PRACTITIONER REVIEW

Hypar: Two Practitioner Reviews Tell the Whole Story

Esperanza Harper, LS3P — Used Hypar on a K-12 school project. Reduced the building footprint to preserve existing trees on site. Six hours of work versus half a week in Revit. “Hypar isn’t about replacing design judgment; it’s about removing friction.”

Martyn Day, AEC Magazine — Pressed CEO Ian Keough on liability: “Who is liable when AI makes errors on healthcare schemes?” Keough admitted “the bits are not totally connected yet” and “it’s very, very hard to design anything real through a chat prompt.”

The gap between the practitioner experience (real acceleration, real value) and the platform limitations (liability unresolved, chat-based design unreliable) is the story of every AI design tool in 2026.

LS3P on LinkedIn  ·  AEC Magazine

VISUALIZATION

AI Visualization: The Phase Problem

Joël Feyaerts of Blacksquid/CGarchitect puts it precisely: “An AI render in the concept phase is exceptional. An AI render in the sales phase, with broken balconies and inconsistent rhythm, is not a workflow. It is a deferred problem.”

Real acceleration is 20–35%, not the 80–90% in vendor marketing. That number comes from Hernán Rodenstein at Spline Dynamics, who actually measured it. Dan Cumberland documented a firm that lost a pursuit because the AI rendering “looked too perfect”—the client found it uncanny.

The uncanny valley cuts both ways. Too rough and it reads as unfinished. Too polished and it reads as unreal. Phase-appropriate fidelity is the skill that matters now.

Spline Dynamics

TOOLS

Claude Code for Architecture: What It Can and Cannot Do

Archgyan tested Claude Code on real architectural workflows. Room schedule export for a 500-room hospital: 15 minutes versus a full day. Spec cross-referencing across documents. Writing pyRevit scripts without knowing Python. The constraint: it cannot open .rvt files directly.

Separately, ALPA published an open-source toolkit for architects: 7 agents and 36 skills covering NYC zoning lookup, CSI spec generation, EPD parsing, and sustainability analysis. The toolkit is live on GitHub.

Archgyan Guide  ·  ALPA Toolkit on GitHub

PRACTITIONER REVIEW

The Honest Verdict

GXGprojects, writing in Forma Magazine, delivered the most category-by-category honest assessment published this year:

Concept generation: “Without a doubt, the top tool.”

Renders: “Bizarre lack of real-world context.”

Floor plans: “If you value precision, keep AI far away.”

Cost estimates: “Very low ceiling on what you can actually trust.”

The pattern is consistent across every practitioner review: AI excels at the fuzzy-early and struggles at the precise-late. That is not a temporary limitation. It is the architecture of these systems.

Forma Magazine

RISK

The Insurance Signal Nobody's Talking About

Verisk has standardized AI exclusion forms CG 40 47 and CG 40 48, effective January 1, 2026. Berkley introduced “absolute” AI exclusion endorsements that name specific tools: ChatGPT, Midjourney, DALL-E. Dan Cumberland is the only practitioner who has surfaced this for the architecture profession.

If your firm is using AI for deliverables, check your E&O renewal language. You may already be uninsured for that work.

This is the most consequential item in this edition. Exclusion endorsements are appearing in renewals right now, and most firms are not reading the language carefully enough to notice.

Dan Cumberland Labs

BUSINESS MODEL

The Billing Model Crisis Is Quantified

The AIA asked publicly. AECOM's CEO named it on an earnings call. The math is straightforward: 20–40% efficiency gains under time-and-materials billing equals 20–40% less revenue. Smaller firms on fixed-fee contracts have structurally solved this—they keep the efficiency gains as margin. Larger firms billing hourly have not.

The firms that moved to value-based pricing years ago for unrelated reasons now find themselves with the only billing model that survives AI acceleration. Sometimes the right answer arrives before the right question.

VENTURE CAPITAL

“Services as Software” Is a Live VC Thesis

Sequoia and a16z are not betting on selling tools to architects. They are betting on firms that perform the work at AI speed and sell outputs directly. Capital is actively deploying against the traditional AEC service model.

The competitive threat is not that your tools get better. It is that someone else uses better tools to do your job faster and cheaper, and sells the result as a product rather than a service.

MARKET

Preconstruction Is the New Startup Battleground

Design-Build is projected at 47% of US construction spend by 2028. Preconstruction workflows have expanded from 6 linear processes to 40+ non-linear ones. There is no clear software winner in this space, which is precisely why capital is pouring in.

ConTech Roundup, May 25

WORKFORCE

Trades Adoption Gap: 90% Say Indispensable, 8% Using

The gap between intent and adoption in the trades is not a willingness problem. It is a training problem. NABTU partnered with Microsoft. DEWALT and NavigateAI are attacking it simultaneously from the tool-maker side. The willingness is there. The infrastructure to act on it is not.

