AI in Engineering: Why Accountability Still Matters

By Mark Coates

The United States, the European Union, and the United Kingdom are taking different routes in governing artificial intelligence (AI), but engineering firms ultimately face the same question: Who remains accountable for the decision?

AI is rapidly moving into engineering practice, changing how information is found, designs are developed, alternatives are assessed, and projects are managed. Yet the most important question for the profession is becoming less about what the technology can do and more about what happens to professional responsibility when AI becomes part of the engineering process.

That is one of the strongest conclusions to emerge from the American Council of Engineering Companies (ACEC) Research Institute's new report, Leading Through AI Risk: The Enterprise Framework for Engineering Firm Leaders. The report raises questions that reach far beyond the U.S. market because the fundamental challenge that it identifies is relevant to engineering organizations around the world: AI may change how work is undertaken, but it does not remove responsibility for the quality, safety, and integrity of the outcome.

The report identifies eight interconnected areas of risk, covering technical reliability, professional liability, data governance, workforce capability, ethics and regulation, cybersecurity, commercial models, and enterprise leadership. More important than the individual categories, however, is the argument connecting them: AI risk is not principally an IT problem to be handed to a technology team. It is an organizational risk that reaches into professional judgement, quality assurance, knowledge management, workforce development, and, ultimately, the way an engineering business is led.

That conclusion is particularly interesting when considered internationally. While the professional challenge is broadly shared around the world, the mechanisms being developed to govern AI are beginning to diverge.

The U.S. is putting greater weight on adoption and organizational responsibility

The federal direction in the U.S. is explicitly pro-adoption. Policies are focused on accelerating innovation and reducing regulatory barriers that could constrain American AI development and deployment. However, state legislatures across the U.S. continue to develop their own approaches to AI, including in private-sector use, discrimination, healthcare, and government applications, leading federal officials to push for a more consistent national approach. For engineering businesses operating across multiple states, the practical picture becomes more varied than federal policy might suggest.

The U.S. has not abandoned risk management: the National Institute of Standards and Technology (NIST) provides a substantial AI Risk Management Framework covering governance, measurement, and management of AI risk. However, that framework is voluntary, in contrast to the comprehensive statutory regime of the EU AI Act.

For engineering firms, this creates an environment in which significant responsibility remains with the organization to determine how AI can be introduced without weakening existing professional controls. That fits closely with the ACEC report, which repeatedly argues that engineers must remain responsible for validating AI-assisted work, maintaining appropriate QA/QC, and exercising professional judgement regardless of the technology used to produce an analysis or recommendation.

There is an important distinction here. A less prescriptive federal regulatory approach does not mean an absence of accountability, particularly in engineering, where professional licensure, contractual obligations, standards of care, and public safety requirements already create significant duties. The current federal approach is better characterised as placing greater weight on innovation within those existing structures, albeit against an increasingly complex state regulatory background.

This may also encourage engineering businesses to move more quickly from experimentation into implementation, although speed brings its own management challenge. As AI becomes embedded within ordinary engineering software and project workflows, organizations need to understand not merely which tools employees are using, but what information is being provided to those systems. They also need to know whether project or client data is leaving the organization's governed environment, where AI is influencing professional decisions, and whether the assurance around those decisions remains proportionate to the consequences of getting them wrong.

That question extends to peer review and independent checking. Engineering assurance has traditionally relied on separation of judgement, particularly where work is safety-critical or technically complex. If the same AI model, source data, grounding information, or underlying assumptions contribute to the original engineering work and its subsequent review, firms will need to consider whether that process provides independent assurance or risks reproducing the same underlying failure modes.

The EU is placing more of the assurance framework into law

The EU has taken a different route. Its AI Act became generally applicable in August 2026, establishing statutory obligations for how AI systems are developed, deployed, and used. The act is being implemented in stages, with the main requirements governing high-risk systems taking effect later. Starting in December 2027, these requirements will apply to specified high-risk uses, including certain AI systems used as safety components in the management or operation of critical infrastructure. A further set of requirements will apply, starting in August 2028, to high-risk AI that falls within the act's product-safety route, covering specified products already subject to EU safety legislation.

The distinction is important for infrastructure because the act does not classify every use of AI on an infrastructure project as high-risk. Classification depends on the system's intended purpose and whether it falls within the act's specified high-risk categories, including certain AI used as a safety component in the management or operation of critical infrastructure.

For engineering and infrastructure organizations, that creates a different governance environment. Depending on the application, organizations may increasingly need to consider not only whether an AI-assisted process produces a technically defensible answer, but what regulatory classification applies, what information must be retained, what transparency requirements exist, and how compliance can be demonstrated.

Establishing common expectations would provide clear advantages, particularly where AI is being deployed in circumstances that could have significant consequences for safety, infrastructure performance, or the public. A defined regulatory framework can provide consistency and improve confidence among clients, regulators, and the wider public, especially as more sophisticated AI systems move into professional workflows.

There is, however, an important distinction between regulatory compliance and engineering assurance. A system could satisfy the relevant governance requirements without the resulting engineering decision being correct. Equally important, a technically sound engineering decision could still expose an organization to risk if the data, AI system, or process surrounding it has not been appropriately governed. The two disciplines, therefore, need to reinforce one another rather than become substitutes.

That distinction becomes increasingly important as AI outputs become more convincing. The ACEC report makes the perceptive point that improving AI accuracy may increase the importance of professional scrutiny. The report predicts that obvious errors could become less frequent while subtle errors, incomplete assumptions, and plausible but incorrect conclusions could become harder to identify.

