5 procedures for AI-driven fabrication, inspection, environmental monitoring, site safety, and financial verification. The first governance framework purpose-built for AI-constructed structures.
Who this is for: Construction safety managers, general contractor compliance officers, structural engineers using AI tools, building inspectors evaluating AI-fabricated components, insurance risk managers, and OSHA compliance consultants.
Prerequisite: Familiarity with OSHA 29 CFR 1926 and building code basics (IBC/IRC). No prior SWT3 knowledge required.
Assessor Notice: SWT3 witness anchors record what happened on a job site or factory floor at a specific point in time. They do not replace building inspections, licensed structural engineering review, or OSHA compliance determinations. A PASS anchor means the witnessed measurement met its configured threshold. A building inspector or licensed engineer determines whether that threshold was the right one. Evidence proves state. Judgment determines compliance.
AI is entering construction from three directions simultaneously. Robotic prefabrication factories use AI to determine stud spacing, nailing patterns, and cut geometry for wall panels that ship directly to job sites. Drone inspection platforms use vision-language models to detect safety hazards, track progress, and identify defects across millions of construction sites. And AI-powered structural analysis tools compute load calculations, seismic resistance, and material specifications that end up embedded in the buildings people live and work in.
Each of these applications generates governance questions that existing AI frameworks do not answer. NIST 800-53 covers audit trails but not concrete cure monitoring. The EU AI Act addresses high-risk AI but not whether a robotic nailing pattern meets IRC fastener spacing requirements. OSHA 29 CFR 1926 mandates frequent inspections but says nothing about how to verify that an AI-fabricated component was built to the parameters the AI determined.
The CON namespace addresses this gap directly. Five procedures cover the full lifecycle of AI-driven construction: structural verification of AI-determined parameters (CON-STR.1), inspection of AI-fabricated components (CON-INS.1), environmental monitoring at deployment sites (CON-ENV.1), site safety for AI-operated equipment (CON-SAF.1), and financial verification for AI-fabricated work (CON-FIN.1). Each procedure maps to specific OSHA citations and building code requirements.
The market context is urgent. A Construction Executive survey found that 79% of construction company boards have approved AI spending, but only 40% have any formal AI governance policy. The Robotics-as-a-Service market is valued at $2.83 billion in 2026 and projected to reach $14.82 billion by 2036. Capital is flowing in. Governance is not keeping up.
When evaluating a construction firm's AI governance posture, the first question is scope: which construction activities involve AI decisions? AI-determined structural parameters, AI-controlled robotic fabrication, AI-powered inspection, and AI-computed environmental monitoring each create different evidence requirements. The 5 CON procedures map to these categories. Ask the firm to identify which CON procedures are relevant to their operations before evaluating evidence completeness.
OSHA's enforcement model has fundamentally changed. The expanded electronic recordkeeping rule now requires construction firms to submit detailed injury and illness data electronically through the Injury Tracking Application (ITA). OSHA feeds this data into its Site-Specific Targeting (SST) program, which uses algorithms to rank worksites by risk and select them for comprehensive inspections.
This creates a feedback loop: the data you submit determines whether OSHA inspects you. Firms with higher reported injury rates, more serious incidents, or incomplete submissions are algorithmically prioritized for inspection. Construction (NAICS 236000) is classified as high-hazard by default.
| Metric | 2026 Value |
|---|---|
| Maximum penalty per serious violation | $16,550 |
| Maximum penalty per willful/repeated violation | $165,514 |
| Inspections conducted (YTD) | 32,000+ |
| Total penalties assessed (YTD) | $123 million+ |
| Construction classification | High-hazard (automatic inclusion) |
For construction firms using AI, the implication is direct: if an AI-operated system contributes to an injury or near-miss, and the firm cannot demonstrate that the AI's decisions were governed, the lack of evidence becomes an aggravating factor. OSHA does not currently have AI-specific standards for construction. Jobsite robots are governed by existing machine-guarding and lockout/tagout frameworks (Subpart O). But the absence of a specific standard does not mean the absence of liability. The General Duty Clause (Section 5(a)(1)) requires employers to provide workplaces "free from recognized hazards." An ungoverned AI system that creates a recognized hazard is a General Duty Clause violation.
