A Survey Team Used AI to Classify Utilities From Drone Lidar in 3 Hours

A Drone, a Point Cloud, and the Question Every Firm Is Asking

AI is used in architecture across six categories today— visualization, generative design, code compliance, document automation, drone lidar and site analysis, and project management. Each category sits at a different point on the maturity curve, and most firm leaders cannot tell them apart.

Picture the workflow most survey teams are now running. A drone flies a corridor at dawn, captures tens of millions of lidar points in a single pass, and an AI classification model sorts those points into ground, vegetation, structures, and utility lines before the operator is back at their truck. Work that used to take weeks of desktop classification collapses into hours of automated processing with human review. FLAI and AISPECO report cutting overall lidar classification time "from months to days," with roughly 50% time savings in urban environments⁸.

That scene is real, but it is only one slice of how AI is actually used in architecture in 2026. Point cloud classification is the most quantifiably mature AI use case in architecture today— and the one most articles ignore. AI adoption in architecture is no longer a slide-deck future. It is a real, uneven, category-by-category transformation.

McKinsey estimates AI can boost construction productivity by up to 20%, reduce costs by up to 15%, and improve project delivery times by up to 30%⁵. Those are upper bounds, not averages. Before we map the categories, here is where the industry actually is.

Where AI in Architecture Actually Is in 2026

59% of UK architecture practices now use AI, up from 41% the year before, according to the Royal Institute of British Architects¹. In the US, the American Institute of Architects² found that 8% of firm leaders have integrated AI into practice, 20% are implementing, and 35% are considering adoption. Globally, 64% of architects have experimented with AI but only 20% have fully embraced it in their workflow, per a 2026 Chaos and Architizer survey⁴.

The headline numbers mask a much shallower reality. Most architecture firms have tried AI. Few have integrated it into a working week.

The firm-size gap is the second number that matters. Large UK practices report adoption rates above 80%, while smaller studios sit at 48%¹. In the US, AIA² reports that firms with 50 or more employees lead adoption. That gap will compound if mid-market firms wait.

85% of architects using AI report efficiency gains⁴. The reported gains are concentrated in concept design, image generation, and text drafting— the categories where the tools are easiest to try. This is the precision the rest of this article needs: "experimented with" (a tried chatbot, a few mood boards in Midjourney) is not the same as "integrated into the production workflow" (a daily-driven tool the team relies on to ship deliverables).

Most of that adoption is happening in one category first. Here are the six in order of practical maturity.

Category 1: Visualization and Image Generation (Where Most Firms Start)

Visualization is where most architecture firms touch AI first, and it is the most prevalent use of AI in architecture today. 48% of architects report their biggest time savings in concept design and ideation, and 43% identify concept and pre-design as the area of highest AI impact⁴.

The tools in this category sort cleanly:

• Text-to-image generators: ChatGPT, DALL-E, Midjourney for mood boards and concept ideation
• Architecture-specific tools: Veras, EvolveLab Glyph for stylized rendering inside CAD/BIM environments
• Enterprise-grade real-time: NVIDIA Omniverse integrated into design toolkits (used by Zaha Hadid Architects)
• Chatbots for text: 79% of US architects report using chatbots³, primarily for narrative drafting and research

Image generators are the gateway tool of AI in architecture— easy to try, fast to evaluate, hard to integrate into a stamped deliverable. Zaha Hadid Architects uses DALL-E and Midjourney for early-stage design ideation alongside NVIDIA's Omniverse Platform inside its ZSPACE design toolkit¹³. These are the tools to know exist; they are also the tools that earn the least defensible competitive advantage.

The demographic gradient inside firms is real. Two-thirds of architects under 35 use AI image generators, compared with 55% of those aged 36 to 50 and 41% of those over 50³. If you are a principal trying to read your firm's AI temperature, the under-35 cohort is already there.

The honest caveat: only about 70% of architects feel AI visuals reflect their design intent "reasonably" well⁴. That 30% gap is the iteration tax, and it is why image generators have not replaced the visualization team. For founders thinking about how AI fits a broader investment strategy, our AI strategy services for firms mapping these decisions lay out the same maturity logic.