The U.K. is developing a third route

The regulatory philosophy in the U.K. is distinctly different. Its approach has so far been built around a principles-based framework delivered through existing regulators. Safety, security, transparency, fairness, accountability, and redress are all addressed within the context of individual sectors rather than through a single horizontal law equivalent to the EU AI Act. Government guidance has described that framework as pro-innovation and context-specific, giving regulators considerable flexibility over how the principles are applied within their existing remits.

In transport, the Office of Rail and Road has already provided a clear example of what this approach can look like in practice. In its 2026 Safe AI Innovation Action Plan, the agency states that it does not intend to create new technology-specific rules. Rather, it seeks to enable AI within existing statutory frameworks while maintaining safety, assurance, and accountability. That approach creates a significant distinction: It emphasizes understanding how established regulatory obligations apply when AI enters the system, rather than assuming each new technology requires an entirely separate regulatory architecture.

Professional institutions are also beginning to fill part of that assurance space. The Royal Institution of Chartered Surveyors (RICS) released its professional standard on responsible use of AI, which became effective in March 2026. It establishes requirements covering governance and risk management, professional judgement and oversight, transparency, and client communication. Although surveying and engineering are not interchangeable professions, the direction is relevant because it demonstrates how professional standards can translate broad principles of accountability into expectations applying directly to practitioners and regulated firms.

The U.K. may, therefore, become an interesting test of whether AI can be effectively governed through a combination of existing legislation, sector-specific regulation, and strengthened professional responsibility, rather than primarily relying on broad voluntary frameworks or a comprehensive horizontal AI law.

Different systems, but the same engineering question

It would be too simplistic to describe the U.S. as innovative, the EU as cautious, and the U.K. as sitting somewhere in between. All three are attempting to balance innovation, economic competitiveness, safety, and public confidence. The more useful distinction is the relative weight that each places on organizational governance, statutory requirements, sector regulation, and professional accountability.

For infrastructure leaders, that distinction matters. Engineering firms increasingly operate across all three environments, often delivering multinational projects through common digital platforms and shared engineering teams. A global business cannot sensibly maintain one standard of technical judgement for an American project, another for the U.K., and another for the EU simply because the regulatory architecture surrounding AI differs.

The better approach is to establish an organizational standard capable of satisfying the underlying expectation across jurisdictions: understand where AI is being used, govern the data on which it relies, verify material outputs, maintain meaningful human oversight, preserve appropriate independence in checking and review, record significant engineering decisions, and maintain an unambiguous line of professional accountability.

That brings the argument back to the central finding of the ACEC research. Reliable AI ultimately depends on reliable organizations; professional responsibility cannot simply be delegated to a model or software platform. The report also raises an issue that deserves considerably more attention across the profession: If AI increasingly undertakes the routine calculations, document reviews, research, and technical tasks through which younger engineers have traditionally developed experience, firms will need to redesign how professional judgement is learned rather than assuming that it will somehow develop automatically.

This is not only an issue for early-career engineers. AI literacy will need to become a core professional competency across the workforce, not as an alternative to engineering knowledge but as part of an engineer's ability to understand, challenge, and appropriately use the systems that support their work. Experienced professionals will need to understand enough about AI to recognize its limitations, while younger engineers will still need sufficient grounding in engineering fundamentals to identify when a seemingly convincing output does not make sense.

This may prove to be one of the more consequential risks associated with AI adoption. Engineering organizations can introduce governance processes around individual AI systems relatively quickly—but rebuilding professional capability once it has been eroded would be considerably harder.

Accountability could become the competitive advantage

The debate around AI in engineering has understandably concentrated on adoption rates, productivity, and technological capability, but these may ultimately become poor measures of competitive advantage. Access to powerful general-purpose AI capabilities is becoming increasingly widespread, which means access to AI alone is unlikely to provide a durable advantage. Differentiation will increasingly depend on how firms combine those capabilities with trusted engineering data, domain expertise, integrated workflows, institutional knowledge, and disciplined assurance.

Engineering firms will also need to confront a commercial question. The financial effect of AI productivity will depend partly on how professional services are contracted. Under time-based fee models, reducing the hours required to produce an outcome can reduce revenue unless the saved capacity is productively redeployed; under fixed-price work, the immediate effect may instead be improved margin. Over time, however, competition and client expectations are likely to determine how much of those productivity gains remain with the consultant and how much are returned through lower fees, faster delivery, broader scope, or better outcomes. The larger opportunity may be to compete around the value of the engineering outcome, the quality of professional judgement, and the risk being managed, rather than simply the number of hours required to produce the work.

It does not mean that engineering expertise becomes less valuable. It makes the opposite case. As routine production becomes easier to automate, the elements of engineering that are difficult to commoditize—including judgement, experience, systems thinking, assurance, and responsibility for the consequences of a decision—become increasingly important.

What will remain much harder for competitors to replicate is the combination of trusted engineering knowledge, experienced people, disciplined assurance processes, and an organizational culture that knows when technology can be trusted and when it needs to be challenged.

The strongest engineering firms will not necessarily be those that adopt AI fastest, nor those that construct the largest compliance functions around it. They will be the organizations capable of extensively using AI without weakening the chain of professional responsibility, where the source of important information can still be established, assumptions can be challenged, outputs can be independently tested, engineering judgement remains visible, and somebody is ultimately prepared to stand behind the decision.

The U.S., the EU, and the U.K. may be taking different routes, which are likely to become more pronounced as regulation develops. But for the engineering profession, the destination itself cannot change.

Technology can support the decision. Professional accountability cannot be delegated to the technology.