When reviewing OSHA compliance for a construction firm using AI, check whether the firm's written safety program (required under 29 CFR 1926.20(b)(1)) addresses AI-operated equipment. If the safety program was written before AI was deployed on site, it likely does not. The absence of AI in the safety program is itself a finding. CON-SAF.1 anchors document that AI-operated equipment safety was assessed and certified. A continuous chain of CON-SAF.1 anchors demonstrates that the firm has an active, ongoing safety program that includes AI -- not a policy written once and filed.
Each CON procedure is an attestation-type control. Unlike automated command-probe controls that run on a server, attestation controls require witnessed evidence from a factory cell, an inspection event, or a sensor reading. The evidence comes from the physical world and is anchored in the witness ledger with the same cryptographic fingerprint as any other SWT3 anchor.
Records that a structural measurement computed or influenced by AI was verified against the applicable building code requirement. The anchor captures the code requirement threshold, the actual measured value, and the verification method (AI vision, physical sensor, or licensed inspector).
OSHA reference: 29 CFR 1926.700(a) requires determining structural capability before placing construction loads.
Building code reference: IBC Section 1604 (General Design Requirements), IRC Section R301 (Design Criteria). ASCE 7 minimum design loads apply regardless of whether the load calculation was performed by a human or an AI.
Evidence anchored: Code requirement value, actual measured value, verification method -- 1=AI vision, 2=sensor, 3=licensed inspector.
The verification method field is critical. A value of "1" means the structural verification was performed by AI vision alone. A value of "3" means a licensed inspector verified it. For load-bearing structural elements, most jurisdictions require licensed professional engineer review regardless of how the original calculation was performed. Check whether AI-only verifications (method=1) are used for structural elements that require PE sign-off. The anchor records what happened; the assessor determines whether the verification method was appropriate for the element type.
Records that an AI-fabricated construction component passed inspection against a defined checklist. The anchor captures the number of checklist items required, the number that passed, and whether the inspector's credentials were verified.
OSHA reference: 29 CFR 1926.20(b)(2) requires frequent and regular inspections of job sites, materials, and equipment by competent persons.
Evidence anchored: Required checklist items, items passed, inspector license verification status.
OSHA requires inspections by "competent persons" -- individuals capable of identifying hazards and authorized to take corrective measures. For AI-fabricated components, verify that the inspector documented in the anchor has the appropriate credentials for the component type. A wall panel requires different inspector qualifications than an electrical assembly. Also check the pass rate: a component that passed 47 of 50 checklist items is not a full pass. The 3 failed items should have remediation evidence in subsequent anchors.
Records environmental conditions at the construction site that affect material performance. Concrete curing requires minimum temperature and humidity over a defined period. Steel erection has wind speed limits. Adhesive bonding has temperature ranges. When AI determines that conditions are adequate to proceed, this anchor records the evidence.
OSHA reference: 29 CFR 1926.700(a) requires adequate strength before loading.
Standards reference: ACI 318-19 Section 26.5.3.1 requires minimum curing periods. AISC 360 addresses temperature limits for welding operations.
Evidence anchored: Environmental threshold, actual reading, hours elapsed since placement.
Environmental monitoring is time-sensitive evidence. Concrete curing conditions must be maintained continuously, not just sampled once. A single CON-ENV.1 anchor at the start of the curing period is insufficient. Look for a series of anchors across the full curing window (typically 7 days for standard concrete, 3 days minimum per ACI 318). Gaps in the anchor chain during the curing period indicate periods when conditions were unmonitored. If AI made a "go" decision based on environmental data, the anchor chain should cover the entire monitoring window.
Records that workers interacting with AI-operated construction equipment have verified safety certifications. This covers fall protection training for workers near robotic arms, lockout/tagout certification for maintenance personnel, and PPE compliance verification for all workers in the AI equipment's operating zone.
OSHA reference: 29 CFR 1926.500-503 requires fall protection training and certification for workers above 6 feet. 29 CFR 1926.502(d) requires personal fall arrest systems. Subpart O (Motor Vehicles, Mechanized Equipment, and Marine Operations) applies to robotic construction equipment.
Evidence anchored: Required safety certifications, verified certifications, PPE compliance status.
