What Should AI Automate?

"Can this be automated" is almost always the wrong first question when scoping AI in a design workflow. Most repetitive tasks in architecture and BIM can, technically, be automated. The better question is what happens on the occasions the automation gets it wrong — because that answer determines whether automating the task is actually a good idea, regardless of whether it's possible.

A simple test: cost of being wrong

Some tasks have a wide margin for error. If an AI suggests a slightly different phrasing for an internal note and it's not quite right, someone edits it. Low cost, easy to catch, no real downside to automating it fully.

Other tasks have a narrow margin. If an AI automatically renames every instance of an element type across a model and its logic is subtly wrong, the mistake replicates instantly and silently, and the cost of finding and reversing it can exceed the time the automation was meant to save. Same category of task — repetitive, rule-based — completely different risk profile once something goes wrong.

The right question for any candidate task isn't "is this repetitive" or "could an AI plausibly do this." It's: if the AI gets this specific instance wrong, how expensive is that mistake, and how quickly would anyone notice? Tasks that fail cheap and fail loud are strong automation candidates. Tasks that fail expensive and fail silent are not — at least not without a person checking before the change takes effect.

Rule-based versus judgment-based

A second useful filter: is the task genuinely rule-based, or does it just look rule-based from a distance? Checking whether a naming pattern matches a documented standard is rule-based — there's a defined correct answer, and a system can check for it reliably. Deciding whether a design choice serves the project's actual intent is judgment-based — it depends on context, trade-offs, and priorities that shift from project to project and aren't fully captured in any document.

A lot of tasks that look automatable on the surface turn out to have judgment buried inside them. "Check if this room meets clearance requirements" is rule-based. "Decide if this layout works" is not, even though both involve looking at the same drawing. Automation works well on the first kind of task and works poorly — confidently, in a way that's hard to catch — on the second, because the system will produce a confident-sounding answer to a question that didn't actually have a purely rule-based answer.

Full automation versus flag-and-approve

Even within genuinely rule-based tasks, there's a further distinction worth making: should the AI just do it, or should it flag the issue and let a person approve the fix?

Full automation makes sense when a task is rule-based, low-stakes if wrong, and high-volume enough that requiring approval on every instance would create more friction than it's worth. Fixing an obviously duplicate warning, for instance, might reasonably not need a person's sign-off every time.

Flag-and-approve makes sense for anything with a meaningful cost if wrong, anything where "wrong" isn't always obvious at a glance, and anything that affects a part of the model other people depend on. Most of what actually matters in BIM and design coordination falls into this category — which is why flag-and-approve, not full automation, tends to be the right default for the tasks worth an AI's attention in the first place.

The scoping exercise worth doing

For any task under consideration for automation, three questions are worth asking in order: Is this genuinely rule-based, or does it have judgment hiding inside it? If the automation gets it wrong, how expensive is that mistake, and how quickly would it be noticed? And given the answers to those two, does this task deserve full automation, or does it deserve a flag with a person's approval before anything changes?

Most of the value in applying AI to design workflows doesn't come from automating the most tasks possible. It comes from automating the right ones — the narrow, cheap-to-verify, genuinely rule-based slice of the work — and leaving everything else, correctly, in the hands of the people accountable for getting it right.