The AEC Intelligence Report — Software Deep Dive

AI Construction Software?
Where Is It?

A complete map of every AI tool claiming to transform AEC — what they actually do, what they can’t do, and the brutal programming reasons why 3D BIM still refuses to be automated

By The AEC Intelligence Editorial Team  ·  Published May 2026  ·  Last Updated May 2026  ·  28 min read

Editorial Note: Software capability assessments in this article are based on publicly available documentation, user community reports, and technical analyses current as of May 2026. Software changes rapidly; readers should verify current capabilities directly with vendors. This article contains no affiliate relationships with any software company mentioned.

Quick Answer

It exists — in fragments. Dozens of AI tools are live and used daily across AEC workflows in 2026. But none of them have solved the core problem: generating a fully coordinated, multi-discipline 3D BIM model directly from design intent. The reason is not a lack of funding or ambition. It is a set of deeply specific programming and computational geometry problems that current AI architectures are structurally ill-equipped to solve. This article maps every major AI tool in the AEC stack, compares what they actually deliver, and explains the exact technical barriers that keep true 3D BIM automation out of reach.

If you search “AI construction software” in 2026, you get hundreds of results. Vendors promise AI-powered project management, AI clash detection, AI design generation, AI cost estimating, AI site monitoring. The marketing is confident. The investment money is real — over $4 billion flowed into construction technology in 2024 alone.

And yet, the BIM coordinator is still manually routing the duct. The structural engineer is still fixing the coordinate offset by hand. The quantity surveyor is still correcting the model export before the numbers make sense.

This article does two things. First, it gives you a complete, honest map of the AI tools that are actually being used in AEC right now — what they do, who makes them, and what their real limitations are. Second, it explains, at a technical level, exactly why building a true AI-to-3D-BIM pipeline is so much harder than it looks — and why the software industry keeps producing tools that solve 20% of the problem while leaving 80% untouched.

Part 1: The Complete Map of AI Tools in AEC (2026)

We have organized every significant AI tool in AEC into seven functional categories. For each tool, we assess: what the AI actually does (the underlying model type), what it integrates with, what it cannot do, and how mature the technology is.

Category 1

AI-Assisted Design Generation & Space Planning

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
TestFit	TestFit Inc.	Procedural constraint solver + ML optimizer	Generates residential/commercial unit layouts meeting site, zoning, and program constraints in real time	Revit export (geometry only, limited attributes)	No structural or MEP content; output is massing shell only	Production-ready
Maket.ai	Maket	Generative AI (diffusion + rule-based layout)	Generates multiple floorplan options from text/parameter input; stylistic variation	DXF/PDF export; no native BIM link	2D floorplan output only; not a BIM model	Early commercial
Giraffe	Giraffe	Parametric + GIS-linked constraint engine	Site feasibility and massing studies with zoning data overlay; development yield analysis	IFC export (limited); Revit not native	Conceptual massing only; no structural logic	Production-ready
Spacemaker (Autodesk)	Autodesk	Generative design + multi-objective optimizer	Generates site layout options optimized for daylight, wind, noise, and density simultaneously	Revit link (massing); Forma platform	Output is conceptual volumes; no building systems content	Production-ready
nTop (nTopology)	nTopology	Implicit modeling + topology optimization	AI-driven structural topology optimization for components (facades, structural nodes); primarily AEC manufacturing crossover	STL/STEP export; limited BIM link	Component-level only; not a project-scale BIM tool	Mature (manufacturing-crossover)

