Table stakes, judged on the details.
Many AI products already ship some version of these seven patterns. What separates them is a handful of decisions: what happens after you stop, which claims get a source, and who made which change. Each section below names the problem, the decision, and how it's built. All screen content is example data.
Stop is not the end.
Hitting Stop can leave the answer frozen or broken, so users wait out a wrong answer or start over.
Stopping starts steering. Everything written so far is kept, and only the part you point at is redone.
Stop aborts the stream, and resuming sends the kept text back with your new instruction.
For monthly specialist visits, Plan B likely costs less over the year. Its specialist copay is lower, and you'd hit the deductible on Plan A before savings kick in.
On premiums, Plan A is cheaper per paycheck, but that gap closes after
Thanks for sending the renewal terms. I'm happy with the scope, but I'd like to push the start date to March and revisit the rate after the first quarter so we can
Only the highlighted sentence is regenerated. Draft 1 is saved, and the change is shown before it replaces anything.
For monthly specialist visits, Plan B likely costs less over the year. Its specialist copay is lower, and you'd hit the deductible on Plan A before savings kick in.
On premiums, Plan A is cheaper per paycheck, but that gap closes after|
Show me where that came from.
About one in four AI search citations don't support their sentence, and people tend to trust a citation without checking it.
Cite each sentence, not the whole answer, open the exact passage in place, and say plainly when there is no source.
The model returns each sentence with a source id and the quoted span, which is matched against the document.
Yes, outpatient MRIs need prior authorization1, and your doctor's office usually submits it. Approval typically takes a few business days.
Most participants compare plans only during open enrollment3, and several felt unsure what a deductible covers5.
“…honestly I don't look at any of it until the enrollment email shows up. That's the only time I compare plans, and I just go with whatever's cheapest…”
Exact passage highlighted · 1 of 2 supporting quotes ›
Yes, outpatient MRIs need prior authorization1, and your doctor's office usually submits it. Approval typically takes a few business days.
Diagnostic tests and lab work are covered at 80% after deductible.
Advanced imaging (MRI, CT, PET) requires prior authorization when performed in an outpatient setting.
Emergency imaging does not require approval.
Confidence without numbers.
People over-trust wrong answers, and plain hedging like "I'm not sure" reduces that better than long explanations.
Mark confidence on each claim by where it came from, and give every weak claim a way to check.
Each claim comes back tagged as in your documents, one source, or inference.
Yes, PT after surgery is covered under your rehab benefit. There's a 30-visit yearly limit, and you probably don't need a referral.
Yes, PT after surgery is covered under your rehab benefit. There's a 30-visit yearly limit, and you probably don't need a referral.
Your plan summary doesn't mention referrals for therapy.
Change by change.
Ask the AI to fix one paragraph and it quietly changes others too.
One change is the unit of consent, each with a one-line reason, and the AI never adds facts you didn't give it.
The model returns a list of changes (from, to, reason), shown as a diff with accept and reject on each.
I am writing to complain about appeal the denial of claim #A-1029.
The procedure was really necessary because my doctor said so medically necessary, as documented in Dr. Lee's note dated May 2.
Please fix this ASAP reconsider within 30 days.
Sincerely,
Jordan
Every change can be undone.
Undoing one AI edit can wipe far more than that edit, even in tools built for it.
History is grouped by intent, shows whether you or the AI made each change, and undoes one change without losing the later ones.
Each change is stored with the instruction that caused it, so undo removes just that change.
I am writing to appeal the denial of claim #A-1029.
The procedure was medically necessary, as documented in Dr. Lee's note dated May 2.
I understand mistakes happen and appreciate your help. This line goes away if you undo “Softened the opening.”
Please reconsider within 30 days.
Starters that read the room.
A blank box makes people guess what the AI can do before they can ask anything.
Three suggestions at most, built from what's on screen, each saying why it's there.
Screen data goes to the model, which returns three questions, each with its reason.
| Billed | $1,180.00 |
| Plan paid | $940.00 |
| Status | Partly denied |
Reasoning on request.
Showing the full reasoning raises trust without improving accuracy, and in one model, full traces made accuracy worse.
Collapsed by default, opened automatically when the stakes are high, and shown as assumptions you can fix.
The model returns its assumptions as editable fields, and changing one reruns the answer.
Plan B looks better for you, by about one month of premiums over the year.
Change an assumption and the answer updates.
I flagged 2 items as personal, so your report total is lower.
Plan B looks better for you, by about one month of premiums over the year.
| Plan A | Plan B | |
|---|---|---|
| Premiums | Lower | Higher |
| Specialist copay | Higher | Lower |
| Estimated year | More | Less |