I didn’t know this when I started, but in many parts of the US you can’t legally sign or submit technical drawings or reports for public projects unless you’re a licensed Professional Engineer. It’s a risk thing whereby you have to prove technical competence. I don’t want to drive over a poorly designed bridge.

What this means with AI is that it’s likely going to take a long time until we see AI completing and delivering full designs. As engineers, we still need to review and investigate any design it creates to sign it off and defend it if we are ever sued. AI doesn’t hold this risk, legally we do and that fact influences how AI systems are defined for construction.

For us, we are seeing these three worlds emerge when working with AI:

Here AI does part of the work but can’t finalize a decision without an engineers approval.

In this scenario AI is more autonomous. It can make and execute a decision but we still supervise, monitoring outcomes, auditing and intervening as needed.

A good example is Waymo where the car self drives but there’s a remote team that can take control.

This is the point where we start getting scared about our jobs. Here a system is fully autonomous and completes tasks automatically. AI isn’t really there yet for a full design or even to manage my inbox but honestly, I don’t think I want it here.

The value of this framework is to understand where we are moving towards. Each level of control is a tradeoff between factors such as safety, speed, risk and learning. With that we can gauge which parts of our work can be delegated to AI, and where we still need to be accountable for the result.

It’s a dial, not a switch

These three levels aren’t fixed. The same workflow can sit at different points depending on the risk and it can move over time as trust in the system grows. A new tool might start fully in the loop and then shift to on the loop once it is proven across a few projects.

Deciding where a task should sit comes down to five trade-offs:

So which of my tasks fit where?

A simple way to sort out where your work sits is to list out the task and ask two questions:

Automate (out of the loop)
Low consequence, fast to verify, highly repeatable.

Examples include formatting reports, renaming and filing drawings, transcribing site meeting minutes, and populating transmittals. If AI gets the answer wrong, nobody gets hurt and it’s easy to fix.

Supervise (on the loop)
Medium consequence, high volume.

Examples include first-pass submittal reviews against spec, model QA checks, and tracking RFI responses. AI runs the process and you can review the exceptions it flags.

Assist (in the loop)
High consequence or anything you sign.

Examples include design calculations, structural checks, value engineering options, and anything going out under your stamp. AI can draft, check or suggest alternatives, but you make the call.

Keep it human
Judgment, relationships and accountability. This covers negotiating a variation, explaining a design decision to a client, or deciding whether a site condition is safe.

The bottom line

AI isn’t replacing engineers but, in the future, every engineer will be using AI. Most of us will spend less time producing and more time reviewing and supervising.

The engineers who are able to get the most done will be those who know which bucket their tasks below to and sets up these tools accordingly.

Start in the loop, build trust and turn the dial gradually.

Got a better way of doing this on your projects? Just reply — we read everything.