Category 2

AI Clash Detection, Model QA & Coordination

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
Construction IQ	Autodesk (BIM 360 / ACC)	NLP + supervised classification (issue prioritization)	Ranks RFIs, submittals, and clash reports by predicted cost/delay impact using historical project data	Native ACC / BIM 360	Prioritizes existing clashes; does not find or resolve them	Production-ready
Navisworks Clash Detective	Autodesk	Rule-based geometry intersection (not true ML)	Geometric intersection detection between federated model elements; outputs clash report	NWC/NWD from all major BIM tools	Not AI; no resolution suggestions; high false-positive rate	Mature (not AI)
Solibri Model Checker	Solibri (Nemetschek)	Rule engine + ML-assisted classification (recent versions)	IFC model validation against configurable rule sets; space analysis; accessibility checking	IFC-native; all major BIM authoring tools	Rule sets must be written and maintained manually per project type	Mature
Predictive Safety (Autodesk)	Autodesk	Computer vision + supervised classification	Analyzes site photos for safety hazards (missing PPE, proximity violations); integrated in ACC	ACC / BIM 360 site photos	Photo analysis only; does not interface with 3D model	Commercial beta

Category 3

AI Energy, Sustainability & Performance Analysis

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
Cove.tool	Cove.tool	ML surrogate model (trained on EnergyPlus simulation data)	Real-time energy analysis from BIM geometry without running full EnergyPlus simulation; early-design optimization loops	Revit plugin; IFC import; Rhino link	Surrogate model accuracy degrades on unusual building geometries outside training distribution	Production-ready
Sefaira	Trimble	Physics simulation + ML optimization (post-2022)	Live energy and daylighting feedback in Revit and SketchUp during design; design option comparison	Native Revit + SketchUp	MEP systems not included in model; envelope analysis only at schematic stage	Production-ready
Insight (Autodesk)	Autodesk	Cloud simulation + ML benchmarking	Building performance analysis (energy, solar, wind) from Revit model; LEED/BREEAM benchmarking	Native Revit	Requires fully attributed Revit model to run; useless on incomplete geometry	Production-ready

Category 4

AI Code Compliance & Permit Screening

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
Archistar	Archistar	NLP rule extraction + GIS constraint engine	Automatically extracts zoning rules from planning documents; checks proposed massing against those rules	Massing geometry via API; not a BIM authoring link	Zoning/planning rules only; building code (structural, fire, accessibility) not covered	Production-ready (AU/US)
UpCodes AI	UpCodes	LLM (RAG over building code corpus)	Natural language queries against full US building code database; returns relevant code sections with citations	No BIM integration; document query tool only	Cannot check a BIM model against the codes it retrieves; query-only	Production-ready
Joist AI	Joist AI	NLP + classification over permit data	Predicts permit approval likelihood and flags likely objections based on local authority historical data	No BIM integration; works on project description text	Prediction accuracy varies significantly by jurisdiction data coverage	Early commercial

Category 5

AI Cost Estimating & Quantity Take-Off

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
Togal.AI	Togal.AI	Computer vision + object detection (PDF drawings)	Extracts quantities from 2D PDF construction drawings via visual recognition; faster than manual digitizing	2D PDF only; no direct BIM model link	Not a BIM tool; works on drawings that come out of BIM, not the model itself	Production-ready
CostX	Exactal	BIM geometry parser + rule-based QS mapping	Extracts quantities directly from BIM model elements; maps to cost codes; integrates historical pricing	IFC, Revit, ArchiCAD, Tekla	Accuracy entirely dependent on model attribute completeness; broken attributes = wrong QTO	Mature
Procore Estimating	Procore	ML cost prediction + historical bid analysis	AI-suggested cost ranges from historical project data; integrates with BIM-linked takeoff tools	Procore CDE; BIM model link via partner tools	Prediction quality requires large local cost database; poor on unusual project types	Production-ready

