AI in civil engineering in Mexico with judgment and evidence

An assistant produces a polished engineering report in seconds. Its tables are organized, its conclusions sound convincing, and it includes technical references. Before celebrating the time saved, one question remains: can someone reconstruct the reasoning, check the inputs, and take responsibility for the result? That question should guide a useful approach to artificial intelligence in civil engineering.

For a Mexican consulting firm, public agency, or university team, the starting point can be a small, verifiable task. We propose evaluating AI through the quality of the complete workflow, including review, and the consequences of an error. Generation speed alone provides an incomplete measure of its benefit.

Professional responsibility still belongs to a person

ASCE’s position is explicit: AI can support engineering activities, while responsibility and professional judgment remain with the engineer. Policy Statement 573 was adopted in July 2024 (American Society of Civil Engineers [ASCE], 2024). In Mexico, this is a professional practice reference; it does not have the status of a Mexican legal provision.

This suggests an operating rule: reviewers should know which information comes from measurements, which comes from a validated model, and which is a generated proposal. When a number moves from a draft into a project document, its source, units, date, and reviewer should remain identifiable. Labeling a document “AI assisted” does not establish that traceability.

Choose a first application that can be checked

As a pilot, we suggest organizing previously anonymized inspection observations, comparing versions of a specification, or preparing an index of public documents. Select a task whose correct output can be evaluated against examples reviewed by the team. A pilot provides little value if it merely demonstrates that the system can produce lengthy answers.

Different technologies also require different evaluations. Classifying photographs with a trained model does not raise exactly the same testing questions as asking a generative assistant to explain a standard. NIST’s generative AI profile identifies specific risks, including confidently presented false content and information exposure (Autio et al., 2024). A persuasive answer still needs independent verification.

A practical starting example

Suppose a consultancy wants to classify maintenance reports. It could assemble an authorized dataset, remove identifiers, and ask the system to assign categories defined in advance. A reviewer would then compare each classification against a reference assessment. This is a hypothetical example, rather than an observed success rate or a guarantee that another dataset will work equally well.

Before expanding the pilot, we suggest separating errors by consequence. Confusing two administrative labels requires a different response from overlooking an observation that might warrant urgent inspection. Acceptance should reflect those differences. An overall accuracy average could conceal precisely the cases that the responsible reviewer needs to identify.

Turn an experiment into a controlled procedure

NIST’s AI RMF organizes risk management around four functions: govern, map, measure, and manage (Tabassi, 2023). To turn that reference into an office exercise, we propose assigning a pilot owner, describing the intended use, preparing evaluation cases, and deciding in advance when the tool will be suspended. These organizational decisions should be recorded.

Each test record can identify the tool version, inputs, output, human review, and corrections required. If the supplier changes the service, those records help determine which tests need to be repeated. Saving only the final document would make it harder to explain why a new version produces a different answer.

Measure savings from information preparation through approval of the output. We suggest comparing total time, rework, and consequential errors against the established process. If preparing instructions and correcting results takes longer than completing the task directly, the pilot still provides a useful finding: that application needs a different approach.

Bring the Mexican context into the inputs

For a Mexican project, the team should identify the jurisdiction, contractual documents, and applicable technical references before consulting the system. We suggest avoiding vague requests such as “apply the current standard.” Instead, specify the document, its edition, and the operation to be reviewed, without automatically assuming that a foreign reference has local authority.

The same discipline applies to units, terminology, and site conditions. An evaluation exercise can deliberately include changes between meters and millimeters, similar material names, and incomplete background information. The purpose is to discover how the tool responds to problematic inputs before using it in an actual project file.

For information control, we recommend beginning with public documents or authorized datasets. Client drawings, sensitive locations, and personal information require the treatment established by the organization. Tool selection should consider where information is processed and who can access it, alongside the functions displayed to users.

A learning opportunity for students and teams

A student chapter could organize a comparison of AI answers to an instructional case. Teams would identify assumptions, verify references, and explain what they would accept or return for correction. The most valuable deliverable would be the verification record. This proposed activity develops technical reading and communication without treating an automatically generated output as a completed design.

The issue is timely. In September 2026, ASCE reported on the preparation of AI RACE, its roadmap for AI applications; that article announced a later release, rather than an already mandatory standard (Degges, 2026). NIST also states that its Framework 1.0 is being revised (NIST, n.d.). Keep a record of the edition consulted.

Start with a decision that can be audited

The next innovation meeting could end with a concrete assignment: select a process, an authorized dataset, a reviewer, and an acceptance condition. Then compare the pilot with established practice. Using AI with judgment means being able to explain what improved, which errors appeared, and why the team decided to continue, adjust, or stop its use.

Sources and recommended reading in APA 7 format

American Society of Civil Engineers. (2024). Policy statement 573—Artificial intelligence and engineering responsibility. https://www.asce.org/advocacy/policy-statements/ps573---artificial-intelligence-and-engineering-responsibility/

Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.600-1

Degges, I. (2026, September 14). ASCE leads AI vision in civil engineering with ‘AI RACE’ roadmap. American Society of Civil Engineers. https://www.asce.org/publications-and-news/civil-engineering-source/article/2026/09/14/asce-leads-ai-vision-in-civil-engineering-with-ai-race-roadmap

National Institute of Standards and Technology. (n.d.). AI risk management framework. Retrieved October 17, 2026, from https://www.nist.gov/itl/ai-risk-management-framework

Tabassi, E. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.100-1

Continue exploring

Leer este artículo en español · More civil engineering articles · ASCE Mexico webinars · Community and student chapters

Illustrative photograph: ThisisEngineering / Unsplash.

Contact

We are always here to help

SOCIAL MEDIA

LINKS OF INTEREST

[email protected]