Visualization shows what a project could look like. Generative design starts asking which version best meets the constraints.

Category 2: Generative Design

Generative design uses AI to produce and evaluate many design options against constraints like zoning, daylight, site geometry, and operational carbon. Autodesk reports these tools can reduce preliminary design time by 30 to 50 percent⁶.

Generative design is not "AI drawing a building for you." It is AI evaluating thousands of design variants against your project constraints in the time it takes you to make coffee. This is a different family of tools from the image generators in the last section, and the confusion in casual coverage is part of why firm leaders feel lost. If you want a clean separation of terms, what generative AI actually is covers the distinction.

The tools in this category include:

• Autodesk Forma (the dominant platform, with a Revit two-way bridge)
• Maket (residential planning and zoning)
• Finch (early-stage massing and layout)
• Architechtures (residential building generation)
• TestFit (site planning and feasibility)
• Snaptrude and qbiq (collaborative early-stage modeling)

If you only learn one platform in this category, learn Forma. Autodesk Forma now embeds Neural CAD for Buildings— Autodesk's first AEC-focused foundation model (an AI trained specifically on the patterns of building design)— and any change made in Forma is visible in Revit and the reverse⁷. That two-way Revit bridge is what makes Forma production-relevant rather than just exploratory. Daylight, sunlight, and operational carbon analysis run inside the same environment.

This category is moving fast. A firm that ignored it 18 months ago can no longer credibly say it has mapped the AEC AI landscape.

Both visualization and generative design help upstream. Now: where AI still struggles.

Category 3: Code Compliance, Specifications, and the Inefficiency Gap

Code compliance and specification writing are where architects most want AI help— and where AI tools have made the least progress. AIA³ research found that current AI adoption among architects does not align with the tasks they find most inefficient: cost estimation, project takeoffs, product libraries, and technical specifications.

AI is most adopted where it is easiest, not where it would help most. The AIA 2025 finding is unambiguous on this.

Tools are emerging:

• Codes.IQ (US building code research and verification)
• CodeComply (automated compliance checking)
• Ichi (code-aware design assistance)
• EvolveLab Glyph (documentation automation alongside its rendering features)

Why is this category lagging? Code compliance demands citation accuracy and liability-grade reasoning. Large language models hallucinate. Specification sections require traceability to product data and standards that change. 94% of US architects cite inaccuracy as a top concern about AI³, and the spec/code arena is the one where that concern bites hardest.

What firms can do today: use LLMs to draft the narrative sections of specifications with mandatory human review. Do not let an LLM autonomously cite codes. If your firm's biggest bottleneck is specifications, the AI category that solves it is still emerging— Codes.IQ, CodeComply, and Ichi are the names to watch through 2026.

Now to the category that earns its "mature" label.

Category 4: Drone Lidar and Point Cloud Classification (The Most Mature AI Use in AEC)

AI-based point cloud classification is the most quantifiably mature AI category in architecture today. Specialized deep learning models (KPConv, PointNet++, and transformer architectures) classify lidar point clouds into ground, vegetation, structures, and utility features. FLAI and AISPECO⁸ report this reduces processing time from months to days, enabling a survey-to-site-plan handoff that used to take weeks.

A single Leica RTC360 scan generates roughly 200 million points in 1.5 minutes; mobile mapping systems can analyze kilometers of motorways in minutes versus the days required by manual workflows¹⁰. The US Geological Survey⁹ uses transformer-based deep learning to classify points across its national 3DEP lidar dataset— a federal-scale signal that this is no longer a pilot category.

The drone-to-site-plan workflow now reads like this:

1. Drone with lidar payload flies the corridor or site
2. Raw point cloud is exported (hundreds of millions of points typical)
3. AI classification model sorts points by semantic category automatically
4. Human reviewer audits classifications against site conditions
5. Classified cloud feeds the architect's existing-conditions model, site plan, and BIM

A representative modern stack:

Hardware is keeping pace. The DJI Matrice 400, launched in June 2025¹¹, gives a survey team:

• 59 minutes of flight time (a single battery covers most site flights)
• 6 kg payload capacity (room for a heavy lidar + camera package)
• 520,000 points per second from its integrated rotating lidar
• 360-degree horizontal field of view

Pair that with a Leica RTC360 ground scanner and FLAI's classification pipeline, and a survey team can deliver a classified existing-conditions model in a fraction of the time the same job took two years ago.