Category 6

AI Site Monitoring, Progress & Safety

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
OpenSpace	OpenSpace	Computer vision + SLAM (Simultaneous Localization and Mapping)	360° photo capture mapped to BIM model floorplan; AI compares photos to BIM for deviation detection	Revit / IFC model for spatial reference; ACC integration	Detects visual deviations; cannot interpret structural significance or update BIM model automatically	Production-ready
Reconstruct	Reconstruct Inc.	Computer vision + 4D BIM schedule comparison	Compares drone/photo capture to 4D BIM schedule model; flags behind-schedule areas; progress reporting	Navisworks 4D model; ACC	Progress flagging only; does not update BIM model or reschedule tasks automatically	Production-ready
Trimble SiteVision	Trimble	AR overlay + GPS-anchored BIM positioning	Displays BIM model overlaid on physical site in AR; field crews can verify placement against model	Trimble Connect; IFC import	Display and verification only; no model writing capability	Production-ready

Category 7

AI Project Management, Scheduling & Generative LLM Copilots

Tool	Company	AI Model Type	What It Actually Does	BIM Integration	Hard Limit	Maturity
Revit Copilot	Autodesk	LLM (GPT-based) + Revit API function calling	Natural language commands to modify Revit model parameters, run tools, query element properties, generate code snippets	Native Revit (2025+)	Command execution only; cannot create geometry or reason spatially; hallucinates on complex multi-step tasks	Beta / rolling
Procore Copilot	Procore	LLM (RAG over project documents)	Answers natural language questions about project documents (RFIs, submittals, specs); draft RFI responses	Procore CDE; no BIM model access	Document QA only; no model interaction	Beta / rolling
Alice Technologies	Alice Technologies	Combinatorial optimization (not neural net)	Generates and optimizes construction schedules from BIM WBS; evaluates billions of sequencing combinations; scenario analysis	IFC model import; Primavera P6 / MS Project export	Requires detailed, pre-structured WBS input; does not auto-generate WBS from model	Production-ready
Hypar	Hypar (WeWork spinoff)	Parametric function library + ML element classifier	Cloud-based building generation functions (floors, cores, facades) combinable via workflow scripts; BIM-compatible output	IFC export; Revit importer; Rhino/Grasshopper plugin	Each function must be individually scripted; not a unified AI model; high setup cost	Active development

The Pattern Across All 7 Categories

Every tool above automates a slice of the workflow. None of them connect. A massing generated by TestFit does not automatically become a coordinated Revit model. A clash prioritized by Construction IQ does not automatically get resolved. An energy analysis from Cove.tool does not update the MEP model. The tools exist. The connections between them do not — and building those connections is where the real programming problem lives.

Part 2: Why AI Cannot Just “Generate” a 3D BIM Model — The Real Programming Reasons

This is the part that gets skipped in most AI-in-AEC coverage. The claim that “AI can’t yet generate full BIM models” is frequently stated but rarely explained at the level that actually matters — the code. Here is a precise account of why the problem is so hard, broken down by the specific programming and computational geometry barriers involved.

Barrier 1: BIM Is Not an Image — It Is a Relational Database with Geometry Attached

Most people visualize a BIM model as a 3D image. Programmers know it is something fundamentally different: a relational database where every object has geometry, properties, relationships to other objects, and constraints that govern how it can change.

A Revit model is not a mesh or a point cloud. It is a structured data graph. A wall element, for example, contains:

Wall {
  geometry: solid_extrusion(profile, height, offset),
  type_parameters: { material, fire_rating, thermal_resistance },
  instance_parameters: { base_constraint, top_constraint, location_line },
  joins: [Wall_A, Slab_3, Column_7], // topological constraints
  hosted_elements: [Door_21, Window_14, Opening_5],
  phase: { created: "Phase 1", demolished: null },
  relationships: { room_bounding: true, structural: false }
}

When you move that wall, every join, every hosted element, every constraint must resolve simultaneously without creating geometric errors. This is handled by a constraint solver — a system that maintains mathematical consistency across the entire model graph as changes propagate.

Current AI architectures — transformers, diffusion models, GANs — are trained to predict tokens or pixels. They do not natively represent or enforce relational constraints. An AI that “generates” a BIM model by predicting geometry without enforcing the relational graph will produce outputs that look like a BIM model but break the moment any element is moved, hosted, or connected.

