Engineering in the age of AI

Dr.-Ing. Lena Golgolnia · 21 October 2026 · 10 min read

In short

AI is entering engineering practice fast: in a VDI survey, more than three quarters of engineers and programmers said they use language models, and 59% of UK architecture practices use AI.^1,2
The value is greatest early in a project, because errors grow expensive later: fixing a requirements error in operation can cost from 29 to more than 1,500 times as much as in the requirements phase.⁶
AI already helps with requirements review, design exploration and documentation. Automated compliance checking remains limited, because regulations are complex, legal texts are ambiguous and training data is scarce.¹⁰
Responsibility does not move to the machine. Whoever signs the documents remains liable, and AI results must be checked for plausibility.¹⁴
The engineer’s role shifts from producing documents to specifying, verifying, integrating across disciplines and deciding.

The problem AI meets

Engineering for the built environment has a long-standing productivity problem. The McKinsey Global Institute found that construction labour productivity grew by about 1% a year over two decades, against 2.8% for the world economy and 3.6% for manufacturing.⁴ Much of an engineer’s time goes into producing, coordinating and checking documents rather than into the decisions those documents record.

This is the work generative AI is good at, and engineers have noticed. In a 2025 VDI survey of 468 participants, mostly VDI members, more than three quarters of engineers and programmers said they already use language models, and 75% of respondents expected noticeable time savings, especially on routine tasks.¹ In the UK, the share of architecture practices using AI rose from 41% to 59% within a year.² In Germany, adoption among architecture offices is lower: 29% of members of the federal chamber of architects used AI tools in daily work in 2025.³

>75%of engineers and programmers in a VDI survey use language models¹

59%of UK architecture practices use AI²

29%of German chamber architects use AI tools in their daily work³

Early results from capital projects show what is possible. McKinsey reports a European transmission system operator whose AI-based parametric design tool needed up to 88% fewer design and procurement hours, and an AI-assisted value-engineering review of more than 500 specifications in a $500 million mechanical, electrical and plumbing package for a factory, which found savings of 15–20% against the initial baseline.⁵ These are individual cases, not averages. They show the direction, not a guarantee.

Why early phases matter more than ever

Every engineer knows the rule that mistakes get more expensive the later they are found. NASA quantified it: if fixing a requirements error in the requirements phase costs one unit, the same error costs 3–8 units in design, 7–16 in construction, 21–78 in integration and test, and from 29 to more than 1,500 units in operation.⁶

Figure 1. The later an error is found, the more it costs to fix.⁶

This curve cuts both ways for AI. Used early, to find gaps and contradictions in requirements or to explore many design options before committing, AI has its largest economic effect. Used without checks, it can introduce errors early that are only found late, at the most expensive point. The value of AI in engineering is therefore inseparable from the quality of verification around it.

AI across the engineering lifecycle

The maturity of AI differs strongly from one phase to the next. A realistic view separates what already works in practice from what is still research.

Table 1. Where AI helps in the engineering lifecycle, and how mature it is
Phase	What AI does	Maturity
Requirements	Scores requirement quality and suggests improvements; generates test cases⁹	Mostly research; few production uses⁷
Concept and design	Explores many options quickly through surrogate models; parametric design^5,13	Proven in research and individual projects
Compliance checking	Checks designs against codes and standards	Limited; complex, ambiguous rules^10,12
Specification and review	Reviews large sets of specifications for value and consistency⁵	Emerging in large projects
Documentation	Drafts reports, descriptions and correspondence	Widely used today¹

Requirements: promising, but early

A systematic review of 238 studies on generative AI in requirements engineering found that over 90% were early-stage research and only 1.3% reached production-level integration. The most cited problems were reproducibility and hallucinations, each named in about two thirds of the studies. The authors recommend AI as an assistant for validation, not as an autonomous tool.⁷ A 2024 workshop of the Systems Engineering Research Center reached a similar conclusion, warning of over-reliance and naming the right placement of human involvement as the open challenge.⁸

Design exploration: faster options

Surrogate models, machine-learning models trained on detailed simulations, can estimate the performance of hundreds of thousands of building designs within seconds. Their limit is approximation error, which is why research focuses on surrogates that state their own uncertainty.¹³ For design teams, this changes the economics of exploring options: comparing many variants early becomes cheap, and the detailed simulation is reserved for the few that matter.