OSHA's Fall Protection Focus Four lists falls as the #1 cause of construction fatalities. When AI-operated equipment is present on a job site (robotic arms, autonomous material handling, drone operations), verify that the safety zone around the equipment has been defined and that all workers within the zone have current certifications. CON-SAF.1 anchors provide evidence of certification counts at a point in time. Compare the "verified" count against the actual headcount in the zone. A mismatch means someone was working near AI equipment without verified safety credentials.
Records that a construction milestone was verified as complete before payment was authorized. In AI-fabricated construction, the milestone verification may involve AI-generated progress reports, drone-captured site imagery, or sensor data confirming component installation. This anchor records the completion percentage and lien waiver status.
Standards reference: AIA A201-2017 General Conditions requires the architect to certify that work has progressed to the point indicated in the application before the owner authorizes payment.
Evidence anchored: Required completion percentage, verified completion percentage, lien waiver status.
Financial verification is a construction fraud control. When AI generates progress reports (from drone imagery or sensor data), and those reports trigger payment authorization, the AI's accuracy directly affects cash flow. A CON-FIN.1 anchor chain across project milestones creates an auditable record of what completion percentage was reported at each pay application. If a dispute arises over work quality or completion status, the anchor chain shows exactly what was verified and when. Compare the AI-reported completion against the architect's certification for consistency.
The 5 CON procedures map to three regulatory frameworks that apply to every AI-driven construction operation in the United States.
| Procedure | OSHA 29 CFR 1926 | IBC/IRC | AIA/ACI/AISC |
|---|---|---|---|
| CON-STR.1 | 1926.700(a) | IBC 1604, IRC R301 | ASCE 7 |
| CON-INS.1 | 1926.20(b)(2) | IBC 110 (Inspections) | -- |
| CON-ENV.1 | 1926.700(a) | -- | ACI 318-19 26.5.3.1 |
| CON-SAF.1 | 1926.500-503, Subpart O | -- | -- |
| CON-FIN.1 | -- | -- | AIA A201-2017 |
For firms operating in the EU, the CON procedures also map to the EU Machinery Regulation 2023/1230 (robotic fabrication equipment), the EU Product Liability Directive 2024/2853 (AI-fabricated components as products), and ISO 10218:2025 (robot safety in factory settings). See the Dual Compliance guide for the EU regulatory framework.
On January 1, 2026, Verisk (through its ISO subsidiary) published three endorsement forms that allow carriers to exclude AI-related claims from commercial general liability (CGL) policies:
| Form | Coverage Part | What It Excludes |
|---|---|---|
| CG 40 47 | Premises and Operations | Bodily injury and property damage arising out of generative AI during operations |
| CG 40 48 | Auto Liability | AI-related claims under auto coverage |
| CG 35 08 | Products and Completed Operations | Bodily injury and property damage arising out of generative AI in delivered products, including autonomous systems, robotics, and AI-directed physical operations |
CG 35 08 is the critical form for construction. It removes coverage for property damage caused by autonomous systems, robotics, and AI-directed physical operations in the products and completed operations hazard. For a general contractor delivering AI-fabricated wall panels, this means the CGL policy may not cover a structural defect in a delivered component if the defect arose from an AI decision.
At least six major carriers have filed their own AI exclusions with state regulators since January 2026. The exclusions are already appearing on policy renewals. Many contractors will not discover the exclusion until they file a claim.
Insurance underwriters price risk based on available evidence. A construction firm that can demonstrate continuous AI governance evidence -- structural verifications, inspection attestations, environmental monitoring chains, and safety certification records -- presents a materially different risk profile than one that cannot. Witness anchors do not guarantee coverage. But they give the underwriter evidence to evaluate, and they give the claims adjuster a timestamped record of what was governed when the loss occurred.
On large projects -- subdivisions, commercial builds, infrastructure -- insurance is typically structured as an Owner-Controlled Insurance Program (OCIP) or Contractor-Controlled Insurance Program (CCIP). These "wrap" policies are designed to cover all enrolled parties (GC, subs, and sometimes the owner) under a single program. The assumption is that everyone on the project is insured. But if the wrap policy includes the Verisk AI exclusions, every subcontractor using AI-operated equipment or AI-fabricated components on that project could be unknowingly uninsured for AI-related claims. The GC who structured the wrap assumes all enrolled parties are covered. The robotics sub who brought the prefab cell assumes the wrap covers their work product. Neither checked the endorsement schedule for CG 35 08. In a loss event involving an AI-fabricated structural component, both parties discover the gap at the same time -- when the claim is denied.