Why this is hard to solve in code:

Building a constraint solver for architectural geometry requires formal logic systems (CSP — Constraint Satisfaction Problems), not gradient descent. Merging a neural generator with a constraint solver requires the AI output to be differentiable through the constraint system — an active research area with no production solution as of 2026.

Barrier 2: 3D Spatial Reasoning at Millimetre Precision — A Fundamentally Different Problem Than Image Generation

Image generation AI (Stable Diffusion, Midjourney, DALL-E) works by learning the statistical distribution of pixels in image datasets. It produces outputs that are plausible in the sense that they look like real images. It does not need to be correct to within a millimetre — visual plausibility is the entire standard.

BIM geometry operates under a completely different correctness standard. Consider a hospital MEP coordination problem:

// Structural beam: bottom of steel at Z = 3450mm
// Main chilled water pipe: centerline at Z = 3385mm, diameter = 200mm
// Therefore: pipe top edge at Z = 3485mm
// Clash: pipe top (3485mm) > beam bottom (3450mm) by 35mm
// Required clearance: 50mm minimum (fire code + insulation)
// Resolution: lower beam by 100mm OR reroute pipe to adjacent grid bay
// Downstream impact: lowering beam triggers re-check of 847 connected elements

An AI system that resolves this clash must: (1) parse the 3D geometry of both elements at millimetre precision, (2) understand the legal clearance requirement from building code, (3) evaluate both resolution strategies, (4) propagate the chosen resolution through all 847 connected elements without creating new clashes, and (5) do this across a federated model where structural, architectural, and MEP data come from different coordinate systems.

Current neural network architectures represent numbers as floating-point tensors. Floating-point arithmetic is inherently imprecise at the scale required for construction geometry. A cumulative rounding error across a 200m building can produce positional errors of several millimetres — enough to invalidate clearance calculations. Production CAD and BIM systems use exact arithmetic kernels (like ACIS or Parasolid) precisely to avoid this. There is no neural network equivalent.

Why this is hard to solve in code:

Geometric deep learning (PointNet, Graph Neural Networks, NeRF) can learn 3D representations, but these representations are statistical approximations — they cannot substitute for the exact Boolean geometry operations (union, intersection, difference) that clash detection requires. A hybrid system that uses GNNs for high-level spatial reasoning and ACIS for exact geometry verification would be theoretically correct, but the interface between probabilistic AI outputs and deterministic geometric verification is an unsolved engineering problem.

Barrier 3: The IFC File Format — Why “Open BIM” Is Not Actually Open to AI

IFC (Industry Foundation Classes) is the ISO-standard open format for BIM data exchange. In principle, any AI system should be able to read IFC files, learn from them, and generate new IFC files as output. In practice, IFC is one of the most hostile data formats for machine learning that exists.

Here is what an IFC file actually looks like:

#1=IFCPROJECT('2O2Fr$t4X7Cv2yVd0Nm5fA',#2,'My House',$,$,$,$,(#20,#23),#7);
#7=IFCGEOMETRICREPRESENTATIONCONTEXT($,'Model',3,1.E-5,#8,$);
#8=IFCAXIS2PLACEMENT3D(#9,$,$);
#9=IFCCARTESIANPOINT((0.,0.,0.));
#245=IFCWALL('0LV8Wt0X5Fe8hFSmpGbD1f',#2,'BasicWall:Interior - 135mm Partition',$,'BasicWall:Interior - 135mm Partition:346673',#246,#253,'346673');
#246=IFCLOCALPLACEMENT(#247,#248);
#247=IFCLOCALPLACEMENT(#15,#16);
#248=IFCAXIS2PLACEMENT3D(#249,#250,#251);

Every entity reference is a pointer to another entity. The geometry of a single wall element is typically spread across 40–200 separate entity records, each referencing others. To reconstruct the 3D geometry of one wall, a parser must resolve a recursive chain of references that can reach 15–20 levels deep.