Compliance checking: the hardest step

Automated checking of designs against building regulations has been a goal for decades. A 2025 review concludes that it remains largely manual and that a comprehensive automated system is still a distant goal, because of regulatory complexity, ambiguous legal language and scarce training data.¹⁰ Progress is real where regulation and data formats are designed for it. Singapore’s CORENET X digital approval system is mandatory for new projects of 30,000 m² and more since October 2025, and since 1 October 2026 also for projects of 5,000 m² and more.¹¹ In Germany, a feasibility study in North Rhine-Westphalia tested BIM-based building permits on nine real projects and found that they work only when modelling guidelines, pre-checks and review rules operate as one system.¹²

What changes for the engineer

INCOSE’s Systems Engineering Vision 2035 expects AI to make systems engineers more efficient and effective, and the practice to become predominantly model-based.¹⁷ Working across architecture, sensor hardware and software, I see three shifts in what engineers actually do.

From draftingto specifying

When AI produces drafts, the precision of the input decides the quality of the output. Clear requirements, constraints and acceptance criteria become the engineer’s main product.

From producingto verifying

Output volume rises; the critical skill is to judge whether a result is plausible, complete and safe, and to reject it when it is not.

From disciplinesto systems

AI speeds up each discipline. The value lies in the interfaces between structure, services, energy, software and use, where errors are expensive.

Figure 2. Three shifts in the engineer’s role.

None of these shifts makes the engineer less important. They move the engineer’s contribution from volume to judgement, which is harder to automate and more valuable to clients.

Responsibility, liability and data

The professional bodies are clear on one point. The German Federal Chamber of Engineers states that results of AI-based calculations cannot replace human judgement, that engineers must check AI-based outputs for plausibility, and that whoever signs planning documents remains liable, including towards the client.¹⁴ VDI strongly advises against processing sensitive technical data on open platforms.¹

The legal framework is moving in the same direction. Under the new EU Product Liability Directive, software, including AI systems, counts as a product; the new rules apply to products placed on the market after 9 December 2026.¹⁵ The EU AI Act requires organisations that use AI to take measures supporting the AI literacy of their staff.¹⁶ For engineering offices, this means documented choices about which tools are used for what, with which data, and how results are checked. This article gives general information, not legal advice.

Principles for practice

Table 2. Six principles for using AI in engineering
Principle	In practice
Use AI earliest where errors cost most	Requirements review and option studies before detailed design
Every AI result has a checker	A named engineer checks plausibility before results are used or signed
Keep the source of truth	Standards, calculations and models remain the reference; AI output points to them
Protect client and technical data	No sensitive data on open platforms; clear terms with tool providers
Design for interfaces	Coordinate across disciplines; AI speeds parts, people own the whole
Record what was generated	Note where AI was used and how results were verified

Engineering has always been the discipline of making things work in the real world and taking responsibility for them. AI changes how much engineers can produce and explore. It does not change that responsibility. The engineers and offices that gain most will be those that use AI to spend less time on documents and more on the decisions the documents exist to support.

Frequently asked questions

Will AI replace engineers?

The evidence points to a change of role rather than replacement. Engineering bodies stress that AI results cannot replace human judgement and must be checked for plausibility, and the person who signs remains liable.

Can AI check building designs against regulations automatically?

Partly. Digital permit systems such as Singapore’s CORENET X and German pilots show progress, but research finds fully automated code checking still a distant goal, because regulations are complex, legal texts are ambiguous and training data is scarce.

Where does AI help most in an engineering project?

Early: in reviewing requirements for gaps and contradictions, exploring design options quickly with surrogate models, and preparing documentation. Errors avoided early save the most cost.

Is it safe to put project data into public AI tools?

Not for sensitive technical data. VDI strongly advises against processing sensitive technical data on open platforms. Use tools with clear data protection terms and keep client data under control.

What skills matter more for engineers now?

Precise specification, systems thinking across disciplines, verification and testing, and the judgement to accept or reject a result and take responsibility for it.