Checking whether a door matches the latest specification sounds straightforward. On a major infrastructure project, the same task could involve reviewing thousands of components against changing designs, contractual requirements and inspection records. This is where enterprise AI faces a more demanding test. Reliable answers require accurate engineering information, traceable sources and strict control over sensitive data.
Southeast Asia’s businesses are already moving towards wider deployment. A February 2026 report by McKinsey, the Singapore Economic Development Board and Tech in Asia found that 46% of surveyed organisations already using AI in the region were scaling or had fully scaled their initiatives, compared with 35% globally. The research covered more than 300 respondents across six Southeast Asian markets.
Why AI adoption does not guarantee business returns
For infrastructure owners, however, adoption brings questions that extend beyond access to powerful models. Can AI interpret specialised project data accurately? Can it operate within strict security requirements? And can its answers help teams make better decisions? The answers depend on how well AI connects engineering information, respects access controls and supports decisions that professionals can verify.
Infrastructure data demands more than a conversational interface
An infrastructure project generates many different kinds of information. Building information modelling (BIM) represents physical components and their properties digitally. Contracts establish obligations. Schedules track planned work, while inspection reports and operational records document what happens on the ground.
The difficulty lies in understanding how these sources relate to one another. A question about whether a component meets the latest specification may require a system to identify the right object in a model, retrieve the current specification and distinguish an approved change from an outdated document. A fluent summary alone cannot establish whether the underlying comparison is correct.
Useful AI therefore needs access to authoritative information, consistent identifiers and reliable methods for querying technical data. Where quantities or engineering rules are involved, the system should use appropriate calculation and checking tools, with results that engineers can inspect. Natural-language interaction can make these processes easier to use. Its value depends on the quality of the information and tools behind the interface.
Connecting information across the asset lifecycle
Fragmentation creates another obstacle. Designers, contractors, suppliers and asset operators may maintain separate systems, with information transferred between them at different stages. An AI application drawing on an incomplete handover could miss a maintenance requirement. One using an obsolete design could flag a discrepancy that has already been resolved.
Singapore’s Building and Construction Authority describes Integrated Digital Delivery as an approach that connects stakeholders and processes through digital technologies across the building lifecycle. This provides a useful foundation for thinking about infrastructure AI: connected, governed information must underpin the answers. That does not mean every participant should see everything. A shared information environment still needs clear ownership, version control and permissions. AI must respect those boundaries when retrieving source material and generating responses.
Singapore’s Punggol Digital District offers an example of how operational data can support this approach. In July 2026, JTC and Panasonic announced plans to develop and validate AI-enabled facilities management applications using the district’s Open Digital Platform, including systems to detect unusual building conditions and coordinate inspection responses. According to the announcement, data shared for testing is curated and anonymised, while proprietary tenant and project information is isolated through controlled access and separate logical data zones. The example shows how wider use of operational data can be paired with restrictions on what each participant can access.
Bringing domain knowledge into AI applications
LeapThought’s partnership with Google Cloud, announced on 23 October 2026, illustrates this approach. The announcement describes applications such as identifying fire-rated doors that differ from the latest specification, extracting material quantities and identifying patterns in non-conforming work. Such applications could help teams investigate technical questions, provided the underlying records are reliable and the findings can be checked.
Other applications use imagery rather than project documents. Bentley Systems’ Blyncsy applies AI to crowdsourced images for roadway asset detection and inspection. In April 2025, Bentley announced additional capabilities combining Google Street View imagery and Vertex AI to help identify changes in road conditions and roadside assets. This illustrates a different technical requirement: turning visual observations into information that maintenance teams can assess and act on.
Data location is only one part of control
For critical infrastructure owners, making information accessible to AI also raises questions about where it is processed and who can access it. Data residency concerns where information is stored or processed. Broader sovereignty requirements can also involve control over administrators, encryption keys, system operation and external connectivity. Keeping data in a particular country does not, by itself, answer all these questions.
These concerns already affect regional adoption. In Deloitte’s Thailand Digital Transformation Survey 2026, 38% of organisations said the security of their AI infrastructure needed attention. The same proportion identified data readiness as an issue. These findings cover Thai businesses generally, rather than infrastructure operators specifically, but highlight the practical barriers surrounding enterprise AI.
As organisations strengthen their approach to cybersecurity in Southeast Asia, infrastructure owners must also consider how AI handles information about sensitive physical assets, including their design, condition and operations.
Running AI within a chosen security boundary
Deployment choices need to reflect the sensitivity of the information and the controls an organisation must maintain. For some workloads, this may require an environment disconnected from external networks. Google describes its Distributed Cloud air-gapped offering as capable of operating without connectivity to Google Cloud or the public internet, with Gemini available for on-premises applications. This creates an option for organisations that need AI processing within an isolated environment.
Isolation still requires careful implementation. Owners need to establish how documents enter the environment, how updates are approved, where logs are retained and which users can retrieve particular records. Generated answers must not reveal information that the person asking the question lacks permission to view.
The appropriate architecture will depend on the asset, the information involved and the organisation’s obligations. The practical aim is to make security requirements part of deployment design from the outset.
Measuring value through better project decisions
Infrastructure owners can start with a specific task and a clear baseline. Does AI shorten specification reviews while identifying genuine discrepancies? Can it extract quantities accurately or flag risks early enough for teams to act? Those gains must justify the cost of preparing data, integrating systems and maintaining the deployment. Where findings affect safety, compliance or contracts, qualified professionals must retain oversight.
Checking a door against its specification offers a simple test: can the system identify a mismatch, show the current requirement and give an engineer enough evidence to act? For Southeast Asia’s infrastructure owners, enterprise AI will earn its place through reliable answers that improve decisions under real operational and security constraints.