Could AI design a bridge?

The concept of AI designing a bridge triggers mixed feelings. Excitement about new tools, worry about job security, and concerns about safety and liability.

In the last two years, AI has already reshaped law, finance, and software. Architecture, engineering and construction (AEC) will not be far behind.

Design automation was the topic of my MSc and doctoral thesis. Since then, I've worked across the industry – onsite and in the design office as well as in academia and startups.

In this article, we'll look at practical ways to implement AI in engineering design and identify where it can add real value.

Technical approach

Full autonomy is seen by many as some inevitable goal upon the horizon, but I really don’t think it is where we should be aiming.

I do chuckle when I read things like “AI will do the design work and humans will sign it off” - always written by somebody who doesn’t really understand design.

Construction isn’t a simple industry to automate. Linking multiple difficult problems just creates a fragile and opaque system.

There is, however, a lot that we can do today if we recognise the value in knowledge access and workflow optimisation that is available right now.

Optimising our time

Design involves complex activities performed by specialists and generalists across many interfaces.

High-value activities like analysis and calculations make up only about 15% of the design process.

Low-value activities take up much more time:

• Searching for information (20%)
• Report writing (25%)
• Administrative work (10%)

If we focus on automating the low value (and often tedious) tasks, there are many benefits:

• More time for concept exploration (where costs generally become fixed on sub-optimal options)
• Targets specific pain points
• Enables companies to learn AI in low-risk scenarios
• Greater professional satisfaction
• More time to engage in meaningful stakeholder communication
• Less errors – which often happen due to rushing design tasks

Over the next few years, assurance frameworks will emerge to enable compliance with standards and best practices.

This will naturally evolve toward more complex tasks as we gain confidence and capability. Eventually, it will transform the way we do everything.

AI as a design assistant

Consider this convoluted (but painfully familiar) example of a bridge engineering workflow:

1. Searching through technical reports and email threads for soil parameters.
2. Checking the Basis of Design document, multiple standards and Craig’s Soil Mechanics for guidance.
3. Updating parameters and geometry in your PLAXIS model, running and… waiting.
4. Trying to remember whether "Foundation_Design_FINAL_v4.xlsx" or "Foundation_Design_APPROVED_Rev3.xlsx" is the correct version to update.
5. Update the technical report with new plots, quantity estimates and check the conclusions are still valid.
6. Sign-off and then email to the client.
7. Only to find there is a factor buried in an email thread somewhere that requires a redo.

This involves many low-value and time-consuming steps of finding and moving data to facilitate the high-value work.

This process could be enhanced significantly. Imagine instructing a design assistant:

"Follow these steps and provide a summary of all assumptions, data sources and the outcome for each one:

1. Pull the latest soil parameters from phase2 GI emails and note changes.
2. Crosscheck against the Basis of Design, Eurocodes, our reference library and past projects; flag any risks.
3. Update the PLAXIS model. Optimise three options—material efficient, fast to build, and tolerant to soil variability. Run ±20% cohesion sensitivity and present as a decision matrix.
4. Draft a memo with assumptions, results, and version history; link raw data and all analyses.”

In this scenario, the engineer is using AI to automate workflows and deal with transferring data between them.

It’s not hard to imagine small teams of talented engineers collaborating to create designs that meet requirements, are auditable and can be easily reconfigured to accommodate evolving project requirements.

However, there are substantial challenges.

Validation

AI systems that can act independently (agentic systems) are hard to build, scale, and validate.

Engineers need subject knowledge and communication skills to explain tasks and set validation metrics.

Connecting multiple AI workflows is technically challenging with many interfaces to navigate.

Companies rushing to build design agents risk poor governance or missing important edge cases.

End users may not have the technical skills to validate AI systems. This means they depend on the agent's creator, which might be an early-stage startup.

User acceptance

Beyond technical validation, user acceptance poses formidable challenges. If AI seems difficult to use, threatens jobs, or appears untrustworthy then adoption is a non-starter.

Nationally critical design projects are delivered by tier 1 consultants through multi-year framework agreements on a billable hours basis. Autonomous AI threatens their core business model.

Furthermore, will regulatory approval be needed and if so, what would this look like? Do we need standards covering how to use AI for design tasks? What is a design task? If we follow CDM 2015 then nearly everything is.

Then there is liability. One of the reasons clients are happy to pay more money for a human is because there is somebody to blame when it goes wrong. What is the value of this peace of mind?

My guess is that the technical challenges will be dwarfed by regulatory, legal and perception challenges. AI-related construction disputes will be common in 3-5 years, with unclear liabilities.

We’re still finding our feet with AI

Could AI design a bridge? Theoretically, yes, but we know that theory doesn’t equal practice, and execution is everything.

The practical path forward isn't to immediately pursue autonomous design. Rather, it's to systematically implement AI in lower-risk, higher-value tasks with clear returns on investment.

This will establish the technical and governance frameworks needed for more advanced applications.

As an industry, we are still finding our feet with generative AI.

But we have a real opportunity to push the limits of engineering and reinvent our industry. The decisions we make today will shape the physical world for decades or beyond.

Our responsibility is to ensure that this legacy reflects the best of human wisdom and technological capability.