During insurance renewal reviews, ask the contractor three questions: (1) Does your current CGL policy include CG 35 08, CG 40 47, or CG 40 48 endorsements? (2) If yes, what AI-related claims are excluded? (3) Can you demonstrate continuous AI governance evidence to your underwriter? Many contractors do not know whether their policy includes the AI exclusion. The endorsement may have been added at renewal without explicit disclosure. Review the actual policy endorsements, not just the declarations page. For OCIP/CCIP programs, ask whether the wrap policy's endorsement schedule was reviewed for AI exclusions before enrollment. If it was not, every enrolled party using AI on the project may have an uncovered exposure that the wrap was assumed to address.
A factory uses KUKA robotic arms controlled by an AI translation model to convert CAD files into wall panels. The AI determines stud spacing, nailing patterns, and sheathing cut geometry. Panels ship directly to residential construction sites.
CON procedures in play:
Assessment value: When the building inspector at the subdivision receives the wall panel, the CON-STR.1 anchor chain shows exactly which code requirements the AI targeted, what values it achieved, and how each was verified. The inspector does not need to reverse-engineer the AI's decisions. The evidence chain is already there.
For prefabrication facilities, the chain of custody matters. The factory inspector (CON-INS.1) and the site inspector are different people, potentially in different jurisdictions. Verify that the CON-STR.1 anchor references the specific code edition enforced by the receiving jurisdiction. A panel built to 2024 IRC in a factory in State A, shipped to State B that still enforces 2018 IRC, may meet different (possibly stricter) requirements. The anchor should reflect the applicable code for the installation location, not the factory location.
A general contractor uses drone-mounted cameras with vision-language models to perform daily safety and progress inspections. The AI identifies PPE non-compliance, fall protection gaps, scaffold deficiencies, and construction progress relative to the project schedule.
CON procedures in play:
Assessment value: OSHA 29 CFR 1926.20(b)(2) requires "frequent and regular inspections." Daily drone inspections with CON-SAF.1 anchors demonstrate a level of inspection frequency that exceeds what manual patrols can achieve. The anchor chain provides OSHA with timestamped evidence of continuous safety monitoring, not just spot-check records.
AI-powered drone inspections can detect hazards faster than human patrols, but they also create a record of hazards the firm was aware of. If the AI identifies a fall protection gap on Monday and the gap is not corrected by Wednesday, the CON-SAF.1 anchor chain shows the firm had constructive knowledge of the hazard for two days. Under OSHA's General Duty Clause, knowledge of a recognized hazard without corrective action is a violation. Verify that the firm has a response protocol: AI detection leads to anchor, anchor triggers corrective action, corrective action is verified by a follow-up anchor.
A structural engineering firm uses AI to compute load calculations, seismic resistance parameters, and material specifications for a commercial building. The AI's output is incorporated into the structural drawings that the licensed PE stamps and submits for building permit review.
CON procedures in play:
Assessment value: The PE stamps the drawings, but the AI computed the values. If a structural failure occurs, the question is whether the PE verified the AI's output or simply relied on it. CON-STR.1 anchors with verification method "3" (licensed inspector/engineer) prove the human verified the AI's work. Anchors with method "1" (AI only) show the AI's output was not independently verified by a licensed professional.
This is a professional liability question. A PE who stamps AI-computed structural calculations without independent verification may be exposing their license. Check whether CON-STR.1 anchors for load-bearing structural elements consistently show verification method "3" (licensed professional). A pattern of method "1" (AI only) for critical structural elements is a red flag for both the building department and the PE's professional liability carrier.