The practical consequences for AI training:

Problem	Technical Detail	Impact on AI Training
Recursive entity resolution	No entity is self-contained; all geometry requires full graph traversal	Cannot be tokenized like text; sequence models break on circular references
GUID instability	Every element gets a new GUID on export; same element has different IDs in different files	Cannot track element identity across model versions; diff-learning is impossible without pre-processing
Schema version fragmentation	IFC2x3, IFC4, IFC4.1, IFC4.3, IFC5 (draft) have different schemas; no backward compatibility guarantee	A model trained on IFC2x3 data does not generalize to IFC4 data without extensive re-preprocessing
Vendor-specific extensions	Revit, ArchiCAD, Tekla all export non-standard property sets that are not in the IFC schema	Training data from mixed-vendor sources is structurally inconsistent; AI learns vendor artifacts, not building semantics
No standardized coordinate origin	Models from different disciplines often have different project base points; federation requires manual offset correction	Training data from real projects contains systematic spatial offsets; AI learns wrong positional relationships between disciplines

Why this is hard to solve in code:

Building an IFC parser robust enough for AI training use (not just data reading) requires constructing a full in-memory graph database from the STEP-encoded file, resolving all references, normalizing coordinates, and standardizing property sets across vendors — before a single training example is usable. The open-source IFC parsers (IfcOpenShell, xBIM, IFCjs) do not produce AI-ready tensor representations. This preprocessing layer is itself a multi-year engineering project.

Barrier 4: Multi-Discipline Coordination — The Context Window Problem at Scale

A federated BIM model for a large building contains millions of elements across structural, architectural, and MEP disciplines. To reason about whether a proposed MEP routing is feasible, an AI system must simultaneously hold in context:

• The 3D geometry of all elements in the relevant zone (potentially 50,000+ elements)
• The structural logic governing which elements cannot be penetrated or moved
• The MEP routing rules (minimum bend radii, clearance requirements, support spacing)
• The fire compartmentation boundaries that the route may not cross
• The sequence constraints (what has already been built and cannot change)
• The spatial relationship between the current element and all adjacent disciplines

The largest current LLM context windows (1 million tokens in some models) can hold approximately 750,000 words of text. A medium-complexity building IFC file encoded as a token sequence would require 200 million to 2 billion tokens to represent completely — 200–2,000 times the largest available context window. No transformer-based model can reason about an entire building model in its context window. It would need to be chunked, with the AI making local decisions without global context.

// Rough token estimation for a 10,000m² hospital IFC file:
// Total IFC entities: ~800,000
// Average tokens per entity (geometry + properties): ~250 tokens
// Total tokens required: ~200,000,000 (200M)
//
// GPT-4 context window: 128,000 tokens
// Claude 3 context window: 200,000 tokens
// Gemini 1.5 Pro: 1,000,000 tokens (largest available 2024)
//
// Gap: 200x to 1,500x larger than any available context window
// Chunking introduces: loss of global spatial context,
// boundary condition errors, inconsistent constraint resolution

Why this is hard to solve in code:

Hierarchical or chunked processing of BIM data means the model makes local decisions without seeing global constraints. A duct routed correctly through Zone A may create an impossible condition at the Zone A/B boundary when Zone B is processed. Solving this requires either (a) a context window large enough to hold entire building models, which is 2–3 orders of magnitude beyond current technology, or (b) a multi-agent architecture where local agents share a consistent global state model — which is itself a distributed systems problem without a standard solution.

Barrier 5: Proprietary API Lock — Why You Cannot Train Directly on Revit Data

The world’s BIM data — the training corpus that would teach an AI what buildings look like at the model level — is almost entirely locked inside proprietary file formats. Revit’s native .rvt format is a binary SQLite-based database that Autodesk does not fully document. The only supported way to access Revit data programmatically is through the Revit API, which requires a running instance of Revit to be installed and licensed.

