We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    The AI-Native Project: The Enterprise Model Stack for Construction and Infrastructure

    A practical map of site speech, progress intelligence, document RAG, schedule risk, resource optimization, safety, equipment, and commercial controls.

    Conscious Engines

    Construction produces enormous amounts of information but rarely a single current version of reality. Drawings change. Site conditions differ from the model. Progress is estimated. Field observations arrive late. Contracts, requests for information, change orders, schedules, and subcontractor records live in separate systems.

    This makes construction a strong market for task-specific models. Speech captures field truth. Vision measures installed work. Document models extract obligations. RAG retrieves the controlling version. Forecasts identify schedule and cost risk. Optimization helps allocate labor and equipment.

    The market gap

    The RICS Artificial Intelligence in Construction Report 2025 received more than 2,200 responses. It found:

    • 45% reported no AI implementation.
    • 34% were in early pilot phases.
    • Just under 12% used AI regularly in specific processes.
    • Only 1.5% used AI across multiple processes.
    • Fewer than 1% reported organization-wide scaled use.
    • 46% cited lack of skilled personnel as a top barrier.
    • 37% cited integration and 30% cited data quality or availability.

    Respondents ranked progress monitoring and scheduling highest for potential impact at 36% each, followed by resource optimization and contract or document review at 30% each. The market is interested but under-instrumented.

    Evidence from projects

    A Buildots case involving NCC reports 70% less time spent on manual reporting and a 2.3 times increase in tasks completed on time after AI-based progress tracking. NCC expanded the pilot from an 8,000 square-meter residential project to three additional projects covering 68,500 square meters. An NCC service manager said: "We spend less time running around the site."

    On a £105 million NHS rehabilitation project, Sir Robert McAlpine and VINCI's IHP joint venture used 360-degree site imagery against BIM and schedule data, then combined progress with biometric turnstile information to examine labor deployment.

    These are company and vendor reports, not controlled experiments. They demonstrate a deployable pattern: objective progress data linked to schedule and workforce decisions.

    The construction model stack

    The construction model stack

    Voice-first field reporting

    Supervisors, engineers, inspectors, and trades can dictate daily reports, observations, delays, quantities, and actions.

    Progress and reality intelligence

    Vision models compare 360-degree images, drones, or scans with BIM and planned work. The system estimates installed quantities and flags deviations.

    Drawing, specification, and contract RAG

    A permission-aware knowledge layer retrieves the current drawing, specification, method statement, contract clause, RFI, submittal, and change instruction.

    Schedule and cost risk

    Forecasting models can predict activity delay, productivity variance, cash flow, and cost at completion.

    Resource and equipment optimization

    Optimization can allocate crews, cranes, plant, deliveries, and work fronts under dependencies and safety constraints. The model proposes feasible options.

    Safety intelligence

    Speech, vision, and document models can capture observations, identify required controls, retrieve method statements, and prioritize leading indicators.

    Voice-first field reporting

    Supervisors, engineers, inspectors, and trades can dictate daily reports, observations, delays, quantities, and actions. Domain ASR must recognize subcontractors, locations, drawing numbers, materials, equipment, and measurements. An SLM converts speech into the project schema and asks for missing fields.

    Progress and reality intelligence

    Vision models compare 360-degree images, drones, or scans with BIM and planned work. The system estimates installed quantities and flags deviations. It should preserve capture time, location, and confidence.

    Measure capture coverage, activity-level precision and recall, reporting time, percent plan complete, delay lead time, and verified billing differences.

    Drawing, specification, and contract RAG

    A permission-aware knowledge layer retrieves the current drawing, specification, method statement, contract clause, RFI, submittal, and change instruction. It must understand document status, revision, package, location, discipline, and contractual authority.

    Schedule and cost risk

    Forecasting models can predict activity delay, productivity variance, cash flow, and cost at completion. Inputs include progress, labor, weather, approvals, supply events, and change history. Historical replay must prevent future data leakage.

    Resource and equipment optimization

    Optimization can allocate crews, cranes, plant, deliveries, and work fronts under dependencies and safety constraints. The model proposes feasible options. The project manager remains responsible for the plan.

    Safety intelligence

    Speech, vision, and document models can capture observations, identify required controls, retrieve method statements, and prioritize leading indicators. Consequential safety decisions require competent-person review.

    Commercial and claims intelligence

    Document models extract notices, dates, obligations, quantities, valuations, and change evidence. An SLM can build a draft chronology with source links. It should not decide entitlement.

    Equipment, fuel, and logistics

    Time-series and optimization models track utilization, idle time, maintenance, delivery queues, fuel receipts, and consumption. These use cases connect construction directly to industrial and logistics capabilities.

    Reference architecture

    1. Project graph: owner, contractor, package, location, activity, asset, drawing, and change.
    2. Evidence layer: field speech, images, schedules, BIM, documents, sensors, and commercial records.
    3. Specialist models: ASR, vision, document extraction, forecast, anomaly detection, SLM, and optimizer.
    4. Knowledge layer: controlled revisions, permissions, status, effective dates, and citations.
    5. Workflow layer: common data environment, scheduling, ERP, safety, quality, and field systems.
    6. Control layer: approvals, legal privilege, labor privacy, audit, and model monitoring.

    KPI framework

    Use caseModel metricProject metric
    Field voicecritical-entity accuracyreport time, missing fields
    Progress visionactivity precision and recallpercent plan complete, billing accuracy
    RAGcurrent-source recall, citation precisionsearch time, RFI cycle
    Schedule riskcalibration, lead timeprevented delay, schedule variance
    Cost forecasterror and biasforecast accuracy, contingency use
    Safetyhazard recall, false alertclosure time, repeated observation
    Equipmentutilization forecast, anomaly precisionidle time, fuel per unit

    A 90-day starting point

    Choose one active project and one workflow, such as daily reporting, document retrieval, or progress validation. Establish a baseline. Build an evaluation set with real accents, drawing numbers, superseded files, ambiguous locations, and permission tests. Run the system silently, then through a small field group.

    A useful gate might require 95% accuracy on project identifiers after confirmation, at least 95% retrieval of the controlling document, and a 30% reduction in reporting or search time without increased missing critical information.

    The conclusion

    Construction AI becomes valuable when it connects field evidence to project authority and action. The specialist model is important, but the durable asset is the project graph, controlled document history, and evaluation set.

    The strongest first sale is not autonomous construction. It is a measurable reduction in reporting, search, delay discovery, and coordination work.

    Research note

    Research is current through September 5, 2026. RICS data is survey evidence. Customer results are company or vendor reported. Contractual, safety, privacy, and labor controls require project-specific review.

    Continue the research

    Building a Production-Ready System

    Conscious Engines builds construction AI solutions around project documents, drawings, schedules, field language, progress evidence, and approval workflows. Speech, document intelligence, private RAG, forecasting, and task-specific models turn site evidence into faster decisions without replacing engineering or contractual authority.