This is the most actionable gap in the industry. The demand signal is unambiguous. The training supply is nascent. Someone is going to build the dominant trades AI training platform, and right now nobody owns it.

ARCHITECTURE

The Single Most Important Rule

Anthropic's canonical advice: “The most successful implementations weren’t using complex frameworks. They were building with simple, composable patterns.” Start with a single LLM call with retrieval. Only add agent scaffolding when you genuinely need it.

Every framework you add is a dependency you maintain. Every abstraction layer is a debugging surface. The bar for adding complexity should be high.

Anthropic Engineering

FAILURE MODES

The 88% Problem

88% of agent projects never ship. The failure mode is remarkably consistent: broad mandate, good first demo, scope expands, evaluation deferred, edge cases emerge, token bills climb, and by month five it is a “research project” nobody uses.

The fix: write 200+ test cases before writing agent code. If you cannot define what “correct” looks like for 200 inputs, you do not understand the problem well enough to automate it.

Banana Labs

ECONOMICS

The 100x/Month Rule

A use case worth building an agent for needs to happen 100+ times per month. Below that threshold, the engineering cost does not pencil out.

In AEC, what qualifies: RFI logging. Submittal tracking. Daily report generation. What does not: one-time bid preparation. The distinction is frequency, not complexity.

Team400

RISK

Silent Failure Is the Real Risk

Agents that crash are fixable. Agents that return “plausible but wrong” answers are dangerous. In AEC, a confident hallucination about a code section or a cost figure looks exactly like a correct answer until someone builds it wrong.

The dangerous failure mode is not the agent that stops working. It is the agent that keeps working and is quietly wrong. Every production agent needs a verification layer that is independent of the agent itself.

Dev.to

CONCEPTS

Context Engineering > Prompt Engineering

The 2025–2026 conceptual shift. The question is no longer “what is the right prompt?” but “what configuration of information is in the context window at each step?” For a submittal review agent processing a 400-page spec, context management is the entire game.

Prompt engineering optimizes a single interaction. Context engineering optimizes the information architecture around every interaction. The difference is the difference between writing a good sentence and structuring a good argument.

Anthropic Engineering

CASE STUDY

Procore + Level 10 Construction

Submittal review agent: 10 reviews completed in 1 hour, a process that previously took 12 hours. The agent caught a critical error that would have caused a week of delays on site.

Procore Press

CASE STUDY

Turner SafeT Coach

AI safety consultant trained on Turner's EHS framework with tens of thousands of site interactions. Now being released to the broader industry as both a free and premium offering. Safety AI as a public good.

Turner is the largest general contractor in the US. When they open-source their safety AI, that is an industry-shaping decision, not a marketing gesture.

Construction Dive

CASE STUDY

Kreo Caddie

Autonomous estimating agent. The key insight was unexpected: the bottleneck was not AI accuracy. It was humans re-doing work the AI had already completed in a different format. Half of estimator time was wasted on copy-paste between systems.

The lesson generalizes. In many AEC workflows, the bottleneck is not intelligence but interoperability. The agent does not need to be smarter. It needs to output in the format the next system expects.

Kreo

CASE STUDY

Build.inc — Real Estate Site Screening

Multi-agent orchestration running overnight across entire geographies. “Running the junior analyst work while we sleep.” A site screening task that previously took days was completed overnight.

Build.inc

WARNING

ServiceNow's Warning

ServiceNow deployed agents internally across IT, HR, and finance. 90% of IT tickets were resolved autonomously. But unchecked agent proliferation created “token cost spirals”—costs ballooned from uncoordinated agents making redundant calls. The fix: they built an AI Control Tower for governance.

The success story and the warning are inseparable. Agents work. Ungoverned agents are expensive. Every organization deploying multiple agents needs a coordination layer, and most do not have one yet.

CX Today

DATA

The Macro Numbers

Median hours saved: 6.4/week per worker (up 64% year-over-year).

Cost per AI support ticket: $0.46 vs. $4.18 human.

Cost per AI code review: $0.72 vs. $48 senior engineer.

But: 88% of pilots never reach production.

And: 19% of deployments never reach payback.

The unit economics are compelling. The completion rate is not. The gap between pilot and production is where most of the value is lost.

REGIONAL

Three Hyperscale Data Centers Within 25 Miles of Downtown

Birmingham is becoming a hyperscale data center corridor. Three facilities are now in various stages of development within 25 miles of downtown:

Nebius Oxmoor — $40M in permits filed. Hoar Construction building.

Project Marvel, Bessemer — $14B total investment.

Columbiana campus — $1.1B. Digi Power X and Cerebras partnership.

The South now accounts for more than 50% of the $73.1B national data center pipeline, per ConstructConnect (May 27).

Bham Now

REGIONAL

Auburn Notes

BuilderHelp startup launching June 1 out of Auburn.

RFID Lab groundbreaking May 21 — Goodwyn Mills Cawood as architect, Bailey-Harris as general contractor.