This creates a fundamental training data problem:

Format	Access Method	Public Availability	AI Training Feasibility
.rvt (Revit)	Revit API only; undocumented binary	Proprietary / closed	Requires licensed Revit instance per extraction; not scalable
.pln (ArchiCAD)	GDL + ArchiCAD API	Proprietary	Same barrier as Revit
.ifc	IfcOpenShell, xBIM, IFCjs	Limited public datasets	Parseable but inconsistent (see Barrier 3); small training corpus
3D mesh (.obj, .gltf)	Standard 3D libraries	Broad availability	No semantic data (no element types, properties, or relationships) — not BIM
COBie (spreadsheet)	CSV / Excel parsers	Some public datasets	Asset data only; no geometry

Why this is hard to solve in code:

The companies that own the training data (Autodesk, Nemetschek, Trimble) are also the companies building AI products. This creates a structural data moat: any external AI company trying to train on real BIM data must either (a) do deals with every major AEC firm to access their proprietary model libraries, (b) use IFC exports that lose significant data in translation, or (c) synthetically generate training data — a hard problem because synthetic BIM models must be architecturally plausible, structurally valid, and MEP-coordinated to be useful. Synthetic BIM data generation is itself an active research problem.

Barrier 6: The Hallucination Problem Is Catastrophic in Construction — Not Just Embarrassing

In text-based AI applications, hallucination — the model producing confident but wrong output — is an inconvenience. A chatbot that gives wrong information can be corrected in the next message. In construction, hallucination has direct physical and financial consequences.

Consider what an AI-hallucinated BIM model error looks like in practice:

// AI generates duct routing through Zone B
// AI assigns: duct_diameter = 400mm (hallucinated — should be 500mm per design spec)
// AI assigns: fire_damper_required = False (hallucinated — required by code at this boundary)
// AI assigns: support_spacing = 2500mm (hallucinated — spec requires 1800mm for this weight)
//
// If uncaught by human review:
// → Fabrication order placed for 400mm duct
// → Fire damper omitted from installation
// → Insufficient supports: duct sags, creates clash with adjacent services
// → Discovered during commissioning inspection
// → Rectification cost: $180,000–$450,000 (rip-out, re-fabricate, re-install)

The problem is not merely that AI makes mistakes. It is that current AI architectures have no reliable mechanism to distinguish between what they know confidently and what they are guessing. A transformer model does not output “I’m not sure about this fire damper requirement.” It outputs a value with the same confidence it uses for everything else. Calibrated uncertainty in domain-specific technical output is an open research problem in AI safety, not a feature available in current commercial models.

Why this is hard to solve in code:

Solving hallucination in a BIM context requires the AI to have a verifiable ground truth to check against — the building code, the structural specification, the fire safety standard. Retrieval-augmented generation (RAG) partially addresses this by grounding responses in documents, but RAG systems still hallucinate when the retrieved context is ambiguous or when the required inference crosses multiple documents simultaneously. A production BIM AI would need verifiable constraint checking against the full specification envelope of a given project — essentially a formal verification layer on top of the generative model, which no commercial system provides.

Part 3: What Architecture Could Actually Work — The Emerging Research Directions

The barriers above are specific. That means the solutions are also specific. Here is where the research community is actually making progress toward a system that could generate coordinated BIM models.