A practical caveat is worth saying out loud: training data locality matters. In one widely shared example, Las Vegas palm trees were mistaken by an AI classifier for columns until the model was retrained on local vegetation. Field accuracy is not benchmark accuracy. Specs and tolerances are still your responsibility.

Why this matters for the architect: site plans, existing-conditions models, utility location, and topographic baselines all start with point cloud data. AI classification feeds the architect's design directly— and it does so today, in production, with measurable savings.

Outside the survey trailer, AI is reshaping the back office too.

Category 5: Document and Project Automation

Document and project automation is where AI delivers measured ROI today. Bluebeam's 2025 industry survey¹² found that 68% of early AEC adopters of AI have saved at least $50,000, and 46% have saved between 500 and 1,000 hours, primarily through AI-assisted document review, RFI triage, contract automation, and project reporting.

Document automation rarely makes the demo reel. But it is where early AEC adopters are reporting the largest dollar savings.

What is actually in this category:

• RFI / submittal review and triage (LLMs summarizing inbound questions, flagging ambiguities)
• Contract automation (clause extraction, redlining assistance)
• Invoice and budget triage (anomaly detection, line-item categorization)
• Project reporting (status drafting from BIM and PM data)
• AI-augmented meeting notes (25% of US AIA respondents already use transcription or meeting assistants³)

The flip side is the underinvestment story. Only 27% of AEC firms currently use AI for automation, problem-solving, or decision-making¹²— meaning the operational opportunity is still wide open for any firm willing to invest first. Lack of skilled personnel is the most-cited barrier. Two thirds (65%) of AEC firms invest less than 10% of their technology budgets on training¹², which is the line item most likely to cap a firm's return.

The optimism is real. 84% of architects say they are optimistic that AI can automate manual tasks to save time³. The gap between that optimism and the 27% actually using AI for automation is exactly the opportunity a mid-market firm can capture by sequencing investments deliberately. The hidden costs of AI projects we see most often covers why under-trained teams undercut tool ROI faster than any other failure mode.

Roll those operational gains up to the firm level, and the picture gets larger.

Category 6: Project Management, Analytics, and the McKinsey Numbers

At the project portfolio level, McKinsey⁵ estimates AI can increase construction productivity by up to 20%, reduce costs by up to 15%, and improve project delivery times by up to 30%. Approximately $50 billion was invested in AEC technology between 2020 and 2022, an 85% increase over the preceding three years⁵.

McKinsey's 20%/15%/30% figures are the most-cited stats in AEC AI. They are upper bounds, not averages. Always cite them as "up to." The $50 billion in AEC tech investment marks an 85% jump over the prior period— the capital has already shifted.

What this category covers in practice: portfolio analytics, schedule risk prediction (nPlan-style platforms), resource allocation across projects, and cost forecasting. Most firms do not see the upper-bound numbers, and the reason is straightforward: integration cost and the training gap. When 65% of firms invest less than 10% of their tech budget on training¹², the tools may be in place but the team is not. This is where how to measure AI success in your firm becomes the question that separates real ROI from line-item waste.

That is the map. Now the honest part.

The Honest Limits — Where AI in Architecture Still Falls Short

94% of architects cite inaccuracy as their top concern about AI, with unintended consequences (94%), privacy and security (93%), authenticity (90%), and lack of transparency (90%) close behind, according to AIA 2025 research³. These are not abstract worries. They are the reason AI adoption has clustered in low-stakes categories first.

AI in architecture is real, uneven, and easy to oversell. The firms that lead are the ones honest about what AI still does not do.

Output quality is the next limit. 48% of architects cite inconsistent output quality as the biggest challenge of working with AI, and 33% cite poor software compatibility as a key integration obstacle⁴. Both are practical bottlenecks. Inconsistent output means more hands-on review. Compatibility friction means tools that look powerful in a demo run aground against your actual Revit-plus-Bluebeam-plus-PM-stack reality.

