Where AI Belongs In Plan Review

There is a natural temptation when people start to think about applying AI to plan review: Why don't we just put the smartest AI we can build at the very beginning of the process and ask it to catch everything before a reviewer ever sees the application? 

The appeal of this front-end-heavy approach is obvious. If AI can identify missing information, compare a submission against requirements, and flag potential problems before the application reaches a reviewer, applicants get feedback faster and reviewers spend less time catching problems that could have been prevented upstream.

Generally, we think that's the right idea.

We just don't think every part of plan review belongs there.

At e-PlanSoft, we tend to start with the job someone is trying to accomplish, rather than the technology or feature we want to apply to it. When we apply that thinking to AI, an important distinction emerges.

Some parts of plan review are based on decisions that are relatively clear, repetitive, and objective. They’re black and white.

But there are also lots of “grey area” decisions that local government agencies run into everyday. Those depend heavily on the experience, expertise, and judgment of a professional reviewer.

And it's our belief that AI should handle those two kinds of work differently.

Automate The Simple Stuff 

Here's our first basic belief about AI in plan review: start with the easy stuff.

Is the file readable? Are required sheets present? Is the document locked? Are pages numbered correctly? Is required information missing?

These are important questions, but they usually don't require a plan reviewer's attention or judgment. That makes them great candidates for automation. The same principle can extend deeper into the review process.

If a jurisdiction allows a certain amount of signage per square foot of building frontage, for example, determining the proposed sign area and comparing it against a clearly defined limit is largely a measurement problem.

Or think about what happens when a corrected plan set comes back for another round of review.

A reviewer may need to pull up the previous submission, find the comments they made, locate the corresponding changes in the new plans, and figure out what the applicant changed in response.

AI can do a lot of that hunting. It can compare the two submissions, identify where something changed, and help connect those changes back to comments from the previous review. These are different kinds of tasks, but they have something important in common: there is usually a knowable answer.

At intake, AI can inspect a submission, identify obvious gaps, and help applicants correct them before the application reaches a reviewer. During review, it can find, organize, compare, and surface information that a reviewer would otherwise have to hunt for manually.

Applicants get feedback sooner. Review teams spend less time opening submissions that aren't actually ready to review or searching through documents for information the technology can find for them.

We've written before about why a complete submission is not necessarily the same thing as a reviewable submission. The goal here is to remove predictable friction before a reviewer has to deal with it.

The harder question is what to do when the work stops being so predictable.

Assist Where Judgment Matters

Take a resubmittal. AI can compare the new submission with the previous version, show the reviewer what changed, and reconnect those changes to the comments from the prior review.

That can eliminate a lot of searching and page-flipping.

But identifying a change is not the same thing as deciding whether the change resolves the original issue.

Maybe the applicant revised the detail that prompted Comment 17. AI can put the old detail, the new detail, the original comment, and the relevant requirement in front of the reviewer.

But the reviewer still has to decide: is this issue fixed now? That distinction gets at the second part of our approach to AI.

Plan review isn't just a checklist. Reviewers interpret drawings in context. They understand local requirements. They recognize unusual conditions and know when an exception matters. AI can make that kind of work faster. It can find relevant requirements, surface information buried in a plan set, suggest comments, and point reviewers toward something that deserves attention.

But there is an important difference between helping a reviewer make a decision and making the decision for them. And in judgment-heavy work, putting that decision on AI too early can actually make the process worse.

Imagine AI reviews an application at intake and tells the applicant that a particular condition is acceptable. The applicant moves forward believing that issue is settled. Then the application reaches a professional reviewer, who looks at the full context and disagrees. Now the applicant has two answers.

Worse, the reviewer (the person responsible for making the final determination) has become the person slowing the application down. From the applicant's perspective, the technology said yes and the human said no. That's exactly the kind of friction we want AI to remove, not create. Our preference is to give intake AI the repetitive, straightforward work it can do reliably, and then use AI to make the reviewer faster and better informed when professional judgment is required. That creates a better experience on both sides. Applicants can trust the feedback they receive early in the process. And when a judgment call has to be made, the reviewer is equipped to make it faster rather than being forced to contradict a decision the technology already presented as settled.

That's why we've consistently argued for using AI to support plan reviewers rather than trying to replace their judgment.

For work that requires professional judgment, we think AI should sit beside the reviewer rather than in front of them. The distinction can be pretty simple:

• Find what changed on resubmittal? Let AI do more of the work.

• Decide whether the change the applicant made actually resolves the comment? Help the reviewer get to the answer faster, but let the reviewer make that call.

Learn Where To Focus Next

This is where the approach gets more interesting.

Suppose AI begins by flagging a particular condition for a reviewer. The AI says, in effect: Based on what I'm seeing and how the code is written, you may want to look at this. Then, the reviewer decides whether it actually matters.

Now imagine that happening across hundreds or thousands of reviews inside a single jurisdiction.

Over time, a pattern emerges and historical data is captured. Let's say that jurisdiction’s reviewers agree with one particular type of AI recommendation almost every time.

That's useful information. Something that jurisdiction initially treated as a judgment call may actually be predictable enough to handle earlier in the process, or more automatically with AI. Maybe reviewers discover that when a particular condition appears, they reach the same conclusion 99 percent of the time.

Now there is a reasonable question to ask: does this kind of decision still need to wait for a reviewer? Or can AI handle it on its own?

With enough evidence (and an ability to interrogate their own data), the jurisdiction could decide to tune that part of their plan review workflow so that AI can flag the issue and an applicant can be given the option to address it before completing their submission.

Another AI recommendation may turn out differently. Let’s say reviewers agree with it only 40 or 50 percent of the time.

That level of disagreement is a signal that the reviewer is still applying important human judgment. The task should stay with the reviewer, while AI continues helping them get to the answer faster. The important point is that we don't have to decide permanently, in advance, whether each plan review task is handled by AI, by a human, or by some combination of the two. We can learn from the work itself.

And this becomes much more useful when AI is connected to the actual plan review workflow. The system doesn't just know what AI recommended. It can learn what the reviewer ultimately decided, which comments were applied, what changed on resubmission, and where the same decisions keep showing up again and again.

That creates a feedback loop between the work being done today and where AI can create more value tomorrow.

Our AI philosophy can be summarized pretty simply:

• Automate the simple stuff.

• Assist where judgment matters.

• Help us learn where to focus next.

The goal isn't to decide today how much of plan review AI can eventually do. (We don’t think it’ll ever do the entire job of an experienced plan reviewer, for what it's worth.) It's to keep getting better at identifying the work where AI can help, while preserving the experience and human judgment that make plan reviewer valuable in the first place.