Someone asked Josh for help picking a private health insurance plan. He had talked to a bunch of PPO providers a few years back, so they figured he would know.
He did not, really. It had been three or four years and he couldn't give specifics anymore.
So instead of guessing, he told them what to do. Get Perplexity Pro. Find the deep research function. Take the documents the insurance people sent you and put them in. Then start asking.
And then he walked them through what to ask, which is the part worth paying attention to.
Give it the context first
Not "is this a good plan." Tell it the situation. I am looking for an insurance plan, I am looking at these PPO plans.
Then the specific questions. What is the limit? Is it a lifetime limit or a single occurrence limit? What are the co-pays? What are the benefit limits?
Then the question that actually matters:
What are the gotchas?
What are the things most people miss when they're looking at a document like this? Do you know any gotchas specific to these providers?
Why that last question is the good one
Because it's the question you can't ask a search engine, and it's the question an expert would answer for you if you happened to know one.
The specific facts in those documents are findable. Tedious, but findable. What a normal person lacks is the pattern across hundreds of similar documents that tells you which clause is standard and which one is unusual and going to cost you.
That pattern is exactly what these models have. Not because they understand insurance, but because they have processed an enormous amount of text about it, and "what do people usually get caught by" is a pattern question.
Josh summed up the value like this: you can go through tens or hundreds of pages of documents and distill it into what a normal person actually needs to know before they buy insurance.
It generalizes
He immediately extended it himself. It doesn't have to be insurance. It could be a car. What are common problems with this model, what are the recalls.
Same shape every time. A long document or a complicated purchase, a normal person with no specialist to call, and a set of things that experienced people know and beginners get burned by.
Lease agreements. Loan terms. Employment contracts. Medical consent forms. Warranty documents. The stack of paperwork at a closing that everyone signs without reading because reading it would take four hours and you would not understand it anyway.
What this reveals about where AI is useful today
I keep coming back to this story because it cuts against how AI usually gets sold.
The pitch is normally about doing more. More output, more content, more velocity.
This is the opposite. Nobody produced anything. Somebody understood something they otherwise would have signed blind.
That is closer to the real current strength: compression and orientation. Taking something too large to hold in your head and making it graspable, so you can make a decision as an informed person instead of a hopeful one.
There is also something worth noticing about who benefits. The person who already has a lawyer, an accountant, and an insurance broker in their phone did not need this. The person who doesn't, now has something between them and a document written by people whose interests aren't theirs.
That is a genuinely good thing, and it is the kind of good thing that rarely makes the marketing because it does not scale into a business model.
The caution, stated plainly
It will get things wrong. On something with real money attached, verify anything that will drive your decision against the actual document. The AI's job here's to tell you where to look and what to ask, not to be the final word.
And it doesn't know your situation. It knows what usually matters. Whether it matters to you, with your family and your health and your budget, is still yours to work out.
The version I would give a friend
Take the document. Put it in. Tell it your situation. Ask what you should be looking at, then ask what people usually miss. Then go read those specific sections yourself, carefully, in the actual document.
Twenty minutes, and you walk in knowing which three questions to ask instead of nodding along.
That isn't a productivity gain. It is closer to not being taken advantage of, which I think is a better use of the technology than most of what it gets used for.
From a conversation with Josh on J&C Unscripted.