Authenticity and intellectual property are the third concern, and they are particularly sharp in design-led practices. Two-thirds of UK architects (67%) are concerned AI increases the risk of their work being imitated— a creative IP question the legal frameworks have not caught up with yet¹.

The reassuring number from the same RIBA report: only 4% of UK architects believe AI will eliminate the need for creativity in design, and only 18% anticipate job losses from AI adoption¹. The existential framing is not the practical one. The practical framing is where to put AI to work first.

The firms that lead are honest about these limits— and act anyway, in a deliberate sequence. Here is the sequence for a mid-market AEC firm.

What This Means for a $20M–$100M AEC Firm

Mid-market AEC firms— principals and COOs running practices between $20M and $100M in revenue— sit in the strategic window. Larger firms lead adoption (above 80% in the UK¹, and US firms with 50+ employees lead²) and smaller studios lag (48%¹). Mid-market firms can leapfrog by sequencing investments where the ROI is most quantified— not where the demos are most exciting.

The right sequence for a mid-market AEC firm: visualization first, point cloud and site analysis next, document automation third, generative design fourth, code and specs last.

Training is the gating constraint, not the tool budget. 65% of firms invest less than 10% of their tech budget on training¹². Under-investment in training will cap any tool investment, regardless of vendor. This is the line item to defend in next year's budget.

The honest reframe: only 4% of UK architects believe AI will eliminate creativity in design¹. The practical question is not whether AI replaces the architect. It is where the firm puts AI to work first. AI amplifies the architect's expertise. It does not substitute for the architect. The firms that get this right invest in the architect's thinking, not just the tools. The AI decision framework for founders we use walks the same logic step by step, and what a fractional AI officer does covers the role that typically owns sequencing at this firm size.

Sorting these categories into a sequence that fits your firm's projects, team, and existing software stack is the work a fractional AI officer or implementation partner does in the first 60 to 90 days. If mapping the right AI investments to your firm's specific bottlenecks would help, Dan Cumberland Labs works with AEC firms on exactly this kind of sequencing.

A few of the questions that keep coming up when firm leaders work through this:

Frequently Asked Questions

How is AI used in architecture today?

AI is used in architecture across six categories: visualization, generative design, code compliance and specifications, document and project automation, drone lidar and site analysis, and project management analytics. Visualization is the most common entry point. AI-based point cloud classification is the most quantifiably mature use case, with FLAI⁸ reporting overall lidar processing reduced "from months to days" and the USGS⁹ using transformer-based deep learning across its national 3DEP dataset.

What percentage of architecture firms use AI?

59% of UK practices use AI as of 2025, up from 41% the year before¹. In the US, 8% of firm leaders have integrated AI into practice, 20% are implementing, and 35% are considering adoption². Globally, 64% of architects have experimented with AI but only 20% have fully embraced it in their workflow⁴.

Can AI replace architects?

The RIBA 2025 report found that only 4% of UK architects believe AI will eliminate the need for creativity in architectural design, and only 18% expect job losses from AI adoption¹. Design intent, stakeholder judgment, ethical responsibility, and stamped accountability remain human work. AI amplifies the architect's expertise rather than substituting for it.

What does generative design actually do?

Generative design uses AI to produce and evaluate many design options against constraints like zoning, daylight, site geometry, and operational carbon. Autodesk reports these tools can reduce preliminary design time by 30 to 50 percent⁶. Autodesk Forma now embeds the Neural CAD for Buildings foundation model and connects two-way with Revit⁷— a step that moves generative design from exploratory tool to production workflow.

How does AI classify drone lidar point clouds?

Specialized deep learning models (KPConv, PointNet++, and transformer architectures) classify lidar points into categories like ground, vegetation, structures, and utility lines automatically. FLAI reports cutting manual classification time by approximately 50% in urban environments and 25% in rural environments, reducing overall processing from months to days⁸. The USGS uses a transformer-based pipeline on its national 3DEP lidar dataset⁹, and Esri's ArcGIS Pro ships a pre-trained transmission power line model out of the box¹⁴.

What are AI's biggest risks in architecture?