CON procedures use the same SWT3 SDK as AI inference witnessing. The evidence comes from the physical world (sensor readings, inspection results, certification counts), but the witnessing mechanism is identical.
from swt3_ai import Witness
witness = Witness(
endpoint="https://your-instance.example.com/api/v1/witness",
api_key="axm_your_key",
agent_id="factory-cell-01"
)
# Structural verification: stud spacing vs IRC R602.3
witness.witness_event(
procedure="CON-STR.1",
factor_a="16", # Code requirement: 16" on-center
factor_b="15.98", # Actual measured spacing
factor_c="2" # Verification method: sensor
)
# Inspection attestation: end-of-line QA
witness.witness_event(
procedure="CON-INS.1",
factor_a="50", # Required checklist items
factor_b="50", # Items passed
factor_c="license_verified"
)
# Environmental monitoring: concrete cure temperature
witness.witness_event(
procedure="CON-ENV.1",
factor_a="50", # Minimum temp (F)
factor_b="62", # Actual reading
factor_c="72" # Hours elapsed since placement
)
witness.flush()import { Witness } from '@tenova/swt3-ai';
const witness = new Witness({
endpoint: 'https://your-instance.example.com/api/v1/witness',
apiKey: 'axm_your_key',
agentId: 'drone-inspection-01'
});
// Site safety: daily drone assessment
witness.witnessEvent({
procedure: 'CON-SAF.1',
factorA: '12', // Required safety certs for zone
factorB: '12', // Verified certs
factorC: 'ppe_compliant'
});
// Financial verification: milestone payment
witness.witnessEvent({
procedure: 'CON-FIN.1',
factorA: '25', // Required completion %
factorB: '27', // Verified completion %
factorC: 'lien_waiver_received'
});
await witness.flush();For full SDK documentation and additional language examples, see SDK Documentation.
The International Code Council publishes new editions of the IBC and IRC on a 3-year cycle. The 2027 I-Codes, finalized in August 2026, include a significant change for AI-fabricated construction: formal incorporation of ICC/MBI 1200, 1205, and 1210 for offsite construction. These are the first building code provisions specifically designed for factory-built components.
This matters because AI-fabricated panels, trusses, and modular units are manufactured offsite and installed on job sites in jurisdictions that may be enforcing the 2018 or 2021 edition. The factory operates under one code edition. The installation site may operate under another. The structural requirements may differ. The inspection requirements certainly differ.
A CON-STR.1 anchor records which code requirement was used as the verification threshold. For a wall panel fabricated in Arizona and installed in Nevada, the anchor shows whether the AI targeted the 2018 IRC (Nevada's adopted edition), the 2024 IRC (Arizona's adopted edition), or the 2027 IRC (not yet adopted anywhere). The building inspector receiving the panel can verify that the structural parameters were computed against the code edition their jurisdiction enforces, not the factory's local code.
Code edition mismatches between factory and installation site are a real and growing problem as offsite construction scales. When reviewing CON-STR.1 evidence for prefabricated components, verify: (1) which code edition the AI targeted, (2) whether that edition matches the jurisdiction of installation, and (3) whether the factory's QA program accounts for multi-jurisdiction shipments. A factory that ships to 15 states may need to target 15 different code editions. The anchor chain should reflect the installation jurisdiction, not the factory jurisdiction.
A conventionally framed wall has a carpenter who can testify about the construction process. An AI-fabricated wall has a model version, a set of input parameters, and a manufacturing log. The evidence requirements are fundamentally different. If the only governance artifact is a QA checklist signed by a human who watched the robot work, the AI's decision-making process is undocumented.
OSHA has not published AI-specific construction standards. But the General Duty Clause, machine guarding requirements (Subpart O), and the frequent inspection mandate (1926.20(b)(2)) all apply to AI-operated equipment on construction sites. The absence of a specific standard does not create a compliance exemption. It creates a General Duty Clause exposure.
A one-time safety assessment when the robot is commissioned does not satisfy 1926.20(b)(2)'s "frequent and regular" inspection requirement. AI-operated equipment should have continuous CON-SAF.1 anchors showing that safety certifications are current and PPE compliance is maintained across every shift the equipment operates.
The Verisk CG 35 08 endorsement is already appearing on CGL policy renewals. Many contractors will not discover the exclusion until they attempt to file a claim for an AI-related defect. Review your policy endorsements now, not after a loss event. If the exclusion is present, discuss it with your broker and consider whether supplemental coverage (such as technology E&O or a manuscript endorsement) is available.
The factory witnesses structural verification (CON-STR.1) and QA inspection (CON-INS.1). The job site witnesses safety compliance (CON-SAF.1) and environmental conditions (CON-ENV.1). If these evidence streams are disconnected -- different systems, different formats, no shared identifiers -- the full lifecycle from fabrication to installation is opaque. All 5 CON procedures land in the same witness ledger, creating a single evidence chain from factory floor to finished structure.