Fix the submission before anyone has to review it

Most document projects I hear about start at the reviewer's desk, after the application has already arrived with a page missing, and someone now has to notice it, write back, wait for the reply, and then read the whole thing again when it comes back. So the AI gets pointed at the reviewer, to read faster and summarize and flag problems, which does help... but by then the bad submission has already happened. I want the check earlier than that, and the examples I've found most interesting this year start with the person who is actually filling in the form.

Three teams that moved the check upstream

In Honolulu, StateScoop reports, citing Census Bureau figures, that the county gets about 20,000 building permit applications a year. The planning department now uses a tool that walks applicants through code requirements before their plans ever reach staff, pointing out possible code issues and missing information while the applicant can still fix them. The director, Dawn Takeuchi Apuna, compared it to tax software ("It's kind of like TurboTax for permitting."), and she says it has cut residential review times by roughly 60%. I'd add two caveats because they matter -- that figure is hers rather than one the article checked, and so far the tool only covers residential single and two-family permits.

Travelers took a different slice of the problem. Insurance Journal reports that its AI claim assistant is initially being used with customers who call to file an auto damage claim, so they started with one kind of claim, not all of them. What stood out to me is that the assistant can help customers make an informed decision about filing a claim, which means it starts working before the claim does, and customers can speak with a live specialist at any point... so the scope is small and a person is always one step away.

The third one is Ping An in China, where a Bloomberg piece republished by Insurance Journal says that five years ago almost no accident and health claims ran without humans, and today nearly 60% are automated. The part I care about most is the scene itself, because when a car accident happens at 3 a.m., an AI agent can respond at once and direct the photo and video capture, so the evidence gets gathered while the car is still sitting there and not pieced together from memory a week later.

All three put the check where the mistake actually gets made. A missing page is cheapest to fix while the applicant still has the file open, and a week later that same page costs an email, a wait, a second read, and an applicant who now thinks you're the slow one. Nobody did anything wrong in that story -- the question just got asked too late.

None of them hand the reviewer's job to the submitter either. They ask the questions a good reviewer would ask, only earlier, in plain language and one at a time, which is what I take from the TurboTax comparison... you don't need to know the rules if the form asks you the right thing at the right moment.

You don't need a new system to start on your own intake, but you do need a list. Pull the last few months of submissions that bounced back and write down why each one came back, in plain words, meaning the actual reason and not the status code. Then for each reason ask whether the person submitting could have caught it if someone had asked them while they were uploading, and a missing signature, an old version of the form, a blurry photo or page four of six usually land in that group. Those reasons become your intake checks, and they go in front of the submitter before they press send. Start with one form, the way Travelers started with one kind of claim, get that right, then move to the next one. And keep a way for people to reach a person, because some of them will want one and they should be able to get it.

Honolulu also mentions a side benefit I didn't expect, which is that the tool helps train staff. Apuna says "It'll highlight and flag a part of the plans that we're focused on," and that gives new employees a clear, consistent way through a plan. That makes sense to me, because you can't write a check down until your team agrees on what it's checking, and that agreement is most of what a new reviewer needs to learn anyway.

What a form can't catch

Some bounce-backs are judgment calls though, and no form is going to catch them... whether a design suits the street, whether an odd situation is covered, whether a document that's technically complete actually says what it needs to say. I don't want an intake app that pretends it can judge those, because it will annoy people, it'll either reject good submissions or hand out a green tick that a reviewer later overrules, and the second one is worse than no tick at all. Be plain on the screen instead -- say the submission looks complete and a person will still make the call. Applicants can live with that. What they hate is being told they're done when they aren't.

So the work gets split -- the form catches what the submitter can fix and reviewers spend their time on what only they can decide. The way we think about it at Workbricks, the submitter's app and the reviewer's app are two different small apps, and both can sit on the same documents.

Fix it at upload, not a week later.