bug
The workflow Structured AI is attacking is so normal that most firms have stopped seeing it as a workflow at all. The company’s own description of it is a five-step loop: model in Revit, export the PDF set, mark up in Bluebeam, email the markups, re-fix by hand in Revit — and then repeat the whole thing at every revision and every milestone.
Each handoff in that loop discards information. A PDF is a picture of a decision, not the decision. Once the set is flattened, a reviewer is looking at lines, symbols and text, and a tool reviewing the set on their behalf is doing optical character recognition and pattern-matching against pixels. Structured AI’s argument is that this caps what a checker can conclude: a PDF check can report that something might be wrong, while a check with access to the model’s objects can traverse the actual duct network and return a definite answer tied to a specific element. The company says its agent runs inside Revit and writes native code against the Revit API, giving it the model’s parameters, tags and schedules across linked architectural, structural, MEP and civil models, and that findings come back as definite fails and items to verify, each with the element ID and level.
Three figures the company puts on the problem, reported with attribution rather than audited: half of billable hours at a typical engineering firm go to checking drawings; three in ten errors are caught by a manual review pass, with the rest surfacing in the field; and a modern set runs 500-plus sheets, re-exported and re-checked at every revision.
One more line from that page matters as much as any of the capability claims: nothing changes in the model until someone applies it.
The Three Doors
Structured AI now offers the same check library through three entry points, which is a more consequential design choice than the checks themselves.
A set can be uploaded directly as PDF, or pulled from the system it already lives in. The company’s integrations page lists Autodesk Construction Cloud and Procore as document sources; for ACC it says reviews run against the current issue, with findings cited back to the sheet. On its homepage the company describes the Bluebeam Studio route as moving markups both ways: the team’s comments pull in, and findings publish back as native PDF markups. And the Revit add-in runs the checks on the model’s live elements and writes approved fixes back to the source file. Outlook, Word, Excel, SharePoint, Google Drive and Gmail are also listed as connected sources.
The company states that files stay in place, that project data is never part of a training set, and that every organisation — pilots included — runs in its own isolated environment.
Meeting a firm where its drawings already are, rather than requiring a migration, is the unglamorous part of selling software to engineering practices, and it is usually what decides whether a pilot becomes a contract. A QA/QC tool that demands the set be exported into it has reintroduced the export step it was sold to eliminate.
The Trust Problem, Restated
When this publication covered Structured AI’s $4.2 million seed in June, the closing observation was that the company would live or die on whether its agents catch what matters without crying wolf — and that the question was execution. This launch is the clearest answer yet, and it answers it in a specific way: by narrowing what the agent is allowed to be uncertain about.
Three design decisions in the platform all point the same direction. Findings are cited to the sheet and detail they came from. Results are returned as definite fails and items to verify, rather than as scores — the company says explicitly that there are no confidence percentages, and that the output is a deterministic list a team can act on and sign off. And nothing is written to the model until a human approves it.
That is a product built by people who understand that the buyer is a professional carrying personal liability for the stamp on the drawing. The constraint is not model capability; it is that a QA/QC manager cannot sign off on a probability. Whether the underlying detection is good enough is still not something an outside reader can verify — and that caveat has not changed since June.
What this publication did not report in June is the named firm behind the endorsement. Structured AI’s site carries a quote from Robert Ioanna — described there as executive principal, chief technical officer and board member at the engineering firm Syska Hennessy Group — saying the firm got a tool custom-built for how it works and that its engineers genuinely like using it. That endorsement comes with a disclosure worth having: ENR reported at the time of the seed round that Syska Hennessy has been testing the platform and helping Structured AI develop it, and quoted Ioanna saying the firm invested in the company as one of its seed-stage companies. This is a development partner and investor speaking, not an arm’s-length buyer. The site also publishes three outcome claims — a full drawing set reviewed in 20 minutes; nine in ten findings triaged by client engineers confirmed and fixed; six times fewer issues reaching the field — which are the company’s figures and are reported here as such.
The Direction of Travel
The interesting thing about this launch is not that an AI tool reads drawings. It is which direction the integration runs.
Most of the drawing-intelligence cohort has moved outward from the document — reading PDFs better, connecting findings to schedules, pushing results into project management. Structured AI has moved inward, from the document back to the authoring environment that generated it. That is the same gravitational pull visible across the category this autumn: Kora Studio shipping a façade manufacturer’s factory constraints as a Revit plug-in, Primepoint putting its drawing reviews inside Autodesk Forma, and Motif building a BIM authoring tool around agents from the start. The checking layer and the authoring layer are collapsing into each other, and the open question is which one ends up hosting the other.
For a startup, the risk in that position is obvious and worth stating. Building inside Revit means building inside Autodesk’s product, on Autodesk’s API, in a year when Autodesk is shipping its own construction agents. Depth in the model is a genuine moat against tools that only see pixels. It is a thinner one against the company that owns the model format.
