June 13, 2026
Date: June 13, 2026 Prepared for: Steve Watts, CEO Focus: AI/Automation Trends Relevant to Commercial Glazing
This week's analysis reveals a decisive shift from AI experimentation to AI orchestration—success now depends on how companies combine models rather than which single model they choose. For PGC, the most actionable development is the maturity of on-device AI, which enables field tools that work without cloud connectivity while keeping client site data private. Spatial reasoning capabilities have advanced enough to automate site measurement and defect detection tasks that currently require manual interpretation. The strategic implication is clear: PGC should build modular AI skills for glazing-specific workflows rather than investing in monolithic systems.
What it is: AI systems built from modular "skills" or micro-capabilities that can be mixed, matched, and reused across workflows. Think of it as the AI equivalent of VS Code extensions—each skill does one thing well, and multiple skills combine to handle complex tasks.
Why it matters: Instead of building or buying a single monolithic AI system, PGC can assemble purpose-built skills for glazing tasks: one skill for reading architectural drawings, another for calculating glass dimensions, another for generating quotes. This approach reduces cost, improves reliability, and allows swapping individual components without rebuilding the entire system.
What it is: Running AI models directly on local devices (tablets, phones, laptops) rather than sending data to cloud servers. Apple's CoreAI framework and similar tools now make this production-viable for real business workflows.
Why it matters: Field technicians can use AI-powered tools on job sites without reliable internet. Site photos and measurements never leave the device, addressing client confidentiality concerns. Latency disappears—no waiting for cloud responses during site visits.
What it is: AI systems that understand and reason about physical space, objects, and their relationships—extracting measurements from photos, identifying defects on surfaces, interpreting 3D layouts.
Why it matters: Directly applicable to site surveys, glass inspection, and quality control. A tablet can photograph a window opening and automatically extract dimensions. A technician can photograph glass upon delivery and identify chips or scratches without expertise.
What it is: Architectural approaches that use expensive, capable AI models for planning and verification, then switch to cheap, fast models for routine execution. This "audit with big, execute with small" pattern minimizes token costs.
Why it matters: Dramatically reduces per-task AI costs. Planning a complex glass installation might use a capable model; generating the 47 standard emails confirming material specs uses a cheap model. PGC can automate high-volume routine tasks without budget shock.
What it is: AI tools that directly generate finished documents (PDFs, presentations, specifications) from requirements rather than requiring manual formatting or conversion.
Why it matters: PGC's proposal and specification workflow involves substantial manual document preparation. AI tools now exist that can generate client-ready PDFs and presentations directly from structured requirements, eliminating the most tedious parts of bid preparation.
What it is: A market structure emerging where small, efficient models handle the majority of production workloads while large frontier models handle only complex planning tasks.
Why it matters: The economic incentive is shifting away from ever-larger models toward optimized small models. This means reliable, predictable pricing for production AI usage. PGC should architect for small models as the workhorses.
What it is: Moving beyond statistical AI detection (which is failing) toward behavioral analysis and cryptographic methods to verify content origin and authenticity.
Why it matters: As AI-generated content becomes indistinguishable from human content, enterprise clients will demand provenance verification. PGC should prepare for requirements to prove documentation authenticity, especially on high-value projects.
What it is: Tools like "humanizers" that modify AI-generated text to pass detection systems are becoming widely available and increasingly effective.
Why it matters: Statistical AI detection is effectively dead. Any content can be made to appear human-written. This accelerates the need for provenance and verification approaches rather than detection.
What it is: Different regulatory environments and technology stacks emerging in different regions—Apple's tools work around Chinese restrictions, benchmarks now cover low-resource languages, localized stacks create different capability sets.
Why it matters: If PGC operates across regions with different data regulations, AI tools may need to function differently in each context. Local-first architecture provides flexibility regardless of jurisdiction.
What it is: Infrastructure and frameworks designed to work with whatever AI model is most appropriate—switching between providers based on cost, privacy, or capability without rewriting integrations.
Why it matters: The open-source-versus-proprietary debate is less important than tool-agnostic design. PGC should prioritize orchestration layers that can route tasks to the best available model rather than committing to a single provider.