Research Direction	What It Addresses	Current State	Key Players	ETA to Production
Graph Neural Networks for BIM	Barrier 1 (relational structure); Barrier 4 (global context)	Research papers (ETH Zurich, TU Delft); no production BIM system yet	Academic; Autodesk Research; Bentley iTwin	5–8 years
Neurosymbolic AI (neural + constraint solver hybrid)	Barrier 1 (constraints); Barrier 6 (hallucination verification)	Early prototypes in academic settings; no AEC commercial deployment	MIT CSAIL, CMU, DeepMind	8–12 years
Differentiable 3D CAD (geometry as differentiable program)	Barrier 2 (precision geometry); Barrier 1 (constraints)	Active research; OpenSCAD-ML hybrids; SketchGraphs (MIT); CAD-as-code approaches	MIT, Stanford, nTopology	4–7 years (component level)
Standardized BIM training corpora	Barrier 3 (IFC hostile format); Barrier 5 (data moat)	BuildingSMART openBIM datasets; UK NDTP; Singapore OneMap 3D; Autodesk Tandem (proprietary)	buildingSMART, UK CDBB, NUS Singapore	3–5 years (dataset); more to trained model
Multi-agent BIM coordination systems	Barrier 4 (context window); multi-discipline coordination	Prototype-stage; each discipline agent operates on partial model; shared state via graph DB	Autodesk Research; academic startups	5–10 years
LLM + BIM API orchestration (current best path)	Partial solution: task automation without geometry generation	Revit Copilot, Hypar; LLMs calling BIM API functions via tool-use	Autodesk, Hypar, multiple startups	Now (limited scope)

Key Takeaways

›	AI construction tools exist across seven functional categories and are genuinely useful — but every tool automates a slice of the workflow, and none of the slices connect into a unified pipeline
›	A BIM model is a relational constraint graph, not an image — current neural AI architectures are optimized for pixel/token prediction and lack native support for constraint satisfaction, making BIM generation fundamentally different from image generation
›	The IFC format, the primary open BIM standard, is structurally hostile to AI training: recursive entity references, GUID instability, schema fragmentation, and vendor extensions make building a clean training corpus an engineering project in itself
›	A complete hospital BIM model would require 200 million to 2 billion tokens to represent — 200 to 1,500 times larger than the largest available LLM context windows in 2026
›	AI hallucination in BIM is not a minor inconvenience — a single hallucinated attribute (wrong duct size, missing fire damper) that reaches fabrication can cost $180,000 to $450,000 to rectify
›	The most promising near-term path is LLM + BIM API orchestration (automating tasks without generating geometry) while the deeper research problems — graph neural networks, neurosymbolic AI, differentiable CAD — mature over 5–12 years
›	The company that solves the BIM-as-constraint-graph problem — not the BIM-as-image problem — will have built something that looks less like a design tool and more like a domain-specific reasoning engine for the built environment

About This Article

This article is part of The AEC Intelligence Report, examining AI, data standards, and digital workflows in architecture, engineering, and construction. Software capability assessments are based on publicly documented features and independent technical analysis. This article does not constitute software purchasing advice.

출처 / Sources

› Autodesk — Revit API documentation, Construction IQ documentation, Spacemaker/Forma product pages (2024–2026) — aps.autodesk.com
› BuildingSMART International — IFC specification (IFC4 ADD2 TC1; IFC4.3); openBIM implementation guides — buildingsmart.org
› IfcOpenShell — open-source IFC parser documentation — ifcopenshell.org
› Solibri — Model Checker product documentation — solibri.com
› TestFit — product documentation and API reference — testfit.io
› Cove.tool — technical documentation and ML surrogate model methodology — cove.tools
› Alice Technologies — construction scheduling AI documentation — alicetechnologies.com
› OpenSpace — SLAM and CV methodology overview — openspace.ai
› Hypar — platform documentation and developer guides — hypar.io
› UpCodes — AI product documentation — up.codes
› ETH Zurich — Graph Neural Networks for BIM research (Bischoff et al., 2023) — ethz.ch
› MIT CSAIL — SketchGraphs: A Large-Scale Dataset for Modeling Relational Geometry in Computer-Aided Design (Seff et al., 2020) — arxiv.org/abs/2007.08506
› Autodesk Research — generative design and AI in AEC publications — autodesk.com/research
› Togal.AI — product documentation — togal.ai
› Archistar — product documentation — archistar.ai