The real opportunity is not more technology. It is fewer decisions made with incomplete, outdated or unverified information.
1. Construction risk begins as an information problem
A contractor may receive a revised detail after procurement has started. A consultant might extract the wrong material property from a manufacturer’s technical sheet. An impressive BIM model can still contain obsolete values, missing classification or unapproved assumptions. A design team may coordinate a duct and beam geometrically yet overlook access requirements for maintenance.
None of these problems disappears simply because a project buys new software. Risk falls only if the right information reaches the right decision-maker in time, is checked against explicit requirements, and leads to documented action.
This is where BIM, digital twins and AI can complement experienced engineering teams: they can make specific information checks more reproducible, expose uncertainty earlier and assist people in prioritizing what requires attention.
2. Start with the risk, not the AI application
Before testing AI, define a bounded problem. Examples include checking drawing revisions before procurement, identifying missing properties in IFC deliveries, examining conflicting product documentation, or monitoring selected building systems after handover.
For each use case, establish a risk register with: the decision being protected; expected evidence; source-of-truth hierarchy; error severity; process owner; approval authority; and what a successful intervention would look like.
A project that cannot define a measurable baseline cannot credibly claim that a pilot reduced risk.
3. BIM is the structured foundation—but the model is not the truth
BIM supports more than geometry. It can link objects, systems, quantities, classifications and asset attributes to operational questions. The ISO 19650 family frames BIM as an information-management discipline across the asset lifecycle, with requirements for how information is organized, exchanged and controlled [1].
Industry Foundation Classes (IFC) provides an open, vendor-neutral schema for built-asset information. buildingSMART lists IFC 4.3 ADD2 (4.3.2.0) as a final release, associated with ISO 16739-1:2024 [2]. For the wider picture of how AI, BIM and digital twins support sustainable construction, see Jarlhalla’s earlier overview article (https://jarlhalla.com/ai-sustainable-construction-bim-digital-twins/); this article focuses narrowly on risk and decisions.
However, exporting IFC is not an assurance process. A valid file can still describe the wrong design, omit critical data or reflect an outdated revision.
buildingSMART’s Information Delivery Specification (IDS), approved as a final standard in 2024, enables machine-interpretable information requirements and automatic checks of specified IFC data [3]. IDS checks alphanumeric information: objects, properties, materials, classification, relationships and values. It does not verify geometry or establish structural, fire or regulatory compliance by itself.
An engineer should distinguish three levels:
- Delivery check: Is the required field or relation present?
- Evidence check: Is the value current, applicable and supported by a trustworthy source?
- Engineering verification: Does the actual proposed solution meet the applicable technical requirement?
These are different decisions and must not be collapsed into one green dashboard status.
4. Digital twins: connect decisions to operational evidence
A static model may provide valuable asset information without being a functioning digital twin. In practical building management, a twin should connect a defined asset or system to a maintained stream of observations, with clear purpose, data provenance and update procedures.
For example, energy meters, indoor temperatures, equipment events and maintenance records may support investigation of operating inefficiencies. But faulty sensors, stale model data and changing occupancy can distort conclusions. Before automating recommendations, teams must check reliability, timestamps, units and the causal limits of the available observations.
The purpose of the twin should therefore be described as an operational decision, not as a 3D visualization feature.
5. Where AI helps—and where it must stop
Useful bounded applications include extracting document references, comparing revision sets, drafting issue lists, explaining deterministic rule-check results and highlighting unresolved information gaps.
An AI assistant can be fluent and still wrong. It can cite a superseded drawing, confuse a declared product characteristic with assembly performance, or interpret a missing property as evidence of noncompliance.
A defensible workflow is controlled source → deterministic check where possible → AI-supported interpretation → qualified professional review → recorded decision.
The voluntary NIST AI Risk Management Framework 1.0 (2023) describes continuous governance, mapping, measurement and management of AI risks, including testing and monitoring under relevant conditions [4]. In an engineering context, those activities must be combined with existing professional responsibilities and applicable rules.
AI should not autonomously sign off a design, approve a material substitution or represent regulatory compliance without authorized technical verification.
6. An illustrative project scenario
Consider a refurbishment project with several hundred doors and a changing fire and access-control schedule. The team wants to reduce missed inconsistencies before procurement.
Baseline: record how many approved schedules and drawing revisions are checked manually, the time required, observed exceptions, and the number of defects discovered downstream.
Intervention: define a controlled property requirement for door objects, validate structured data and use AI only to summarize the resulting exceptions with source links. A qualified reviewer checks whether the underlying documents and design decisions are correct.
Outcome measures: review effort, false alarms, missed known errors, time-to-resolution and subsequent verified rework.
This scenario is hypothetical. No percentage savings or actual project result is claimed. The effect must be measured on real cases before it can support a commercial assertion.
7. Evidence beats demonstration
For a pilot, distinguish task completion from technical correctness. Track data completeness, version correctness, false positives, false negatives, review time, unresolved critical cases and unauthorized actions. Evaluate performance by risk severity, not only average accuracy.
A successful demo of ten carefully chosen documents does not establish reliability across unknown building projects. A credible assessment includes difficult inputs, missing data, conflicting revisions and situations where the correct system response is needs review.
Do not present a reduction in manual effort as a reduction in safety risk unless the relevant safety outcomes have actually been measured.
8. From assessment to pilot to scale
Jarlhalla’s proposed Project Intelligence Assessment starts with a focused 2–4 week diagnostic of one project or process. The output is a risk picture, information maturity assessment, candidate improvements, clear limitations and an evidence-informed prioritization.
A separate pilot can then test one controlled use case, with documented before/after measures and defined approval rights. Only solutions with credible value, manageable risk and appropriate ownership should proceed toward wider deployment.
The sequence is Assess → Pilot → Scale. This is a proposed service model—not a claim that a particular customer has already achieved the outcomes described.
9. Engineering judgment is the point, not the obstacle
Construction projects do not become safer because they contain more models, dashboards or AI-generated text. They become more controllable when assumptions are visible, information is traceable, checks are relevant and qualified people can act before errors become expensive or dangerous.
BIM can structure the evidence. Digital twins can help maintain operational context. AI can help people navigate complexity. The final responsibility remains with the relevant professionals and project organizations.
Discuss a Project: If your team has a specific documentation, BIM or project-information challenge, start by defining the decision and evidence required. A bounded assessment is often a better first move than a broad technology rollout.
References (verified primary/official sources)
- ISO, ISO 19650-1:2018: Information management concepts and principles. Confirm the current edition, national adoption (NS-EN ISO 19650) and contract edition before relying on it.
- buildingSMART International, Industry Foundation Classes.
- buildingSMART International, Information Delivery Specification (IDS).
- NIST, AI Risk Management Framework.
Featured image: AI-generated conceptual editorial illustration; not a real project, certification or construction detail for execution.