The Edge + Spatial Reasoning Intersection
On-device AI (Trend 2) combined with spatial reasoning (Trend 3) creates the highest-value use case for PGC. Field tools that understand physical space AND run locally without cloud dependency enable site survey automation, defect detection, and measurement extraction that work in any job site condition.
The Agent Skills + Design-to-Deliverable Connection
Composable agent architectures (Trend 1) enable the design-to-deliverable pattern (Trend 5). A glazing specification generator can chain together: reading architectural drawings (one skill), extracting requirements (second skill), checking inventory (third skill), generating PDF output (fourth skill). Each component is independently maintainable.
The Cost Orchestration Reinforcing Small Models
Cost orchestration patterns (Trend 4) validate the two-tier model economy (Trend 6). As more companies implement "audit with big, execute with small," demand for efficient small models increases, driving down costs and improving availability of production-optimized models.
The Provenance + Balkanization Pressure
Both behavioral provenance (Trend 7) and regional balkanization (Trend 9) push toward local-first architecture. Organizations uncertain about data handling regulations find comfort in solutions where data never leaves their control.
PGC can validate these trends with minimal investment. Recommend starting with the following experiments:
Time required: 2-3 days Cost: Under $500 (API usage + staff time) Approach: Use a Vision Language Model (VLM) API to analyze site photos, extracting rough measurements and identifying framing dimensions. Test against manual measurements for accuracy.
Success criteria: Model provides usable estimates within 10% of actual measurements on at least 80% of test cases.
What it proves: Spatial reasoning capability readiness for site survey automation.
Time required: 1-2 days Cost: Under $200 (API usage) Approach: Take three recent glazing specifications and prompt an AI model to generate polished PDF documents from structured requirements. Compare output quality and time against manual preparation.
Success criteria: Generated documents require minimal editing and capture 90%+ of manually-prepared content.
What it proves: Design-to-deliverable automation readiness for proposal workflows.
Time required: 4-7 days Cost: Under $1,000 (device + development) Approach: Deploy a pre-trained image classification model on a tablet to identify common glass defects (chips, scratches, cracks) from photos. Test against quality control team accuracy.
Success criteria: Model matches or exceeds QC team defect identification accuracy on a standardized test set.
What it proves: Edge AI viability for quality control workflows without cloud dependency.
Implication 1: Field Tools Are Now Feasible Without Compromise
On-device AI has reached production viability. The traditional trade-off between capability and privacy/connectivity no longer applies. PGC can deploy AI-powered tools to field technicians that work offline, keep site data private, and deliver real-time insights. Recommendation: Prioritize development of field-facing AI tools over back-office automation for maximum competitive differentiation.
Implication 2: Glazing-Specific AI Skills Beat General AI
The composable agent architecture trend validates building specialized skills rather than training staff to use general AI tools. A "glass specification parser" skill, a "structural framing interpreter" skill, and a "defect classifier" skill provide more value than generic AI assistants. Recommendation: Identify the 5-7 recurring glazing tasks that consume staff time and invest in building modular skills for each.
Implication 3: Proposal Automation Offers Fastest ROI
Design-to-deliverable automation directly addresses a high-frequency, high-effort workflow. Converting structured project requirements into client-ready proposals, specifications, and presentations is repetitive and expensive. AI can handle most of this work. Recommendation: Target proposal generation as the first full workflow automation, targeting 50% reduction in preparation time.
Implication 4: Privacy Positioning Differentiates Enterprise Contracts
Enterprise clients with data sovereignty requirements (government, healthcare, financial) increasingly demand local-first solutions. PGC can differentiate on privacy by guaranteeing site data never leaves the device or client environment. Recommendation: Market local-first AI capabilities as a trust differentiator in enterprise bids.
Implication 5: Model Selection Is Less Important Than Orchestration
The two-tier model economy and cost orchestration trends suggest that PGC's AI infrastructure investment should focus on orchestration and routing rather than model selection or training. The ability to route tasks to appropriate models based on cost, capability, and privacy requirements will matter more than any single model choice. Recommendation: Evaluate AI platforms based on orchestration flexibility, not just benchmark performance.
End of Briefing