94% of US architects cite inaccuracy as their top concern, followed by unintended consequences (94%), privacy and security (93%), authenticity (90%), and lack of transparency (90%)³. 67% of UK architects are also concerned about work being imitated¹. 48% of architects cite inconsistent output quality as the biggest practical challenge of working with AI, and 33% cite software compatibility friction⁴.

Which Category to Fund First

AI in architecture in 2026 is not one thing. It is six categories at six different maturity levels. The firms that lead are not the ones with the most tools. They are the ones with the clearest map.

AI amplifies the architect's expertise. The firms that get this right invest in the architect's thinking, not just the software. The practical question, the one worth answering before next quarter's budget, is not whether AI will replace the work you do. It is which of the six categories to fund first, and why.

References

1. RIBA, "Second RIBA AI Report Shows Surge in Usage Among UK Architects" (2025) — https://www.architecture.com/knowledge-and-resources/knowledge-landing-page/second-riba-ai-report-shows-surge-in-usage-among-uk-architects
2. AIA, "New Research Explores Perceptions and Opportunities of Artificial Intelligence in Architecture" (2025) — https://www.aia.org/about-aia/press/new-research-explores-perceptions-and-opportunities-artificial-intelligence
3. AIA, "Architects Are Excited About the Potential of AI, But Concerns Abound" (2025) — https://www.aia.org/aia-architect/article/architects-are-excited-about-potential-ai-concerns-abound
4. Chaos and Architizer, "The State of AI in Architecture: How AI Is Reshaping Architectural Design & Visualization in 2026" (2026) — https://blog.chaos.com/the-state-of-ai-in-architecture-survey-insights
5. McKinsey & Company, "Artificial Intelligence: Construction Technology's Next Frontier" (2024) — https://www.mckinsey.com/capabilities/operations/our-insights/artificial-intelligence-construction-technologys-next-frontier
6. Autodesk, "How AI in Architecture Is Shaping the Future of Design, Construction" (2024) — https://www.autodesk.com/design-make/articles/ai-in-architecture
7. Autodesk, "Autodesk Brings Design and Make Intelligence to the Built Environment with Forma Building Design" (2025) — https://adsknews.autodesk.com/en/news/autodesk-design-and-make-intelligence/
8. FLAI and AISPECO, "Evaluation of FLAI AI Classification on LiDAR Data Collected with the AISPECO Heliux LITE System and RIEGL VQ-580 II-S" (2026) — https://www.flai.ai/post/evaluation-of-flai-ai-classification-on-lidardata-collected-with-the-aispeco-heliux-litesystem
9. U.S. Geological Survey, "Automated Deep Learning-Based Point Cloud Classification on USGS 3DEP Lidar Data Using Transformer" (2024) — https://www.usgs.gov/publications/automated-deep-learning-based-point-cloud-classification-usgs-3dep-lidar-data-using
10. Geo Week News, "The Essential Role of AI in Point Cloud Classification" (2024) — https://www.geoweeknews.com/blogs/lidar-ai-point-cloud-classification-artificial-intelligence
11. DroneLife, "DJI Launches Matrice 400 Enterprise Drone with 59-Minute Flight Time" (2025) — https://dronelife.com/2025/06/10/dji-launches-matrice-400-enterprise-drone-with-59-minute-flight-time/
12. Bluebeam (Nemetschek), "New Bluebeam Report Shows Early AI Adopters in AEC Seeing Significant ROI Despite Uneven Adoption" (2025) — https://press.bluebeam.com/2025/10/new-bluebeam-report-shows-early-ai-adopters-in-aec-seeing-significant-roi-despite-uneven-adoption/
13. Wallpaper, "AI in Architecture: Zaha Hadid Architects on Its Pioneering Use and Collaborating with NVIDIA" (2024) — https://www.wallpaper.com/architecture/zaha-hadid-architects-nvidia-ai-in-architecture
14. Esri, "Classify Transmission Power Lines in Point Clouds Using Deep Learning— ArcGIS Pro" (2024) — https://pro.arcgis.com/en/pro-app/latest/help/analysis/3d-analyst/classify-transmission-powerlines-in-point-clouds-using-deep-learning.htm