VASCOM TECHNOLOGY AUTOMATION CO., LTD
Home › News › AI in Underground Construction and Service Operations: Three…

AI in Underground Construction and Service Operations: Three Developments to Watch

From underground worksites to software systems, AI helps predict risks and resolve incidents under human supervision.

Three announcements on 19/11 show AI being applied to work that requires reading large amounts of data and making careful decisions. In each case, the system helps detect problems while people retain responsibility for assessment and approval.

1. AI helps predict risks in metro and tunnel construction

VnExpress Số hoá reports on AI applications in geological surveys, monitoring and underground construction. Metro tunnels, underground stations and deep excavations are often close to homes, roads and existing infrastructure. Even small changes in ground shape require close monitoring. AI can combine survey data with sensor signals to identify changes that are difficult to observe directly and help engineers predict risks.

In Japan, a research team at the Shibaura Institute of Technology used 13.926 geological borehole records to train a neural network and build a three-dimensional map of Tokyo’s ground conditions. The model predicts soil properties at locations without direct borehole data. In China, a team led by Professor Gang Zheng at Tianjin University developed a system combining AI, sensors and a digital twin to control ground movement with an accuracy of about 0,1 mm. According to the article, the solution has been deployed at about 100 projects across 13 provinces and cities.

Another study by the team used 1.159 data samples and 34 tunnel-boring machine operating parameters to predict water pressure in the soil. That measure is related to the stability of the excavation face and the risk of ground settlement. Test results cited in the article show that the combined model had lower error than the comparison models. These are research results, not a guarantee of accuracy at every worksite.

For construction companies and infrastructure operators, the practical value lies in how data is organized. Survey logs, machine parameters and monitoring data need to be recorded consistently so engineers can compare alerts with conditions on site. A useful prediction must include the measurement point, time and specific construction conditions. A qualified person must remain responsible for decisions to adjust the construction plan.

2. OpenAI describes an agent workflow for performance incidents

OpenAI introduced Performance Factory, a system that uses agent workflows to investigate and propose responses when ChatGPT performance declines. In the described process, an alert opens a task. An agent investigates the cause, records its findings and proposes a fix for an engineer to review. After the fix is deployed, an agent checks whether performance has returned to its previous level. Engineer approval is part of the workflow.

According to OpenAI, over more than three months, the workflow tracked more than 1.700 investigations and identified the root cause in more than 600 cases. The company says the system conducts about 17 automated investigations each day. Before that, the engineering team had built up more than 30 tools for performance analysis. They put their investigation experience into a shared set of instructions, then tested and refined it using past incidents.

For businesses, the point is not to replicate OpenAI’s scale. The workflow structure offers a useful example: which signal opens a task, which data sources an agent may read, where conclusions are recorded, who approves changes and how results are checked. Small and medium-sized businesses often have recurring tasks such as classifying support requests, checking orders or reviewing data errors. When handling rules are written clearly, an agent can prepare evidence and propose the next step for the responsible employee to decide.

The figures published by OpenAI reflect its own system during the stated period. They do not show how many incidents another business could resolve or how much time it could save. Before expanding use, businesses need to measure results against their own processes and data, including cases where an agent identifies the wrong cause.

3. Google AI emphasizes human curiosity and judgment

Google AI published a conversation between Kent Walker and biographer and historian Walter Isaacson about creativity in the age of AI. The introduction to the discussion mentions Leonardo da Vinci, Benjamin Franklin and Ada Lovelace to raise questions about how innovators work. Google’s description focuses on collaboration, curiosity and human judgment as information becomes easier to access.

This is a discussion about ways of working, not an announcement of a new model or test result. For businesses, the relevant message is to identify where people need to ask questions and check conclusions. AI can gather information and suggest options. Employees who understand customers, processes and risks still need to assess which option should be applied. This is especially important when input data lacks context or a decision directly affects customers.

Three steps businesses should take now

  1. Choose one process with clear data and an assigned owner. Small and medium-sized businesses in Vietnam can start with support requests, document checks or equipment monitoring. Record the inputs, expected results and cases that need to be passed to an employee.
  2. Set an approval point before an agent makes a change. The responsible person needs to see the data the agent used, the reason for its proposal and the expected impact. For steps involving customers, finances or operations, businesses should specify who can approve them.
  3. Measure results using the business’s actual cases. Compare the agent’s results with current handling, record errors and update the instructions. Small and medium-sized businesses should expand the scope after understanding the cases the system commonly handles incorrectly.

VASCOM’s view of these developments

VASCOM believes the first step is to review business processes: where the data is stored, who uses it, which steps recur and which decisions require expertise. From there, a business can choose a specific task to test, define how to check the output and assign responsibility.

VASCOM proposes designing agents by department, with access rights suited to each task. A responsible person approves proposals that affect operations or customers. For data that requires strict control, the deployment approach should consider keeping data at the business and clearly defining what an agent may read and write.


Sources

  • VnExpress Số hoá, “How AI is changing metro and tunnel construction”, 19/11/2026, vnexpress.net.
  • OpenAI, “How OpenAI’s Performance Factory keeps ChatGPT fast”, 19/11/2026, openai.com.
  • Google AI, “Walter Isaacson on curiosity and creativity in the AI age.”, 19/11/2026, blog.google.
Technology news Talk to VASCOM
More news

Keep reading

Zalo ☎
VASCOM
Call 0822.577.579 Get a consultation